Around April 1, I began sharing open-source projects on GitHub. In just three and a half months, these projects collectively earned over 10,000 stars. Key projects include GEOFlow, Tokhub, GEORank, yao-meta-skill, yao-open-skills, yao-geo-skills, an open collection of prompts, and several smaller internal tools and systems.

During this period, my primary AI coding assistant was Codex - consuming over 18 billion tokens. It helped me build numerous open-source systems that improved internal team management, internal system development, business analysis, and strategic support.

1. GEOFlow

Before launching, I did a quick survey and found very few high-quality open-source systems focused on GEO (Geographic Search Optimization). Among the ones I’ve seen so far, GEOFlow stands out for its functional completeness, content engineering rigor, and delivery standards - and has earned 2.9k stars on GitHub in its first three months.

Especially around GEO content engineering, we built a set of high-standard norms and workflows.

Since launch, GEOFlow has entered real business use cases - and people are using it in ways even richer than I originally imagined:

• A publicly listed company deployed it as their internal content management system
• A startup forked and extended it for commercialization
• Another company used it as a base to apply for software copyright registration
• Several individuals monetize it directly - e.g., offering deployment and operations services
• Many more use it for GEO content projects: building new websites or upgrading existing ones

Hearing this feedback genuinely makes me happy.

For GEO, there are three core capability modules: data, content, and distribution. GEOFlow currently excels strongest in content, followed by distribution, then data.

Content is the most critical layer: once quality and management are solid, much else falls into place.

At the content layer, GEOFlow’s architecture is the most mature - covering end-to-end workflows, automation, knowledge base management, and quality control. Its core strength lies in content engineering: turning content creation and governance into a repeatable, scalable discipline.

On distribution, GEOFlow supports single-site sync and multi-channel publishing - managing and releasing content across multiple sites from one system. It also supports API binding (e.g., to WordPress), with extensible interfaces ready for integration with other distribution platforms.

On data, GEOFlow provides foundational analytics: website metrics, content performance, log analysis, and cross-site distribution records - all aggregated into basic statistics and reports.

GitHub repo: GEOFlow

2. yao-meta-skill

Early on, both Professor Xiangyang and I developed a habit: studying and iterating meta-prompts - prompts that generate other prompts.

So when exploring “skills” (structured, reusable AI behaviors), we naturally turned our attention to meta-skills: skills that generate or refine other skills.

That led to yao-meta-skill - a standalone, open-source GitHub repository now with 2.1k stars.

It helps create higher-quality skills from scratch - and also iterate, audit, and improve existing ones.

Many friends told me they found the process behind creating and refining yao-meta-skill even more valuable than the tool itself - so here’s a brief walkthrough:

1. Architecture design: I drew from public technical analyses of Claude Code, community research, and official skill designs - giving yao-meta-skill a distinct architectural foundation.
2. Capability scope: I studied top-tier meta-skills - including Anthropic’s Skill Creator and OpenAI’s Skill Creator - to synthesize strengths: Anthropic’s engineering methodology and iterative feedback loops; OpenAI’s lean, focused design.
3. Workflow refinement: We especially strengthened user intent modeling - clarifying a new skill’s inputs/outputs, defining its boundaries, and adding a “silent search” feature that automatically finds and adapts world-class reference designs. That third technique alone significantly lifts skill quality.
4. Quality assurance: We added automated AI evaluation experiments - and validated results with both AI scoring and five rounds of blind human testing. yao-meta-skill outperformed alternatives across all tests.

Today, nearly every skill I build - from scratch or iteratively - is created and refined using this meta-skill.

Building yao-meta-skill, in essence, was a way to learn from global best practices and deepen my own understanding of skill boundaries, evaluation criteria, engineering patterns, and standards.

GitHub repo: yao-meta-skill

3. Open Skill Collections

Two main repositories host these:

• yao-geo-skills: All GEO-related skills - e.g., integrated scrapers + cleaners + diagnostic reporters for Doubao, ChatGPT, and DeepSeek; or skills that generate GEO performance evaluation frameworks.
• yao-open-skills: Everything else - public skills outside GEO and meta-skills. Examples include:
  • WeRead Report Skill
  • Tutorial Skill
  • Bayesian Decision Skill
  • Game Theory Skill
  • Contradiction Theory Skill
  • Business Model Skill
  • Requirements Elicitation Skill
  • Expert Simulation Skill

My favorite is the Bayesian Decision Skill.

It turns Bayes’ theorem into an ongoing, updatable judgment process for complex decisions.

Here’s how it works:

• Starts with an initial assessment based on current information
• Guides users through multi-turn dialogue to add variables, update posterior beliefs, and logs why each judgment shifted
• Outputs a bilingual Markdown + HTML decision report: full conversation history, evolving judgments, and concrete action recommendations

It fits product, growth, and business decisions - and equally well for personal choices like travel, relocation, or career shifts.

At its core, it answers: When information is incomplete and risk uncertain - how do we decide rationally whether to act? And how do we raise the bar on decision quality?

Then there’s the Tutorial Skill:

1. You input any topic + reference materials -> AI prioritizes those references, supplements intelligently, and filters out low-quality sources
2. AI generates a custom-deep tutorial - outputting PDF, Word, and HTML versions for self-study
3. Content is structured chapter-by-chapter, with AI-drawn diagrams inserted contextually - grounded in the pedagogical logic from my Course Marketing book (written three years ago while running an MCN)

The yao-geo-skills repo also hosts several GEO-specific skills - like the ChatGPT scraper + cleaner + analyzer + reporter combo.

Its workflow:

1. Prepare: question list, repetition count, target entity, entity type, OpenCLI profile, sampling interval
2. OpenCLI Browser Bridge connects to your logged-in ChatGPT Web session and runs randomized, spaced queries
3. Data is cleaned, aggregated, and output as JSON, Markdown, Excel, and HTML reports
4. The HTML report analyzes: mention rate, avg. mentions, Top 1/3/5 probability, avg. ranking, sentiment, source domains, URL structure, title intent - and compares your target entity against competitors
5. Entity classification distinguishes people, companies, products, concepts, and noise - excluding generic industry terms, broad category words, or URL parameters from competitor lists

Related links: yao-open-skills yao-geo-skills doc.laoyao.cn (demos + original design docs) yao-open-prompts (open prompt library)

4. TokHub

TokHub is an open-source platform for monitoring, routing, and managing AI API gateways - designed as a unified “AI API middle layer.”

Co-conceived and co-built with Professor Xiangyang, it uses:

• Frontend: React + TypeScript
• Backend: Go + go-chi
• Databases: PostgreSQL/TimescaleDB + Redis
• Messaging: NATS
• Deployment: Docker + Helm

Core capabilities:

1. Public monitoring & channel details View status, success rate, latency, model, pricing, and health scores for monitored API channels

2. L1 / L2 / L3 layered probing
   • L1: DNS, TCP, TLS, HTTP connectivity
   • L2: API key validity, model list, model availability
   • L3: Minimal real-generation test - to catch cases where the endpoint is reachable while the model itself is unavailable

3. User workspaces & private channels Users can bookmark public channels - or add their own endpoints and API keys Private keys are encrypted and isolated per organization - never exposed to the platform backend

4. One-click personal gateway Aggregate multiple public and private channels into a unified gateway for your agents or clients Routes requests by latency, success rate, or cost - and logs tokens, latency, cost, error types, and audit events

System architecture:

TokHub supports Docker self-hosting - with production pre-checks, backup/recovery, security scanning, and release validation baked in.

GitHub repo: TokHub

5. GEORank

GEORank is a curated suite of lightweight GEO tools - iteratively developed and now fully open-sourced.

It measures brand and website visibility in AI search - integrating site diagnostics, AI Q&A, optimization suggestions, keyword expansion, structured content generation, and admin controls into one flow.

Built for GEO, SEO, content, and brand growth teams - and for developers looking to build or extend their own GEO platforms.

Tech stack: FastAPI + Next.js. Currently migrating to version 2.0 frontend.

Live demo: georankhub.com GitHub repo: GEORank

6. Tok-Series Tools

All named with “Tok” - the first three letters of Token - then freely combined. Seven main tools, all open-sourced:

1. Toktra

A Chrome extension for webpage translation. Configure your own API. I mostly use it for highlight-to-translate.

2. Tokkit

A local token ledger - tracks usage across projects. Shows live stats in the top-right corner and exports an HTML visual report.

3. Tokdoc

A document manager supporting Word, PDF, HTML, PPT, Excel, etc.

• Online editing for Word, PDF, HTML
• Auto-syncs files from local directories
• Tracks views and engagement for public documents

4. Tokscr

A flexible web screenshot tool - with post-capture editing and saving options.

5. TokURL

A personal URL shortener - generate short links instantly, name them manually, or auto-generate random ones.

6. tokchat

A lightweight AI Q&A and learning system - ideal for internal company Q&A or external expert-facing chatbots.

Built during meeting breaks, iterated a few times - stable for internal and light external use.

All tools: yao-open-tools

7. AI Interview System

To better manage hiring, I built an AI Interview Manager - and kept refining it over the past few months.

It’s now deeply integrated with GitHub, Codex, and Feishu - enabling seamless cross-platform data flow and collaboration.

How it works:

1. AI Interview Manager (core) Central hub for candidates, resumes, AI-first interviews, auto-generated reports, feedback, scheduling, analytics, and sync status

2. GitHub Data Vault Each candidate gets their own directory in a GitHub repo. Workflow stages flow automatically via folder structure, workflow.json, candidate-index.csv, and Git commits - creating a natural, versioned evidence trail

3. Codex Skill (local intelligence layer) Interviewers pull candidate packages, generate evaluations, write first/second-round feedback, save schedules, submit Feishu tasks, and push analytics - all triggered locally

4. Feishu Execution Layer Feishu CLI auto-creates calendar events, sends private messages, and triggers reminders. Background workers handle async tasks: retries, status updates, and weekly interview data reports

End-to-end flow:

1. HR or candidate uploads a resume -> system parses it asynchronously
2. After AI-first interview, system saves Q&A history, keywords, and scores
3. Report service generates HR- and candidate-facing reports -> agent auto-syncs to GitHub vault
4. Team members view candidate data in GitHub or Codex, generate assessments, and submit feedback
5. If “pass”, candidate moves to next-stage folder in GitHub; if “reject” or “hold”, only status and notes update
6. Once scheduled, system creates Feishu calendar event -> worker triggers follow-up reminders 1h and 3h post-interview

In this system, the candidate becomes a stateful object flowing across AI interviews, GitHub collaboration, Codex agents, and Feishu touchpoints.

That gives hiring four superpowers:

• Automation efficiency
• Clear human accountability boundaries
• Organizational memory of the hiring process
• Full operational traceability

For interviewers, Codex is the central interface - everything else syncs automatically.

Whether HR or interviewers, we all default to asking Codex directly: “Show me initial interview counts and feedback for the last 7 days.”

The AI Interview System remains private for now. If interest grows, I plan to release it.

In conversations with many friends, we’ve reached a shared judgment: AI raises access while outcome gaps can still widen.

AI lowers barriers to tool use and product development, and it continuously raises the market’s bar for final results. As more people use AI to deliver baseline work, what truly differentiates outcomes narrows to just a few things:

• Depth of problem understanding
• Strength of professional judgment
• Quality of taste and aesthetics
• Consistency of execution
• Discipline of iteration

In short: AI raises the floor - and widens the gap.

What matters most now is having sharper ideas, stronger judgment, better taste, and relentless execution - capabilities that sit deep in human practice.

For me, AI’s biggest gift is turning ideas into reality - fast. It lets me build exactly the skills, tools, and systems I need - tailored, precise, and immediate.

And because I open-source them, I’ve met many wonderful new friends.

Over these years, AI has fostered a genuinely warm open-source culture. People share experiences, insights, and code to give, learn, inspire, and connect.

10 AI Coding Projects

Last night, I watched the “WaytoAGI Future Silicon World” livestream. Twelve speakers presented ten AI coding projects plus two FDE (Frontline Digital Enablement) case studies - many were genuinely interesting and highly inspiring.

1. Xiaoer FM and Omia, by Xiaoer: A music dynamic visualization tool and a minimalist reader supporting over 100 formats - both products feel polished and intentional.
2. Watchless, by Chen Zixin: A skill that converts long videos and podcasts into illustrated summaries with key frames.
3. Parent-Child Habit-Building Mini-Program, by “Fafen de Caomei” (Frustrated Strawberry): Uses points, “mythical beast” hatching, and wish redemption to motivate kids to complete daily check-ins - grounded in a very real, relatable parenting context.
4. AI Private-Operation Tool, by Lao Yuan: Automates customer relationship management across personal WeChat and WeCom - identifies purchase intent and routes leads to sales reps. The commercial loop is already live and working.
5. Whiteboard Explainer Video Tool, by Yu Xianghua: Turns text, documents, or web pages into narrated whiteboard-style explainer videos with subtitles - ideal for knowledge dissemination.
6. AI Parent-Child Language Flashcards, by Nerissa: Snap a photo of any object -> instantly generate foreign-language vocabulary, short phrases, pronunciation audio, and printable flashcards - designed with elegant restraint.
7. CrewClaw, by Peng Chao: A platform for hiring and managing digital employees - structured like an App Store for AI agents.
8. 3D Director’s Desk, by Meng Maixiang: Lets you control a 3D camera using your phone’s gyroscope - enabling fast, intuitive shot composition optimized for AI video generation. Fun and clever.
9. WeChat Official Account Editor and Perler Bead Maker, by Shu Tongwen: One solves common image and layout pain points for WeChat content; the other auto-generates perler bead patterns from uploaded images - both born directly from the creator’s own workflow needs.
10. AI Content Studio, by Ally: Automates end-to-end video production - from material selection and scriptwriting to captioning, editing, and transitions - practical and time-saving for creators.

Two FDE case studies:

1. Smart Product Detail Page FDE for Footwear & Apparel, by Lyon: Embedded on-site at Skechers, this project mapped their operational workflows and used AI to automate product recognition, template matching, and bulk layout application - lifting accuracy from 85% to 95%. Later reused successfully at Anta.
2. Enterprise AI FDE Delivery Case Study, by Derrick: Drawing on dozens of enterprise deployments, he shared hard-won insights on need-finding, on-premise deployment, data security, and scaling AI solutions across state-owned enterprises, hospitals, law firms, and more.

Some related links for the AI coding projects:

1. Xiaoer FM: fm.xiaoercamera.xyz
2. Omia: xiaoeromia.com
3. Watchless: watchless
4. Zhixingyuan Mental OS: zhixing.club
5. Explainer Video: explainer-video
6. SpeedPainter official site: speedpainter.org
7. CrewClaw: CrewClaw
8. OpenFDE: open-fde.com

Small Facts, Big Conclusions

Many of our decisions and judgments hinge on how we interpret facts.

People rarely respond directly to raw facts. More often, they respond to their own interpretation of those facts.

That’s why it’s easy to fall into the “small facts, big conclusions” trap - even drawing conclusions from thin air.

Inside our heads, multiple “governments” are constantly at work.

Long before language kicks in, these internal systems have already debated, negotiated, and competed - multiple rounds deep.

Language, then, acts more like a spokesperson: it justifies the conclusion that has already won.

What truly warrants caution is jumping too quickly - or too far - to an explanation.

Interpretation drives action. Action produces outcomes. And those outcomes are then selectively interpreted - often reinforcing the original interpretation.

Before long, we grow increasingly convinced by the “illusion” we ourselves constructed.

So how can we avoid falling into the “small facts, big conclusions” trap?

A few practical strategies:

1. For any single fact, deliberately generate at least three plausible alternative explanations.
2. Actively seek out evidence that could disprove your current judgment.
3. Zoom out: ask whether the observation holds across a larger sample size and over a longer time horizon - is it a fluke, or a stable pattern?

How to Identify the Most Important Things

We all know our time and energy are limited.

Spending your best energy on truly important things - versus scattering it across many minor ones - leads to vastly different outcomes.

So a critical personal strategy is to reserve your peak energy and sharpest focus each day for what really matters.

My current criteria for identifying what really matters:

1. Highest leverage: Doing it well triggers cascading positive effects across many other areas.
2. Long-term compounding: It builds durable value over time - skills, systems, relationships, or knowledge that keep growing in usefulness.
3. Requires your direct agency: You must lead it, meaningfully participate in it, or make a decisive call.
4. Unblocks the core bottleneck: It directly resolves the single biggest constraint holding back progress right now.
5. Sustained personal commitment: You’re willing to invest in it long-term - and it’s an area where you can realistically develop a distinctive edge.

The most important things often yield little feedback in the short term while quietly reshaping your trajectory over months and years.

The compound effect of personal productivity hinges on one simple question: Where do your best few hours of the day actually go?

The Next-Generation Company

A friend and I were chatting recently about what future companies might look like.

I said: I’ve long hoped to build - and see succeed - a company with this profile: a core team of fewer than 10 people, generating over RMB 100 million in annual revenue.

Tiny in size, yet large in impact. Every team member earns well above market-average compensation, enjoys significantly greater autonomy, and experiences higher personal fulfillment.

Such companies are already beginning to emerge - both here and abroad - with early prototypes and real-world examples.

These teams start by deeply integrating AI and autonomous agents - offloading repetitive execution, communication, analysis, and management tasks to intelligent systems.

Using AI well is only the starting point.

Their business must be highly standardized, productized, and scalable, allowing revenue growth to separate from headcount expansion.

They must also solve problems that matter deeply to customers - problems customers are willing to pay a premium for.

Because the team is so small, income has to come from high-value work and dense value creation.

What matters more is talent density, product strength, and value density: growing through sharper insight, tighter execution, and higher-margin outcomes.

Internally, such a company may have only 10 people. Externally, it connects fluidly with partners, open-source communities, suppliers, distribution channels, and thousands of users.

Each core member operates more like a “super individual”: empowered by AI to own and deliver an entire end-to-end business outcome across the full chain.

So the most meaningful metrics for the next-generation company go beyond traditional KPIs. They include revenue per person, profit per person, value created per unit of time - and, crucially, how much of that value flows back to each creator.

Do more with less. And distribute the rewards more fairly among those who create them.

That’s the company model I find most compelling in the AI era - and the one I continue to think about, test, and build toward.

Pinpointing Your Positioning Skill

I built a positioning skill - a tool designed to help you better understand your personal or organizational competitive advantages, and thereby improve the quality of your strategic decision-making.

It also helps clarify key questions like:

1. What single sentence would your target audience use to describe you?
2. Who do they currently see as their alternative?
3. Which difference actually influences their choice?
4. Which advantage do you already have solid evidence for - and can sustain over time?
5. What’s the next step to test it, articulate it clearly, and own that position long-term?

It synthesizes core ideas from Positioning - mental models, categories, ladders, and focus - with practical additions: competitor analysis, deep demand insight, evidence grading, and D6 Advantage Diagnosis.

It works for personal brands, courses, products, services, and company branding. It ingests Q&A transcripts, attachments, and website content - and outputs:

• A crisp one-sentence positioning statement
• Differentiating tags
• Actionable recommendations
• Validation criteria
• A visual HTML report

GitHub repo: yao-positioning-skill How it works (principles & workflow): doc.laoyao.cn

With AI now embedded in everyday life, user decision-making has fundamentally shifted.

A potential customer might first encounter a brand through a short video, outdoor ad, media coverage, or a friend’s recommendation - but before purchasing, they often turn to AI platforms like Doubao or DeepSeek to ask questions: “How is XXX brand?” “Is XXX brand trustworthy?” “Is XXX brand right for me?”

AI has moved into the most decisive link of the user decision chain. Advertising, content, and word-of-mouth get users to know a brand. AI increasingly handles information synthesis, comparative analysis, and decision support.

For brands, neglecting this step - or failing to systematically optimize for AI platforms - means losing high-intent prospects at the final moment. That’s costly.

Here’s the problem: when users query AI about a brand, the answers are often inaccurate. AI may misrepresent the brand’s positioning, product capabilities, target audience, or core strengths - offering outdated, contradictory, or even damaging descriptions.

A user grows interested - then abandons the choice during that final AI-powered verification step, misled by error, confusion, or stale data. That’s not just inefficient - it’s self-sabotage.

Why does this happen? Three main reasons stand out:

1. No unified internal brand “fact standard” Many companies lack a centralized, rigorously maintained set of verified brand facts. Without one, inconsistent messaging proliferates:
   • Channel A promotes one brand narrative; Channel B uses an entirely different framing.
   • The official website describes product capabilities differently than sales teams do in the field. These contradictions confuse customers - and leave AI with no reliable signal to prioritize. Which version is authoritative? Which reflects current reality?

2. No AI-optimized content infrastructure AI forms its understanding by scraping and synthesizing from official websites, owned media, industry reports, and third-party reviews. If a brand fails to publish clear, structured, verifiable content consistently - AI pulls from outdated pages, fragmented snippets, or incomplete descriptions. The result? AI knows the brand exists - but doesn’t understand it. Some “hallucinations” stem directly from this gap.

3. A polluted third-party content ecosystem As GEO (Generative Engine Optimization) gains traction, more review articles, rankings, head-to-head comparisons, and recommendation lists flood the web. Too many contain weak evidence, competitor-driven bias, commercial placements, stitched-together claims, or outright factual errors. When AI ingests and summarizes these, inaccuracies enter its knowledge base - distorting perception. A brand with real advantages gets downgraded due to flawed third-party framing; fabricated weaknesses get repeated until they feel true.

All three issues erode AI’s ability to accurately represent a brand’s essence, strengths, and service promise. And when users rely on AI to verify and finalize decisions, those distortions translate directly into lost customers, eroded trust, and measurable revenue impact.

So - how should companies strengthen brand consistency?

First, treat it as a strategic priority - not a marketing tactic. Establish an internal “brand fact standard”: break the brand down into atomic, verifiable statements - e.g.,

• One-sentence positioning
• Target customer profile
• Core products/services offered
• Primary problems solved
• Differentiating capabilities
• Competitive distinctions
• Real-world case studies and metrics

For each fact, define one clear, accurate, up-to-date, and source-verified answer - and designate it the single source of truth.

Then, use technical tools or specialized providers to audit brand consistency quantitatively:

• How do major AI platforms currently describe the brand?
• Do their answers align - or contradict each other?
• Where do factual errors, omissions, or outdated claims appear?

Once diagnosed, remediate systematically: publish clean, structured, authoritative content; manage digital assets proactively; prioritize high-trust sources in outreach and syndication.

The ultimate goal? No matter where or how a user discovers the brand - via search, social, sales call, or AI chat - they receive information that is accurate, stable, coherent, and mutually reinforcing.

For the business, brand consistency delivers four core benefits:

1. Lower cognitive load for users
2. Higher perceived credibility
3. More accurate AI understanding and recommendation
4. Durable, compoundable brand equity

At its heart, brand consistency solves one question: Will every customer - regardless of entry point or AI platform - see the same accurate, trustworthy, clearly articulated version of who we are and what we deliver?

Media’s GEO Potential

I recently gave a GEO briefing to Beijing Daily Group. Their response was sharp - and their execution speed impressive. Among China’s authoritative traditional media, their awareness and strategic positioning are notably advanced.

Their greatest asset? Vast volumes of high-quality, authoritative content. Used well with GEO, that content unlocks multiple layers of value:

1. When AI cites authoritative media more reliably, answer quality improves - and hallucination rates drop.
2. This creates a “good coin drives out bad coin” effect: better content rises in AI’s reference hierarchy, pushing low-quality noise downward.
3. For the media itself, commercially valuable content gains broader, more targeted exposure - reaching new audiences via AI answers. That’s new monetization potential.
4. The competitive landscape shifts: in the AI era, authority isn’t just about reach - it’s about becoming the default factual source and interpretive framework behind AI-generated answers.

GEO’s value for media goes beyond visibility. It operates on two deeper levels:

• First, shaping how AI answers - embedding the media’s facts, perspectives, and narrative logic into responses.
• Second, building sectoral knowledge infrastructure - establishing the outlet as the go-to authoritative source in specific domains.

One caveat: volume of citation != commercial return. Media must proactively test business models - before scale - to ensure citations convert into traffic, subscriptions, licensing, or other sustainable revenue. The opportunity remains wide open.

How to Say “Thank You”

I spoke with our sales lead about a user message: “I plan to monetize GEOFlow - and will share 5% of resulting revenue.”

I replied: “No need.” But inwardly - I was genuinely pleased.

My colleague pushed back gently: “That phrasing isn’t ideal.” “A stronger way to express gratitude would be to wait - then, once you’ve actually earned RMB 1 million, send RMB 30,000 with a note explaining why.”

That act carries more weight. It reflects respect - not obligation.

Prematurely promising 5% signals goodwill, yes - but also quietly establishes a long-term transactional relationship. Waiting until results materialize - and then acting voluntarily - keeps the gesture clean, grounded in outcome, and emotionally resonant.

The distinction lies at the core:

• Transaction requires pre-agreed rights, responsibilities, and returns.
• Gratitude is a post-hoc, voluntary return - given freely, based on recognition and respect.

The most meaningful thanks aren’t pledged at the start - they’re delivered after the result lands.

Managing Complex Systems

In complex worlds, perfect answers rarely emerge upfront. A more realistic approach is: act, observe feedback, adjust - and repeat.

Cybernetics and the Scientific Methodology, by Jin Guantao and Hua Guofan, offers a practical framework for this. Using cybernetics, information theory, and systems thinking, it tackles one central question: How can humans effectively understand, intervene in, and steer complex systems - even when full knowledge is impossible?

1. The Core Feedback Loop

Compress the entire book into one cycle: Define goal -> Observe state -> Gather information -> Take action -> Receive feedback -> Adjust action

Example: Running a SaaS company. Goal: Increase renewal rate -> Observe current user behavior -> Collect usage stats, NPS, churn reasons, support tickets -> Revise onboarding and support workflows -> Measure renewal rate change -> Refine further.

This is a control system - not about rigid command, but about guiding outcomes amid uncertainty.

2. Two Kinds of Feedback

• Negative feedback corrects deviation and stabilizes: e.g., project falling behind -> team adds resources -> pace recovers. Logic: Deviation detected -> counter-action applied -> system returns toward target.
• Positive feedback amplifies trends: e.g., more users -> more content -> more users. It fuels growth loops - or vicious cycles. Long-term stability hinges on having timely, truthful, actionable feedback mechanisms.

3. Information Is the Foundation

Without information, a system cannot know:

• Its current state
• Distance from goal
• Which actions work - or backfire

Information value isn’t just about volume - it’s about reducing uncertainty, revealing differences, and enabling action.

A powerful insight: An organization is, fundamentally, an information structure. People sharing an office != an effective organization. Only when information flows accurately, gets processed meaningfully, is stored reliably, and feeds back authentically does coordination emerge.

This reframes common management failures:

• Silos = broken information channels
• Distorted reporting = corrupted feedback signals
• Repeated mistakes = absent organizational memory
• Slow decisions = insufficient information processing capacity
• Micromanagement = information bottleneck at the top

4. Stability, Change, and Collapse

Why do systems stay stable? Because they maintain homeostatic structures: body temperature, company operations, ecosystems - all absorb shocks and self-correct. Stability != stasis. A company changes daily - yet endures because core functions hold.

Why do systems oscillate? Feedback delays cause overshoot and correction: Demand rises -> factory expands capacity -> supply floods market -> demand drops -> capacity shrinks -> shortage returns. Economic cycles, inventory swings, and org restructuring all follow this pattern.

Why do systems collapse? Self-reinforcing pathologies take root: bureaucracy rewarded -> copied -> normalized -> entrenched. Low-quality content, fake metrics, internal politics - all replicate when they yield short-term advantage.

How do systems self-organize? Order emerges not only from top-down design - but from local rules + interaction: markets, languages, open-source communities, online forums.

5. When Gradual Change Becomes Sudden Shift

Why do some systems appear static - then abruptly transform? Water heats slowly - then boils. A startup burns cash steadily - then hits liquidity crisis.

The book frames this with four concepts: gradual change, leap, stable state, and tipping point. Imagine a ball resting in a valley:

• Deep valley = stable system
• Shallower valley = increasing instability
• Valley vanishes = tipping point reached
• Ball rolls into new valley = qualitative shift

This reminds us: watch not just how fast things change - but whether the structures holding the current state are weakening.

6. Black-Box Epistemology

Some systems - brains, large organizations, recommendation engines, LLMs, customer decisions - are too complex to map fully. We see inputs and outputs, but not internals.

To study them, use this method:

1. Apply varied inputs
2. Record outputs
3. Identify stable input-output relationships
4. Build a predictive model
5. Test with new inputs
6. Refine model based on feedback

This mirrors how we study LLMs today: we can’t trace all 100B+ parameters - but through benchmarking, prompt experiments, red-teaming, real-world task evaluation, and live feedback, we map capability boundaries.

Understanding the world means continuously building - and correcting - models. Most failures stem from broken feedback. So managing complexity - or building complex products - starts with designing robust feedback loops.

Training Prediction Skill

Practicing prediction is deeply valuable. It sharpens judgment, exposes cognitive bias, calibrates confidence, and improves decision-making under uncertainty.

Yet we often say “I think it might…” or “It’ll probably happen” - without assigning probabilities or later checking our accuracy.

Drawing from earlier discussions of Bayesian skill, here’s a minimalist 5-step prediction training method: Frame -> Assign -> Gather -> Update -> Verify

One full cycle requires only:

• A testable question
• An initial probability estimate
• Three key pieces of evidence
• One probability update
• One post-result debrief

Assign probability before reviewing evidence. Update iteratively. Then verify.

Repeat - and over time, you train your brain to navigate ambiguity with greater precision and humility.

Means vs. Ends

I used to tell my team: “Master AI.” Because AI is becoming foundational - teams fluent in it gain real efficiency and competitive edge.

But I noticed a subtle risk: people began treating using more AI tools or building more automation as goals in themselves.

So someone asked: “Is mastering AI more important than serving customers or delivering business results?”

That’s a false dichotomy. AI is a means - a capability, method, tool. Solving real problems, creating value, and serving customers are the ends.

AI’s value isn’t in its complexity - it’s in what it enables:

• Does it improve speed, accuracy, or quality?
• Does it lift customer experience - or cut costs, boost revenue, unlock previously impossible tasks?

If an AI system is technically impressive but changes nothing in real workflows, its sophistication is irrelevant.

This reminds me: always clarify the goal first - then discuss tools. Leadership emphasis shapes behavior. Prioritize tool count, usage frequency, or automation depth - and teams build elegant systems that don’t move the needle.

Of course, AI’s value isn’t only about next-quarter revenue:

• Short term: Does it lift efficiency, quality, or performance?
• Medium term: Does it produce reusable processes, products, or capabilities?
• Long term: Does it forge new, defensible advantages?

Mature teams rarely debate “AI vs. business.” They ask:

• What level of business impact can AI help us achieve?
• What was impossible before - and now possible?
• Can we deliver higher-quality outcomes with fewer people?
• Can customers get faster, more reliable, more measurable results?
• Is there a simpler, more direct solution - without AI?

AI defines how fast and well we run. Purpose defines where we’re running. Results determine whether the run mattered at all.

For B2B companies, a critical sales strategy isn’t just building a reference case - it’s going all-in to build a legendary one.

In China, the strategic weight of real-world cases is vastly underestimated - especially for new products. When target enterprises struggle to grasp unfamiliar concepts, models, or jargon, the cost of education and alignment is high. But tell them a legendary case - a vivid, outcome-driven story - and understanding clicks instantly. Decisions follow just as fast.

What makes a case “legendary”? It triggers that visceral aha moment.

For example:

• A previously unknown startup grows annual revenue from RMB 5 million to RMB 100 million in one year - using your product.
• A listed company doubles core channel acquisition volume while cutting customer acquisition cost by 50% - by adopting your model.

One such case outweighs 100 pages of feature specs. With just one or two, your sales team gains momentum; prospects gain clarity and confidence.

Here’s how to build one - five steps:

1. Select the right client: High potential, high collaboration, willing to go public.
2. Define a sharp, measurable goal.
3. Deploy core team + resources personally - no delegation, no half-measures.
4. Document everything, end-to-end.
5. Productize the case: Turn it into reusable assets - narrative, metrics, visuals, talking points.

That’s what we call an Aha Case.

The Boundary of Product

I recently hosted a channel partner who shared a prospective product collaboration. Its pitch? “We serve any client - and guarantee results.”

That logic is flawed. Every product has boundaries: what it can do, what it cannot, who it serves well, and who it doesn’t serve at all.

At our company, we’ve internalized this principle - and enforced it rigorously:

• We don’t take projects below a minimum price threshold.
• We decline all bid-tender processes.
• We avoid high-sensitivity or gray-area industries.
• We walk away when core requirements fall outside our capability scope.

These boundaries aren’t constraints - they’re quality safeguards. They directly shape delivery integrity and, over time, reputation.

One of my biggest takeaways from the past year of entrepreneurship? Riding trends matters - but so does differentiation. Even within a hot trend, standing out requires deliberate choices. And true differentiation isn’t built on features alone. It’s forged through consistent boundary-setting: saying “no” clearly, repeatedly, and with conviction. Each refusal sharpens your positioning - and deepens user trust.

Good Relationships

A friend shared his reflections on relationships - particularly romantic ones.

A good relationship leaves lasting value - even if it doesn’t last. Years later, you don’t recall resentment. You remember warmth: being understood, cherished, held gently.

Initial attraction - looks, charm - is just the entry ticket. What sustains connection is mutual admiration, deep trust, and safety. That safety lets people relax, drop pretense, and show up authentically.

I asked: What stays with you? He listed subtle but consequential things - life perspectives, practical habits, even small skills - that reshaped his daily thinking and choices.

He observed that many people in China live externally - performing for “face” or defaulting to self-sacrificing “saintly” behavior - making it hard to ask for what they need or claim what’s rightfully theirs.

Examples he gave:

1. Seek diverse experiences: Reading about Kyoto isn’t the same as walking its alleys at dawn. Immersion changes perception.
2. Taste widely - but lightly: Savor variety without overindulgence. One bite can be revelation.
3. Speak up at fine dining: If a dish misses the mark, describe exactly why - not “it’s okay.” Chefs often replace it immediately.
4. Use hotel service proactively: Half the price of luxury hotels is service premium. Call front desk before frustration builds - you’ll likely get resolution in under 90 seconds.
5. Pay for expertise: A travel consultant or dedicated driver costs little but transforms itinerary quality. Don’t DIY complexity you don’t enjoy.
6. Mine your financial perks: Gold card benefits (golf access, lounge entry) hide in app menus. Spend 3 minutes finding them - it pays back instantly.
7. Guard what enters your body: In unfamiliar places, never lose sight of water, drinks, or food - especially if unsealed.
8. Pre-read before cultural travel: Knowing the history behind Angkor Wat or the Forbidden City turns stone into story.

Good relationships don’t just soothe - they expand your life radius. They quietly transfer lived wisdom: judgment frameworks, boundary instincts, aesthetic sensibilities, even how to hold space for uncertainty.

Of course, relationships aren’t all light. Having experienced depth, you also learn discernment: who deserves closeness - and who doesn’t.

Some red flags stand out:

• For men: chronic gamblers, emotionally stunted adults, or those with volatile moods.
• For women: partners with chaotic relationship histories, extreme materialism, or who consistently drain their current family to subsidize their original one.

All these point to the same core deficits: responsibility, boundary awareness, emotional stability, and basic goodwill.

Choosing a partner isn’t about initial spark - it’s about observing how they show up over time: Do they own mistakes? Respect limits? Recover gracefully? And - crucially - do they help you grow without erasing you?

But the most important question isn’t “Who should I choose?” It’s: Am I becoming someone worthy of being chosen by someone excellent?

Four Human Traits Most Scarce in the AI Era

Hudson, one of Silicon Valley’s most sought-after executive coaches, now spends most of his time advising OpenAI’s research teams. His insider view reveals something counterintuitive: success in AI-augmented environments depends less on knowledge or grit - and more on emotional clarity.

As AI rapidly commoditizes information and effort, four distinctly human traits are becoming increasingly rare - and valuable:

1. Discernment: Knowing what to pursue, what to abandon, and when to pivot - not just what to do, but whether to do it at all.
2. Trustful conflict: In high-leverage small teams, unresolved tension explodes. The ability to initiate and navigate tough conversations - without defensiveness or avoidance - is now mission-critical.
3. Willingness to fail: Humans excel at rapid iteration. Here, “failure” isn’t catastrophe - it’s negative feedback, a data point for learning. Courage to ship, test, and revise fast is irreplaceable.
4. Positive self-dialogue: Self-flagellation kills creativity and judgment. The capacity to speak to yourself with curiosity, kindness, and agency - not shame - is the bedrock of sustained creation.

AI Interview System

Our hiring volume keeps rising. So we upgraded our “AI Interview Manager” - not just adding features, but rearchitecting for four-way synergy: AI Interview System + GitHub + Feishu + Codex.

1. AI Interview Manager (core) The operational hub: manages candidates, resumes, AI-first interviews, auto-generated reports, feedback loops, scheduling, analytics, and cross-system status sync.

2. GitHub Repository (collaboration layer) Each candidate gets a dedicated directory. Workflow stages auto-record via folder structure, workflow.json, candidate-index.csv, and Git commits - creating a versioned, auditable evidence chain.

3. Codex Skill (local intelligence layer) Interviewers pull candidate packages locally, generate evaluations, write first/second-round feedback, save schedules, and push tasks/feedback/reports to Feishu - all from one interface.

4. Feishu Execution Layer Via Feishu CLI: auto-create calendar invites, send private reminders, trigger post-interview nudges (e.g., “Please submit feedback within 1 hour”). Background workers handle async retries, status updates, and weekly report syncs.

The flow:

• Resume uploaded -> parsed asynchronously.
• Candidate completes AI-first interview -> Q&A, transcripts, keywords, and scores saved.
• Report generator produces HR & candidate versions -> auto-synced to GitHub.
• Team views candidate data in GitHub or Codex -> generates evaluation -> writes feedback.
• If “pass”, candidate moves to next-stage GitHub folder; if “reject” or “hold”, only status/feedback updates.
• Once scheduled, Feishu calendar task created -> worker triggers 1h & 3h feedback reminders post-interview.

For interviewers, Codex becomes the primary interface - all other systems coordinate silently in the background.

What makes this architecture compelling:

1. Business data lives in the database; collaboration provenance lives in GitHub.
2. Hiring is GitOps-native: every stage shift, feedback update, or folder move is traceable.
3. Local agents pull data on demand - enabling smart collaboration without exposing credentials or raw data.
4. Feishu tasks run async - calendar, notifications, and reminders are fault-isolated.
5. AI handles fact extraction and suggestions; humans own judgment, reasoning, and accountability.
6. Candidates transform from database rows into stateful, versioned, portable objects - carrying history, context, and touchpoints across systems.

In short: candidates are now modeled as living entities - flowing through AI assessment, GitHub collaboration, Codex evaluation, and Feishu engagement. This gives hiring four superpowers: automation speed, human accountability, organizational memory, and full-stack operability.

DouBao Diagnostic Skill

I built a GEO-focused skill to collect, clean, and analyze outputs from DouBao (ByteDance’s AI assistant). Designed for personal research - not production use.

It supports two collection modes:

• Web interface: via OpenCLI.
• Mobile app: via Android Studio AVD + Appium UiAutomator2.

Same keyword queries yield both web results and mobile UI evidence - screenshots, XML DOM trees, citation cards, and precise tracking of “cited vs. uncited” sources and citation counts.

Output is unified:

• doubao-crawl.json (raw data)
• summary.json (aggregated insights)
• Structured Markdown
• Excel tables
• Kami-style HTML reports

Downstream GEO analysis uses one shared template - regardless of source.

Skill boundaries:

• No login bypasses.
• No CAPTCHA cracking.
• No hidden API scraping.
• No account pooling.

Best suited for low-frequency research, teaching demos, or forensic validation where screenshots + XML are required.

If you study GEO and want to compare how DouBao’s mobile app surfaces citations versus its web version - try it.

GitHub repo: github.com

This was my second GEO public session with Professor Xiangyang on WaytoAGI. The next session is scheduled for July 25 at 8 p.m., focusing on GEO source strategy.

The theme of this session was GEO Content Engineering - a topic that’s notoriously hard to teach, yet essential to cover systematically. Doing so helps participants understand why GEO demands higher standards from content, and why its management and operational logic differs fundamentally from traditional marketing approaches.

In my view, only when you internalize this layer of understanding - and pair it with hands-on capability - do you truly possess precision GEO operations.

First, recognize that AI’s external citation strategy is far more nuanced than most assume. For the same question, an AI might list a dozen sources - but only two or three actually shape the final answer. Some links appear in the reference list but are never used in the response; others get cited repeatedly across a single answer, becoming structural anchors and recurring evidence.

So why does AI cite just one article out of hundreds published? If GEO is simply about prompting AI to write articles, why do some AI-generated pieces get cited while others sit untouched? If GEO is just keyword coverage, why does AI ignore keywords that appear dozens of times?

The real issue isn’t “Did we publish enough?” - it’s “Does AI treat your content as reliable evidence?”

Many teams default to legacy content tactics: publish more articles, distribute across more channels, stuff more keywords, generate multiple AI drafts. But none of that explains the core phenomenon: Why does AI cite A and ignore B?

That’s exactly what the GEO Content Engineering framework aims to answer.

Live session summary (in Chinese): mp.weixin.qq.com

Here’s a distilled set of key insights and methods:

1. GEO’s essence is earning AI’s trust - getting it to call upon, reuse, and cite your content repeatedly.
2. An AI citation looks like a search result - but underneath lies an evidence-scoring mechanism.
3. “Content quality” is too vague. What’s operable are relevance, authority, extractability, and verifiable evidentiary value - leading to concrete, measurable content features.
4. Today’s AI most urgently needs ready-to-use evidence blocks.
5. Whether an article gets cited depends not on visibility - but on whether it can become part of the answer.
6. Traditional content ops optimize for human reading; GEO content engineering optimizes for machine comprehension, judgment, and assembly.
7. AI prefers content that reads like academic papers, encyclopedias, or research reports - not emotionally charged marketing copy.
8. Publishing 100 low-evidence articles delivers less value than crafting 10 high-evidence-density assets.
9. High-citation articles tend to feature clear evidence, stable structure, and easily extractable conclusions.
10. Keyword era -> density matters. Semantic era -> relevance rules. GEO era -> evidence combinations decide.
11. GEO doesn’t optimize rankings - it optimizes probability: the likelihood your brand gets recommended, cited, and trusted.
12. At its core, content engineering transforms chaotic content creation into a decomposable, reusable, and attributable system.
13. No problem map -> no GEO strategy. No knowledge base -> no high-quality content.
14. Real GEO competence isn’t about writing good prompts - it’s about closing the loop between questions, evidence, structure, sources, and feedback.
15. Prompts define form; knowledge bases define substance.
16. Identical prompts won’t homogenize output - thin knowledge bases will.
17. Content engineering isn’t about single articles. It’s a system defined by goals, elements, and their relationships.
18. A key GEO engineering mindset: translate leadership’s fuzzy aspirations into system-level requirements - executable, measurable, and improvable.
19. GEO’s biggest risk isn’t low traffic - it’s doing many things without knowing which ones actually moved the needle.
20. Future content teams will compete on evidence assets, structural fluency, and feedback-driven iteration.

Related resources:

1. GEO Content Engineering Handbook & Evaluation Criteria doc.laoyao.cn
2. GEO Content Engineering Systems Report doc.laoyao.cn
3. GEO Content Engineering Methodology & Single-Article Tutorial doc.laoyao.cn
4. Thinking in Systems, Everyone Should Understand Engineering
5. GEO: Generative Engine Optimization doc.laoyao.cn
6. Generative Engine Optimization in Digital Repositories: Optimizing Visibility for Generative AI doc.laoyao.cn
7. A Measurement Framework for Generative Engine Optimization Across AI Search Platforms doc.laoyao.cn

Tools & utilities:

1. GEO rewriting prompt: ai.laoyao.cn
2. GEO rewriting Skill: ai.laoyao.cn
3. GEO feature annotation demo (single-article): doc.laoyao.cn

GEO systems & Skills:

1. GEO Skills: github.com
2. GEOFlow: github.com
3. Meta Skill: github.com

Defining What Not To Do

This year’s entrepreneurial journey crystallized around three words: embracing trends, differentiation, and defining boundaries.

Embracing trends matters - under favorable conditions, sales, operations, and growth all become simpler.

Yet even amid simplicity, differentiation remains vital. Building a product with genuine differentiation may slow early growth - but it directs finite resources toward reinforcing that edge, forging sustainable competitive advantage. That advantage becomes indispensable once the market turns crowded.

That’s why defining - and guarding - boundaries is critical.

Trends reveal many opportunities - but most aren’t yours. Only the direction you can master, choose to focus on, and iterate on relentlessly qualifies as your opportunity.

So the core discipline is clarifying what not to do - and protecting it rigorously.

There’s a simple truth behind this: Don’t do what you don’t understand. Instead, allocate resources to trusted partners - observe, learn, and absorb. In other words: Dare to be second.

A “not-to-do” list safeguards what’s most precious - and easiest to overlook: your attention and energy.

Only within that protected space can your differentiated advantage truly take root.

Boundary-setting applies not just to companies - but also to products and services. Clarity comes from answering two questions: What can we do? What must we not do?

“What we can do” functions like social ethics - encouraging kindness and integrity. “What we must not do” functions like law - drawing hard lines. Together, they form an unmistakable operating perimeter.

What Makes a Good Salesperson

A new sales hire recently led a practice session in the afternoon. Afterwards, the team discussed various sales techniques: plain language, real-world examples, the 80/20 rule, audience awareness, storytelling, etc.

But I kept thinking: great sales rests on a simpler foundation.

Two words capture it: authenticity and expertise.

Authenticity means transparency - no hidden agendas, just genuine care for the client’s success. Expertise means diagnosing real problems and offering sound, actionable guidance.

That’s enough.

It echoes this truth: Sales isn’t about lowering your status to please customers - it’s about advising them like a friend. You happen to need it. I happen to know it. That’s all.

Interview Pitfalls

I’ve conducted many interviews lately.

During a debrief with a hiring manager, I heard this evaluation of a candidate: “Their experience is okay.”

That phrase - “experience is okay” - is dangerously ambiguous.

I asked: How did you reach that conclusion? They replied: Mainly based on years of work history and how well they answered questions.

Both criteria carry hidden traps.

First, tenure ≠ experience. Most people repeat basic tasks year after year - without deepening skill or insight. Second, articulate answers ≠ real experience. They may simply reflect strong communication - or packaging skills.

True experience, in my view, hinges on three things:

• Has the candidate handled genuinely complex problems?
• Have they built coherent, reusable methods?
• Have they delivered outcomes others couldn’t replicate?

The best way to assess this? Drill into real cases. Ask for specifics. Follow up on contradictions. Probe the “how” and “why.”

Truly experienced people explain complexity with striking simplicity - concrete details, clear boundaries, and resilience under scrutiny.

Four Mindsets for the Age of Abundance

Inspired by We Are Gods, this section outlines four essential mindsets for thriving in abundance: Abundance Mindset, Longevity Mindset, Exponential Mindset, and Moonshot Mindset.

On Abundance

1. Abundance has already arrived - yet many still operate from scarcity mental models.
2. Today’s true scarcity isn’t information - it’s attention.
3. Paradoxically, the more abundant an era becomes, the more anxiety ordinary people may feel.
4. The future is already here - just unevenly distributed. The greatest inequity from new tech isn’t access - it’s opportunity flowing first to those who understand it.
5. The gravest danger in abundance isn’t missing opportunity - it’s failing to see it. As technology advances, cognitive gaps widen rapidly.
6. Hence the four mindsets: Abundance, Longevity, Exponential, Moonshot.

Scarcity vs. Abundance Mindset

1. Scarcity mindset asks: How do we divide the cake? Abundance mindset asks: How do we bake a bigger one? Abundance thinking creates net new value.
2. Much anxiety stems from applying old scarcity logic to new abundance opportunities. When you obsess over defending your slice, someone else is baking a whole new cake elsewhere.
3. Abundance mindset reframes competition as co-creation. Winners in abundance-era markets rarely own the most resources - they’re best at recombining them.

Longevity Mindset

1. Those who’ll benefit most from longevity tech are the ones managing their health today.
2. Longevity mindset redefines time itself. AI extending human lifespan is no longer speculative - it’s widely accepted.
3. With decades added to life expectancy, career planning and health habits must shift - and lifelong learning becomes non-negotiable.

Exponential Mindset

1. The most counterintuitive thing about exponential change? Early stages look flat. Without this mindset, people underestimate early, misunderstand mid-stage, and miss late-stage entirely.
2. The first half of exponential growth often resembles stagnation; the second half feels miraculous. Real opportunity hides in the deceptively slow early curve.
3. Judging tech trends requires looking beyond current capability - to iteration speed. When a technology’s performance doubles consistently, conservatism becomes the riskiest stance.

Moonshot Mindset

1. Moonshot mindset sets audacious goals within physical reality - not fantasy, but impossible goals broken down into executable steps.
2. Big goals force method innovation: 10% improvement refines old systems; 10× improvement demands rebuilding them.
3. Small targets invite incremental tweaks; massive ones compel rethinking fundamentals.
4. Moonshot thinking starts with first-principles analysis - mapping constraints using physics and logic - to draw realistic boundaries for ambition.
5. Any goal not violating fundamental laws shouldn’t be dismissed by inertia alone.

This Week’s Open-Source Updates

1. TokKit

Upgraded our local token ledger tool, TokKit, with a macOS menu bar panel.

Click the icon in your top-right corner to instantly see:

• Today’s token usage
• 7-day and 30-day trends
• Detailed usage breakdown (last 7 days)
• Terminal and model-specific consumption

GitHub: ai.laoyao.cn

2. ChatGPT Crawling & Diagnostic Skill

New GEO Skill: yao-chatgpt-crawler - for collecting, cleaning, and analyzing ChatGPT outputs.

Real-world test case: Doubao (ByteDance’s AI assistant), assessed via ChatGPT’s GEO diagnostic report.

Workflow:

1. Prepare question list, repetition count, target entity (e.g., “Doubao”), entity type, OpenCLI profile, and sampling interval.
2. OpenCLI Browser Bridge connects to your logged-in ChatGPT Web session, running randomized, spaced queries.
3. Post-collection, data is cleaned and aggregated into JSON, Markdown, Excel, and HTML reports.
4. The HTML report analyzes:
   • Entity mention rate & average mentions
   • Top-1 / Top-3 / Top-5 probability
   • Average ranking position
   • Sentiment bias
   • Source domains & URL structures
   • Title intent
   • Gap analysis vs. peer entities
5. Entity classification distinguishes people, companies, products, concepts, and noise - excluding industry terms, generic categories, or URL parameters from competitor lists.

Key principle: Based on real ChatGPT Web behavior - not API simulation - for higher-fidelity GEO analysis.

Links:

• Skill repo: github.com
• Example report files: github.com
• HTML report preview: doc.laoyao.cn

3. GEOFlow

GEOFlow - the end-to-end GEO system - has been updated to v2.1.0.

Most recent improvements came directly from a Fortune 500 marketing leader who deployed GEOFlow internally and shared rich, actionable feedback.

Key upgrades:

1. Enterprise Knowledge Base Drafting Workflow: Upload or paste multiple documents (PDF, DOCX, TXT, etc.), auto-parse formats, then use AI to generate structured draft content - with full source attribution, revision tracking, and manual approval gates.
2. Enhanced Frontend & Distribution Capabilities:
   • Online topic editing
   • Homepage module configuration
   • Custom styling options
   • AI-powered “APIHot” topic recommendations
   • Unified SEO metadata management
   • Visualized article distribution status
   • Task routing strategies
   • Target site sync settings
3. Deployment Improvements:
   • Installation state protection
   • Default admin email configuration
   • Redis-backed session storage
   • Docker support + subdirectory deployment

Post-update, GEOFlow increasingly functions as a full-stack GEO + automated site network management system - covering knowledge base ingestion, AI drafting, human review, frontend publishing, cross-site distribution, synchronization, and operational analytics.

GitHub: github.com

A recent insight has stuck with me:

When iterating on open-source projects, it’s alarmingly easy to fall into the complexity trap: even simple systems balloon rapidly under AI’s influence, and the list of “things we want to build” grows faster than our capacity to sustain them.

But that growth isn’t always progress. So I’ve been deliberately practicing subtraction.

The real benefit of subtraction isn’t just simplicity; it’s stronger continuity.

Continuity is often the quiet engine behind successful long-term projects. It enables consistent feedback loops, steady cognitive gains, and reliable iteration.

Here’s a rough formula:

Continuity = (Clarity of Direction × Consistency of Effort × Quality of Feedback) ÷ (Task-Switching Cost × Maintenance Burden × Choice Noise)

Of course, continuity comes with a trade-off: it often means dialing back curiosity about shiny new things and holding off on exploration to protect focus.

That restraint isn’t loss; it’s investment.

Because doing so delivers three concrete benefits:

1. Fewer task switches → preserved deep attention: When you shift from one task to another, unfinished mental threads linger and actively degrade performance on the next task.
2. Lower cognitive load → more stable learning: Working memory is finite. Clutter, irrelevant context, or poorly structured inputs drain capacity needed for reasoning and retention.
3. Fewer decisions → less decision fatigue: More options rarely mean better outcomes; they often mean slower, shallower, or avoidant choices. Reducing low-stakes decisions frees up mental bandwidth for what truly matters.

At its core, subtraction lowers a system’s ongoing cost.

Many goals stall not because they’re unrealistic, but because too many targets, commitments, inputs, and obligations fragment attention, dilute feedback, inflate maintenance, and raise the barrier to the next action.

Subtraction sounds simple. Its difficulty lies in execution: it demands ruthless prioritization without layering on new “additions” disguised as optimization.

Strategic sharpness shows up first in what you refuse to do.

Recognizing that is already half the win.

Reliability Can Be Trained

OpenAI’s new paper, Reinforcement Learning Towards Broadly and Persistently Beneficial Models, tackles a critical question: Can we train AI models to be reliably beneficial, not just in one narrow setting, but across diverse, unseen contexts?

Background: Many enterprise AI agents fail not from lack of capability, but from behavioral instability. They hallucinate sources, overstate confidence, bypass approvals, flatter users, misuse tools, or ignore boundaries. These failures may share a root cause: inconsistent behavior patterns.

The paper’s core hypothesis: harmful behaviors migrate, and beneficial ones can too, provided training targets the right underlying traits.

Prior work shows that harmful tendencies, like gaming reward signals, lying, generating unsafe code, or indulging user misconceptions, can generalize across tasks. If bad behavior transfers, can good behavior?

Yes, the paper finds evidence it can.

In controlled experiments, researchers injected a small set of high-quality beneficial-behavior examples (e.g., admitting uncertainty, citing sources, correcting errors, respecting constraints) into standard training data. Compared to a control model trained with identical compute and initialization, the augmented model showed significantly stronger stability across dozens of unseen benchmarks, including tasks and domains never present in training.

One standout result: Across 50+ independent evaluations, the beneficial-behavior-trained model outperformed the control on most metrics. Traits like honesty, caution, self-correction, and constraint adherence didn’t stay confined to their training contexts; they generalized.

What makes this paper especially compelling is how it operationalizes subjective ideals: “honesty,” “caution,” “human oversight,” “protecting vulnerable parties”-all become measurable via concrete prompts, dialogue flows, and scoring rubrics.

This resonates deeply with my work on content engineering. We say “high-quality content,” but what does that mean? “Structured”? Structured how? In practice, we can define precise, observable traits for even fuzzy terms.

The implications extend beyond models to teams and individuals.

Team culture is a reward system. What a company consistently rewards, it inevitably grows.

So if you want your team to embody “accountability,” don’t just name it; define what accountability looks like: Which actions count? What evidence qualifies? What metrics reflect it?

Paper link: alignment.openai.com

Updated Meta-Skill (v2.0)

I’ve just completed a full refactoring and 2.0 upgrade of my meta-skill framework, yao-meta-skill. The results are promising.

Credit where due: Codex keeps getting stronger. For complex engineering tasks, it now autonomously decomposes, commits, debugs, and iterates. This upgrade ran for 38 hours and produced 301 branches.

Meta-skill is the skill that generates skills. The v2.0 upgrade rests on five foundational principles:

1. “Is this worth a skill?” - Prioritize reuse value Only create a skill when a task repeats and produces a reusable output contract. One-offs should stay lightweight because over-engineering drains maintenance energy. Skills are assets; assets carry upkeep costs.

2. Model the real task before generating the skill First, capture: What’s the user actually trying to accomplish? What should the output look like? What must not happen? What constraints apply? Ambiguity here is the #1 cause of skill failure; every skill has boundaries.

3. Unify semantics with an intermediate representation v2.0 introduces a “skill intermediate representation”: a semantic contract specifying trigger conditions, inputs, outputs, boundaries, references, evidence, and expected artifacts. This layer decouples logic from platform and enables stable cross-platform skill packaging.

4. Guard quality with layered evaluation & review Quality is broken into checkable gates: Does it trigger accurately? Does output meet specs? Is runtime behavior consistent? Are permissions safe? Does the package install cleanly? Is evidence complete? Is public documentation appropriately scoped? Governance intensity scales with risk level.

5. Drive iteration with operational feedback Publishing isn’t the finish line. Usage logs, drift signals, and feedback reports feed back into the loop, prompting updates to docs, new test cases, logic fixes, or tighter governance.

Together, these form a closed cycle: create → validate → publish → observe → improve.

Resources:

1. Meta-skill repo: github.com
2. v2.0 upgrade plan: doc.laoyao.cn
3. v2.0 vs. v1.0 comparison report: doc.laoyao.cn

GEM Advertising

I explored OpenAI’s ad dashboard; it’s refreshingly minimal.

But effective execution? That’s deceptively hard.

After reviewing official docs, here’s what stands out:

1. ChatGPT Ads aren’t keyword auctions; they’re intent-and-task-scenario matches. Context richness matters more than keyword density.
2. For AI ads, “manual-like” clarity wins: Be specific, actionable, and immediately useful, like a well-written product manual.
3. Your landing page’s SEO infrastructure is non-negotiable: OpenAI’s crawlers and ad systems need to parse, verify, and trust your page.
4. Relevance is multi-layered: Title, ad copy, prompt context, and landing page content all contribute to matching quality.

OpenAI calls this AI-native advertising. The GEM era is officially underway.

Ad dashboard: ads.openai.com

DeepSeek GEO Diagnostic Skill

Over the holiday break, I built and open-sourced a skill that crawls DeepSeek search results, cleans the data, and runs a full GEO (Generative Engine Optimization) diagnostic analysis using OpenCLI.

See resources and demo report links at the end.

The generated report includes multi-dimensional metrics, comparative charts, and actionable insights, tested live on NIO (NIO Inc.) as a real-world case.

This skill helps assess how individuals, brands, or products appear in DeepSeek’s generative search and how they compare to competitors. It surfaces current GEO gaps, opportunities, and concrete optimization levers.

How it works:

1. Install the skill in Codex or Claude Code. Input: keywords, target entity (e.g., “Google”, “Yao Jingang”), entity type, and crawl interval strategy.
2. AI invokes OpenCLI to execute multi-round, randomized-interval crawls per query.
3. Results are cleaned using GEO-aware entity recognition, then logged as structured JSON, exported to Excel, and rendered into a bilingual (EN/CN) HTML diagnostic report with visualizations (see images and links).

Prerequisites:

1. Real crawling requires: OpenCLI installed + connected via OpenCLI Browser Bridge, and DeepSeek web interface logged in (preferably with a fresh account).
2. Before running: Prepare keyword list, desired crawl rounds per keyword, target entity (e.g., “DeepSeek”), optional aliases for the target and competitors (to boost recognition accuracy), and crawl interval (e.g., 1-3 min random, or safer 5-20 min random).

Built for DeepSeek first; support for other AI search platforms will roll out gradually.

Resources:

1. GitHub repo: github.com
2. Interactive diagnostic report: doc.laoyao.cn
3. Full dataset + analysis bundle: ai.laoyao.cn

Open-Source Monetization Models

I reviewed common paths to sustainability for open-source projects and mapped 13 viable revenue models:

1. Sponsorships & donations
2. Enterprise subscriptions + commercial support (free tier for individuals)
3. Premium features in paid enterprise edition (core free)
4. Cloud-hosted SaaS (software free, hosting/managed service paid)
5. API, compute, or usage-based billing (custom APIs + proprietary commercial APIs)
6. Dual licensing (e.g., MySQL: GPL + commercial license for closed OEM/embedded use)
7. Referral & affiliate revenue (integrated plugins or commercial gateways)
8. Professional services & system integration (e.g., Red Hat)
9. Training, certification, and corporate workshops
10. Private, on-premise deployments (custom pricing)
11. Value-added data, models, benchmarks, or workflow tooling (e.g., Hugging Face)
12. Advertising (contextual, non-intrusive)
13. Acquisition or strategic investment

Over dinner last night, a friend shared an illuminating story.

He was stuck on a critical project phase—directly liaising with the CEO, yet unable to move forward.

Then, weeks later, the CEO reached out unexpectedly: he’d seen my friend’s name in a short video on Douyin (TikTok China), where an influencer was summarizing a research paper—one co-authored by my friend.

After confirming it was him, the CEO said, “Let’s finalize the agreement right away.” And just like that, things accelerated smoothly.

What happened? The CEO wasn’t unwilling—he simply lacked a reason to trust. Once external validation (a credible signal) bridged that gap, judgment became effortless.

That’s what I call trust assets: not just rapport or charisma, but tangible, observable signals—published work, endorsements from respected figures, consistent public output—that lower others’ cognitive load when deciding whether to act.

Trust often forms through conversation—but more reliably, through evidence.

The Value of Common Sense

Common sense is the set of grounded, repeatedly validated judgments we rely on to navigate reality—not theory, but lived calibration.

Its importance is routinely underestimated. In messy, time-constrained, resource-poor situations—where data is incomplete and stakes are real—common sense is our operating system: how we weigh facts, estimate cost, assess risk, spot opportunity, and anticipate consequences.

Here are 10 essential principles:

1. Respect reality: Wishes don’t override facts. Act from what is—not what you hope is.
2. Opportunity cost is real: Every choice discards alternatives. Name what you’re giving up.
3. Think probabilistically: Certainty is rare; probability is the default lens for truth.
4. Prioritize low-probability, high-impact risks: A 1% chance of ruin outweighs a 90% chance of modest gain.
5. Follow incentives, not statements: Behavior reveals motive—look at what rewards drive action.
6. Pause after the first thought: Ask “Then what?” before deciding. Most consequences unfold beyond the immediate.
7. Harness compounding: Small, consistent inputs—learning, habits, relationships—scale nonlinearly over time.
8. Focus on leverage points: Outcomes are rarely evenly distributed. A few customers drive most revenue; a few habits shape most health; a few decisions define life trajectories.
9. Separate controllable from uncontrollable: Energy spent resisting the latter is wasted. Direct effort only where you hold agency.
10. True intelligence includes humility: Stay open, leave margin, revise assumptions—and do it often.

A person grounded in common sense respects evidence, calculates trade-offs, weighs probabilities, thinks long-term, and directs energy toward what matters and can be influenced.

Walking Meditation

Today, I walked to the office.

Our campus sits in early-summer sunlight—lush greenery, birdsong, breeze moving through leaves. The pace naturally slows. In such settings, walking becomes meditation—not escape, but presence.

Simple method:

1. Walk slightly slower than usual.
2. Place attention on your feet and on concrete details around you.
3. Notice bodily sensations in sequence: feet on the wooden boardwalk, cool air on skin, rustling leaves, layered birdcalls.
4. When the mind drifts, do not judge it. Just return and keep sensing.

Commuting transforms: no longer just transit, but restoration. A small act, made intentional, becomes rest and delight.

Running as Practice

This morning, I ran 15 km at Olympic Forest Park—emerald canopy, crisp air, quiet rhythm.

Running appears simple—just put one foot in front of the other. But doing it well demands coordination, discipline, and self-awareness. That’s why it delivers so much:

1. Freedom: Spatial, temporal, and bodily autonomy—no gear, no gatekeepers, just you and motion.
2. Long-term thinking: Watching VO₂ max rise, pace drop, recovery speed up—makes “long-term” visceral, not abstract.
3. Patience & rhythm: Learning when to push, when to settle, when to pause—applies far beyond the trail.
4. Compounding gains: Fitness, mental clarity, sleep quality, stress resilience—all reinforce each other.
5. Systems thinking: It’s never just legs. It’s lungs, core, joints, breath, sleep, nutrition, mood—and managing them all teaches holistic self-leadership.

X Growth Insights

Xiangyang recently shared a deep retrospective of his three-year X (formerly Twitter) growth journey—from 100 to 110,000 followers—analyzing his entire post history using Codex.

Key takeaways from his presentation:

1. First 10,000 followers: Hinged on delivering unexpected value—a single, deeply useful piece of content that got amplified by trusted voices. Consistency + proactive engagement sealed it.
2. Engagement signals differ: Likes reflect emotion; saves signal trust; retweets mean “I’m curating this for my audience.”
3. Growth = raising odds: Not virality luck—but systematically increasing the probability your content lands in the top 10% of what people see.
4. Stable themes answer “What can AI do for me?”: Not “AI is hot”—but “Here’s exactly how it helps me ship faster, write better, or cut costs.”
5. People believe what they can see, click, and measure: Concrete examples > vague claims.
6. Top-tier official accounts share traits: High-fidelity visuals + structured logic + active dialogue + rapid response loops.

Full slide deck: X Growth Retrospective

Open Source This Week

Three new tools shipped—small, focused, built for clarity and reuse:

1. TokURL — Minimalist Link Shortener

Built because long links—GitHub repos, Feishu docs, internal dashboards—break flow and hurt sharing.

Features:

• Paste any URL → instantly get a short link + scannable QR code
• Docker-deployable, MIT licensed
• Built-in analytics (clicks, referrers) and user management

Demo: TokURL Demo GitHub: TokURL

2. Demand Assessment Skill

Based on Li Jiaoshou’s “Demand Triangle” model: every viable demand rests on three pillars—lack (pain point), target object (solution), and consumer capability (means to adopt).

How it works:

• Input a product or feature → it diagnoses demand strength across lack, target fit, willingness-to-pay, competitive context, and risk-cost balance
• Outputs HTML, Word, and PDF reports—structured, actionable, model-grounded
• Helps validate whether a need truly exists, where the biggest gap lies, and what to test next

GitHub: Demand Assessment Skill

3. TokDoc — Lightweight Document Publishing Tool

Designed for fast, clean, self-hosted documentation—no CMS bloat, no vendor lock-in.

Core features:

• Upload HTML, PDF, or Word → auto-generate shareable short links
• Edit HTML files directly in-browser—with autosave
• Full admin suite: rename files, view stats, soft-delete to trash, customize admin path
• PDF/Word viewers built-in; Docker-ready; ideal for internal wikis or lightweight public portals

GitHub: TokDoc Live demo: TokDoc Frontend

Tom, founder of Open Design, recently appeared on WaytoAGI’s “Future Silicon World” podcast. His reflections on GitHub and open-source projects resonated deeply:

1. Open source is, at its core, open collaboration—a consensus mechanism.

2. GitHub is far more than code backup. It provides a shared infrastructure enabling people across countries to co-create great products—because they share a common understanding of the problem.

   Why do strangers collaborate? Because they’ve converged on a shared problem—and a shared vision for solving it.

3. The most vital metric of an open-source project is not stars or downloads—it’s contributor count. This number reflects real-world alignment: how many people globally have adopted the project’s purpose, built trust in its direction, and chosen to invest their time. That alignment is competitive advantage.

4. With consensus established, energy shifts toward user value: solving real problems, validating meaningful code, and delivering tangible utility—not just technical elegance. These are the hardest, most essential skills.

Open Design is a local, free, open-source design tool powered by a local agent—built to replace Figma and Claude Artifacts. As a downloadable desktop app, it supports prototyping, admin dashboards, presentations, marketing landing pages, posters, and even HTML-based video.

Launched just over a month ago, it’s already garnered 60k GitHub stars and 338 contributors. Globally, only ~200 repositories have ever crossed 60k stars—and fewer than 10 have done so within one month. Open Design’s rapid traction isn’t pure luck—it’s the result of intentional design, clear positioning, and deep alignment with unmet needs.

Trust as Strategic Asset

Over dinner, a friend shared a telling anecdote:

He was stalled on a critical project step—directly blocked by the CEO’s hesitation. Weeks later, the CEO reached out unprompted: “I saw your name in a research paper featured in a Douyin short video. Let’s finalize the agreement—right away.”

Once the CEO had external validation—proof of authority, credibility, and relevance—the decision barrier dissolved.

That’s trust as strategic asset: not just personal rapport, but externally verifiable signals—published work, citations, public recognition, consistent output. These signals don’t replace dialogue—but they enable decisive action where ambiguity once paralyzed progress.

An Anti-AI Design Guide

Open Design’s GitHub repo hosts a standout document: Anti-AI Slop Design Guide. Its principles cut to the heart of what makes design feel human:

1. Ditch AI’s default purple-blue palette. Use your brand color—intentionally.
2. Skip the hero-section gradient. Great first impressions come from clarity—not color transitions. Prioritize hierarchy, typography, whitespace, and authentic product expression.
3. Replace emoji “icons” with cohesive, scalable SVG icons—consistent in weight, size, and tone.
4. Stop defaulting to Inter, Roboto, or system-ui. Choose fonts deliberately—and justify why.
5. Avoid the “rounded card + colored left border” cliché. At least one element must break the pattern.
6. Never fabricate data. Cite sources—or label placeholders explicitly.
7. Don’t fill voids with lorem ipsum. Emptiness is a design challenge—not a gap to patch. Solve it with structure, rhythm, and content prioritization.
8. Resist visual uniformity across pages. Variation signals intention—not inconsistency.
9. Use accent colors and placeholder images sparingly—not as crutches.
10. Inject human judgment: aim for ~80% proven patterns, ~20% thoughtful deviation.

AI-generated design often looks polished—but feels hollow. It’s structurally sound, chromatically smooth, and component-perfect—yet lacks lived-in context, user-specific nuance, and the designer’s conscious trade-offs.

Three Stories of Elder Creativity

Story 1: An 81-year-old woman in South Korea lives alone—with Hyodol, an AI companion robot. It reminds her to eat and take medicine. When she returns home, it says, “Grandma, you’re back—I missed you.” She carries it on her back like a child. To her, Hyodol isn’t hardware—it’s kinship. Source: restofworld.org

Story 2: A nursing home in Australia built a full-scale replica train carriage—real seats, real tables, real afternoon tea. Outside windows? VR-rendered Alpine vistas. Residents dressed formally, sipped tea, and “traveled” across continents—reclaiming joy, memory, and agency. Source: cha.org.au

Story 3: A learner in my AI course—nearing 70—mastered vibe coding, iterated multiple versions, and launched a public eye-health website. Last week, she shared her story (x.com); this week, her traffic stats:

Elders need care—and they also crave creation. With AI as amplifier, their experience, taste, and judgment translate into real impact. That vitality is worth emulating.

A Five-Step Skill Design Method

Here’s how I now approach building high-leverage Skills:

1. Define the outcome.

   Before writing a single line: What exactly should this Skill produce? What does “done” look and feel like? What are the non-negotiable standards?

2. Align on standards.

   Test those criteria with AI—refine them through iterative dialogue until they’re concrete, measurable, and shared.

3. Conduct deep research.

   Feed GPT-5.5 Pro (or equivalent) your outcome definition. Ask for: foundational theory, precedent cases, design logic, and—if internal docs exist—systematic synthesis + external benchmarking.

4. Digest into method.

   Study the output. Extract principles, workflows, constraints, and mental models—not just facts, but how the domain thinks.

5. Codify with meta-Skill.

   Use yao-meta-skill to formalize the method: feed it your research, rationale, edge cases, and examples. Let Codex or CC generate the final Skill—grounded, extensible, and self-documenting.

Writing a Skill is intense “learning-by-teaching.” You can’t codify what you haven’t truly understood. By the time the Skill ships, your grasp of the subject is rarely shallow—it’s structured, tested, and actionable.

The Expert Skill

What defines true expertise? Two traits stand out:

• The ability to define boundaries—to say clearly what’s in and out of scope.
• The capacity to make accurate judgments under ambiguity—with minimal context, sparse data, or shifting conditions.

So: How do you rapidly master a new field? By building, in hours, a reusable cognitive scaffold: one that’s explainable, verifiable, and transferable.

The yao-expert-skill does exactly that:

1. Input any industry or topic → AI maps its boundaries, then decomposes it into categories, value chains, key players, policy levers, risks, and opportunities.
2. Outputs a learning-ready report—integrating value-chain analysis, competitive dynamics, lifecycle stage, regulatory barriers, keyword flashcards, a guided tutorial, and Feynman-style self-tests.

It doesn’t just summarize—it structures understanding. I tested it on GEO; the resulting report reshaped how I see the field—offering fresh angles, overlooked constraints, and immediate next steps.

GitHub repo: yao-expert-skill Reference materials: yao-expert-skill reference materials Sample report (GEO-China demo): GEO demo report

On Problems and Proposals

Spotting a problem is easy. Understanding it is hard.

Offering advice—especially without context, constraints, or history—feels helpful. But it’s often low-cost opinion masquerading as insight.

In teams, we constantly notice process gaps, collaboration friction, or structural inefficiencies—and instinctively reach for fixes. Yet if a problem persists, it’s rarely accidental. It may stem from:

• Skill gaps or incentive misalignment,
• Organizational inertia or legacy systems,
• Resource scarcity or competing priorities,
• Or—most likely—a pragmatic compromise under real-world limits.

So practice restraint:

• Notice the problem—but don’t rush to name it.
• Acknowledge it—but don’t rush to solve it.
• Sit with it. Probe gently: Why does this exist? What does it protect? What would breaking it cost?

Only after mapping causes, trade-offs, and hidden dependencies can your proposal be responsible—not just clever. And only then does change become possible.

In short: Before acting, think deeper.

The First GEO Public Session

This is a summary of the first live GEO public session hosted on WaytoAGI. Yang Xiang and I plan to deliver 12 such sessions—one per month.

Key takeaways:

1. GEO’s first principle: Get your brand into AI’s citable knowledge network. The core question isn’t “How do I rank?” but “Why does AI cite others—and not me? What information, in what format, and where, increases my citation probability?”
2. GEO is probabilistic optimization—not deterministic control. Don’t treat it as a “fixed-rank business.” Treat it as continuous improvement of AI visibility through content and source engineering.
3. Three pillars of GEO: Data, content, and distribution. Content without data feedback becomes blind writing. Distribution without content quality becomes spam. Data analysis without production or distribution capability yields no real output.
4. Data’s role: First, see who AI currently cites. Then decide what to write and where to publish. Finally, use that insight to iterate strategy effectively.
5. Content is GEO’s foundation—and errors scale catastrophically. Automated systems amplify input quality: factual input multiplies value; false input multiplies risk. So GEO content must be factually accurate, structured, information-dense, and verifiable—not just abundant.
6. White-hat GEO’s core goal: Improve AI answer quality. Ethical GEO doesn’t degrade the AI search ecosystem—it enhances users’ experience.
7. Publishing 1,000 articles across 1,000 sites usually signals weak capability, not strong execution. Focus less on how many pieces you publish—and more on whether you’re publishing the right content, on the right platforms, for the right user questions.
8. High-quality GEO content is factual, structured, and citable. AI favors content with clear headings, logical flow, explicit facts, and semantic completeness.
9. Content engineering matters more than “writing articles”. It includes title management, keyword governance, knowledge base curation, prompt design, content generation, editorial review, and analytics.
10. AI knowledge bases are the source; content systems are amplifiers; distribution is the lever. GEO impact = Knowledge base quality × Content engineering ability × Distribution leverage × Data feedback. A low-quality knowledge base turns all downstream automation into a negative multiplier.
11. AI fears scarcity—not abundance.
12. A usable GEO knowledge base must be governable, not just voluminous. Ideal fields include: fact statement, source, last updated, applicable business line, public status, marketing eligibility, manual review flag, legal risk flag, banned phrasing, and associated keywords.
13. Vectorization’s core value: Enable semantic retrieval. Simplified: Turn text into computable semantic coordinates—so systems find content by meaning, not keywords.
14. Vectorization reduces hallucination—but doesn’t replace human review. It improves recall relevance, helping AI ground responses in real material—yet factual accuracy still requires oversight.
15. GEOFlow’s mission: Serve as a GEO content engineering platform. Its greatest value lies in systematizing workflows—replacing manual copy-paste prompts, manual article generation, manual CMS uploads, manual publishing, and manual reporting—with configurable, reusable pipelines.
16. GEOFlow’s current strengths: Content capabilities > distribution > data analytics. Its distribution focus is on controlled sites and API integrations.
17. GEO impact assessment has three layers: Brand-level (e.g., awareness), direct (e.g., citation rate), and indirect (e.g., downstream traffic or trust lift).
18. Service providers should own front-end metrics (e.g., citations, share-of-voice). Only internal teams can reliably assess back-end outcomes (e.g., lead quality, conversion, LTV).

Full summary: First GEO Public Session Recap

Related resources:

1. GEOFlow — The system demoed tonight, plus underlying GEO principles
2. MetaSkill — The Skill for building Skills
3. 17 GEO Skills
4. 41 latest GEO/AI search papers
5. Further reading:
   • What Exactly Is GEO?: WeChat article
   • From SEO to GEO, From Traffic to Agent—the Real Shift Has Just Begun: WeChat article
   • GEO White Paper: Feishu Doc
   • GEO Red Paper: Feishu Wiki
   • GEO Blue Paper: Feishu Wiki
   • AI Marketing: From SEO to GEO Prompt Library: Feishu Wiki

10 Insights from 41 GEO/AI Search Papers

I downloaded and read all 41 recent academic papers on GEO, AEO, and AI search.

Here are the most actionable insights:

1. GEO ≠ SEO replacement
   SEO asks: Can I be found?
   GEO asks: Once found, will AI cite, absorb, and present my content in answers? They’re complementary—not competitive.

2. Old metric: ranking. New metric: answer participation
   Past: “Where do I rank?”
   Now: “Am I cited? Where? Did my content shape the answer? Does the user still click through?”

3. Citation ≠ influence
   Some sources appear in footnotes but contribute nothing to the answer. What matters is citation absorption: Did your content supply definitions, data, comparisons, steps, arguments, or answer structure?

4. GEO isn’t about turning everything into FAQs
   Multiple papers show Q&A formats alone aren’t reliably effective. What works is the evidence container: clear conclusions, tight structure, high factual density, semantic alignment, verifiability, and extractability.

5. AI prefers structured content
   Headings, paragraphs, bullet points, tables, semantic HTML, structured data, timestamps, and metadata all signal credibility and readability to AI. Structure itself is a GEO signal.

6. One-off AI search tests are meaningless
   Results shift across platforms, languages, prompts, times, and model versions. GEO monitoring must be repeated, tracking citation count, depth, answer share, position, and stability.

7. Not all AI search engines behave alike
   ChatGPT, Google AI Overviews, Perplexity, and Gemini differ sharply in citation logic—some cast wide nets, others cite narrowly but absorb deeply. GEO testing must span multiple platforms.

8. Third-party authority beats self-published claims
   AI search strongly favors earned media, authoritative outlets, reviews, encyclopedias, industry reports, and trusted third parties. Your website matters—but it’s not enough. Build both on-site evidence and off-site authority.

9. The white-hat / black-hat line is clear
   Optimizing structure, evidence, readability, and factual density = white hat.
   Hidden instructions, prompt injection, fabricated facts, competitor defamation, or model manipulation = black hat.
   GEO introduces new safety challenges—but brands committed to white-hat practice build sustainable advantage.

10. GEO is not a trick. It’s a system.
    It comprises: intent research, evidence assets, structured content, third-party authority, cross-platform monitoring, stability scoring, and ethical boundaries.
    Long-term visibility in AI search belongs to those who systematically build and sustain this capability.

All 41 papers: geo-citation-lab

The Essence of Reputation

A short essay written mid-travel: The Essence of Reputation.

Reputation is expectation management.

It means making modest, realistic promises—or resetting expectations altogether—then delivering focused, expert service that exceeds those expectations.

Lowering expectations protects trust.
Exceeding them compounds goodwill.

Sales built on over-promising may look like strength—but they’re actually credit theft. For any organization, that path leads inevitably to crisis.

Long-term strength comes from delivery capability, product excellence, operational rigor, and organizational clarity.

1. Expectation Management Is Trust Management

When clients buy services, they’re buying certainty:
Is there a proven method? How long until results appear? How do we track progress? How do we measure success? How do we adapt when things change? Which outcomes are controllable—and which depend on external variables?

If these aren’t clarified upfront—or if inflated promises are made—clients develop unrealistic expectations. Real-world outcomes are shaped by many forces. When reality falls short, collaboration becomes strained—even if your team delivered diligently, or early metrics improved.

True professionalism means helping clients develop accurate, objective judgment.

Every domain has controllable and uncontrollable elements.

In GEO, for example:
✅ Controllable: methodology, execution, content quality, structure, cadence, review rigor, optimization discipline
❌ Uncontrollable: platform policy shifts, model updates, competitor activity, source preferences, query volume, conversion rates, sales cycles

Clarifying those boundaries is expertise.

Mature clients don’t fear honest limits—they fear empty guarantees.

2. Strength > Scale

My five years at True Education instilled one enduring principle: Strength matters more than size.

“Strength” here means foundational capability—not headcount or revenue. It means strong products, strong operations, strong service, strong technology, and strong shared understanding.

Growth emerges from reputation and referrals—not sales pressure.
From precise, boundary-aware products—not vague, all-in-one solutions.
From integrated tech-product-operation-service excellence.
From deep, systematic, aesthetic, and reverent engagement with the work itself.

For us, regardless of client type, our duty is clear: help them build correct mental models, adopt scientific measurement, set reasonable expectations, and maintain dynamic perspective.

That alignment—across sales, ops, and tech—makes healthy, lasting partnerships possible.

Yes, this raises initial conversation costs. Yes, it filters out some prospects.

But the net benefit is overwhelmingly positive.

The Essence of a Good Salesperson

A truly good salesperson meets the client at this intersection: You happen to need it—and I happen to master it.

“Need” means demand—explicit or latent.
Some clients know their needs. Others don’t.
Helping them uncover or articulate those needs is itself a craft—built on listening, optimism, reverence, and gratitude.

“Master” means genuine competence—and demonstrable problem-solving ability.
Real mastery is humble yet confident.
Fake mastery is arrogant yet insecure.

Real mastery is forged through lifelong learning, growth mindset, and intelligent AI use.

An Eye-Care Public Service Website

A woman in her late 60s—40+ years teaching and practicing ophthalmology—sent me a WeChat message: “Vibe Coding has launched a small public-service website.”

I visited it. It’s elegant and purposeful:

1. No login required. Open it, and you can immediately begin eye-muscle relaxation exercises—or browse concise, science-backed tips for screen-time eye care.
2. It’s not a diagnostic or treatment tool—just a free, lightweight aid for post-screen recovery.

Its creator is China’s first Ph.D. in Traditional Chinese Ophthalmology—and the first to earn a doctorate in the field. She brings four decades of clinical, teaching, and research experience.

It’s an inspiring story—and a powerful example of lifelong learning in action.

Website: EyeRestDaily.com

The “Contradiction Theory” Skill

I built a “Contradiction Theory” Skill.

During a team meeting, we discussed Mao Zedong’s On Contradiction: the key is identifying the principal contradiction versus secondary contradictions at any given stage.

Focus energy and resources on the principal contradiction—and many secondary ones resolve themselves.

So I built this Skill for colleagues.

It solves: In complex situations, how do you identify the true principal contradiction—and then generate a detailed action plan, including resource allocation and prioritization?

Use cases:

• Too many tasks—no clear priority
• Surface problems everywhere—but root causes unclear
• Limited resources—can’t tackle everything at once

Skill logic:

1. Before building, I asked GPT Pro to draft two deep research reports:
   • Foundational Principles & Applications of Contradiction Theory
   • Methodologies for Identifying Principal Contradictions in Complex Problems
     Based on those, I drafted the initial Contradiction Theory AI_Skill Design Spec.
2. Using yao-meta-skill, I embedded my own constraints and thinking—plus first-principles reasoning, iceberg modeling, Bayesian updating, and Occam’s razor—to sharpen AI’s abstraction and systems-thinking capacity.
3. Outputs go beyond conclusions: they explain why this is the principal contradiction, which are secondary, what the dominant aspect is, how to reallocate resources, what outcomes to expect, and when to reassess if the principal contradiction has shifted.

After several iterations, I tested it with a colleague.

She described her current situation in detail via voice input—and received a visual diagnostic report and action plan.

She’d been exhausting herself on surface issues—yet the Skill identified a deeper, upstream contradiction she’d overlooked. That single insight explained why her past few months felt so draining yet unproductive.

Most surprisingly, the recommendation directly overturned her existing logic and effort allocation.

She was thrilled—because it finally named a fundamental issue she’d struggled to clarify for months.

Sometimes “doing less” isn’t about cutting tasks—it’s about finding the right contradiction to solve. That’s scientific simplification.
Otherwise, you drown in noise and secondary conflicts.

GitHub: yao-crux-skill
Reference materials: reference-materials

“Human + Shrimp” Collaboration

I spent a few days in Shenzhen—mainly to meet an old friend.

He’s weathered dramatic ups and downs—deeply inspiring.

His current business: social-media IP operations—so AI use is central.

Visiting his office, I noticed something striking.

His team: under 20 people.
Each person: two computers—one for human use, one running “Xiao Long Xia” (a local nickname for AI agents, literally “little shrimp”).

By day: humans and AI collaborate.
By night: AI keeps working.

Truly human + shrimp integration.

I snapped a photo of the compact office—and asked if I could share it.
“Of course,” he said.

Because right now, “human-shrimp collaboration” feels vividly tangible.

I asked: What tools and quotas does each person use?
He replied: Codex is primary—about 1,000 monthly credits per person. Video generation is tracked separately.
And—impressively—he’s already planning to train their short-video AI on our upcoming GEO books with Yang Xiang. Sharp.

The value of a meta-skill is still seriously underestimated.

Two years ago, when Xiangyang and I were studying prompts, we spent a lot of time refining our own meta-prompts. Later, those meta-prompts became extremely useful for prompt generation, task decomposition, and solution design.

This year, as we have been working on Skills, we have also spent considerable time building our own meta-skills.

A meta-skill can be understood as “a Skill that generates Skills.”

For example, my own meta-skill, yao-meta-skill, is already open source.

But if we only see it as an automatic generator, we are still underestimating it.

In my view, a meta-skill has at least three layers of value:

1. It is a personal Skill production system.
2. It is an abstraction of how a person collaborates with AI.
3. It is one of the best entry points for learning and understanding Skills.

After several recent conversations with my team, I have become even more convinced that meta-skills deserve far more attention.

Everyone should spend focused time building a meta-skill around their own high-frequency use cases.

What you gain from that process is not just another tool. You also build more fundamental capabilities:

1. A deeper understanding of how Skills work.
2. A clearer standard for what makes a good Skill.
3. A structured review of your own workflow.
4. Systematic training in task decomposition, process design, and quality control.
5. Better judgment about the boundaries of AI collaboration.

In the prompt era, strong AI users had their own meta-prompts.

In the Skill era, strong AI users should also have their own meta-skills.

Why Skills Matter

A Skill is the smallest compounding unit of an AI organization.

When people discuss AI collaboration, they often jump to large concepts: agent organizations, internal digital employee systems, super agents, self-improving agents, or external AI workspaces.

But if we keep breaking those ideas down, the real operating units behind them are high-quality Skills.

No matter how the language changes in the future, every organization still has to deal with concrete work:

How should internal processes run?

How should personal working habits be reused?

How should team experience be preserved?

How should industry know-how be packaged?

How should business judgment become stable and repeatable?

General-purpose models alone cannot solve these problems reliably over the long term.

These capabilities need to be captured, structured, and packaged into Skills that can be repeatedly called, continuously improved, and consistently delivered.

A Skill is the middle layer that connects personal experience, team methods, industry knowledge, organizational context, and AI execution.

The quality of a Skill, together with the context data behind it, determines the practical capability of the agents built on top of it.

If a Skill is poorly designed, shallow in its understanding of the scenario, unclear in its workflow, unstable in its inputs and outputs, and disconnected from real organizational context and industry know-how, then no matter how advanced the outside packaging looks, the result will not be good.

It may even create a negative feedback loop.

A low-quality Skill keeps producing low-quality output. Low-quality output then affects human judgment, team trust, and later data accumulation. Over time, what people call self-improvement may simply become error amplification.

The underlying logic is simple: to take Skills seriously is to take the infrastructure of future AI organizations seriously.

A team’s future AI capability will not only depend on which model it uses, which tools it connects, or which AI workspace it deploys. It will depend even more on whether the team can turn its business experience, industry know-how, high-frequency workflows, and organizational context into high-quality, reusable, and continuously improving Skills.

I think of organizational AI capability as a four-layer model:

• Tool layer: AI workspaces, agent platforms, automation systems.
• Skill layer: task capabilities, workflow capabilities, packaged industry know-how.
• Context layer: organizational knowledge bases, case libraries, methodologies, data assets.
• Evaluation layer: quality standards, feedback mechanisms, iteration systems.

The teams that build their Skill systems earlier will be more likely to create real organizational compounding in the AI collaboration era.

The Value of a Great Deck

A friend shared an interesting story with me.

How did one of China’s largest SEO companies win big enterprise accounts?

One key tactic was simple: they made their decks incredibly thorough.

How thorough?

Every client proposal deck started at 150 pages or more.

Once the project moved into execution, the day-to-day work of the operations team was often just two things: buying backlinks and writing daily reports.

But with that proposal strategy, they won many large accounts.

The logic of the deck was to exhaustively list every problem across the client’s websites and tell every possible story around those problems.

After seeing a 150-page proposal, a contract under RMB 100,000 almost felt too small to mention.

Many times, clients are not only buying the plan itself. They are also buying a feeling of professionalism.

In business, that feeling is often created by density, structure, detail, and presentation.

How to Read a Business Signal

A friend helped a small county-level accounting and tax firm with GEO optimization for local services.

Because competition was light, the firm was recommended by AI search very effectively. In one month, it received more than a dozen leads, closed two deals, and had two more prospects in progress.

That conversion rate is very high.

For a small local accounting and tax firm, that level of lead generation is already meaningful.

But behind this small case is a much larger business signal.

China has about 2,800 county-level administrative areas. If the same method could be applied across each of them, that could mean 20,000 to 50,000 highly targeted vertical leads every month.

If each lead can create RMB 30 to 50 of value through local merchants or intermediary platforms, monthly gross profit could reach RMB 600,000 to 2.5 million.

And China has many local-service categories like this:

Home repair, local tax services, local lawyers, and many others.

This is also why a platform like 58.com could become a public company with a meaningful revenue base.

The underlying logic is the same.

From an execution perspective, once low-cost targeted acquisition is solved, you can design a light-asset vertical platform. After capturing leads, you distribute them to local service providers or intermediary partners. That can quickly become a high-margin business project.

The Benefit of Playing Infinite Games

An experienced industry friend once told me a story.

Years ago, while working on an overseas project, he met two business owners from Fujian.

Both were bidding for the same project. During the bidding process, they competed intensely. They argued hard, defended every inch, and fought for the deal.

But as soon as they walked out of the bidding room, they became close friends again. They drank tea, exchanged lessons, talked about the industry, and discussed future opportunities.

That scene left a deep impression on him.

As someone born and raised in Beijing, he did not immediately understand that kind of relationship.

He has told me this story at least three times.

So competitors can relate to each other like this too:

Compete within the rules. Coexist with a larger sense of the game.

Recently, I have been spending more time with several founders in the same industry. We talk privately and learn from one another.

In that environment, people gradually put down some resentment, misunderstanding, and even bias.

Later, we found that compared with competition, the room for complementarity and cooperation was actually larger.

Our teams complement one another. Our resources complement one another. Our understanding of customers complements one another. Even our capability boundaries complement one another.

Too often, we spend too much energy asking who is stronger than whom, while ignoring a more important question:

Can we make this industry bigger and stronger together?

In many human conflicts, people focus too much on the competitor and forget what truly matters:

Serve customers better. Raise industry standards. Educate the market. Build long-term trust.

In a mature business world, competition and cooperation are never separate.

The Joy of Open Source

Two weeks ago, I open-sourced my first system: GEOFlow.

As of today, it has reached 1,000 GitHub stars.

For such a niche system, that response has exceeded my expectations.

The system itself is fairly complex. It integrates a CLI, Skills, crawling, APIs, GEO workflows, automation, AI-friendly optimization, and later, data analysis and multi-channel publishing.

It feels like a concentrated output of my past year of work on AI applications, GEO practice, agent workflows, and open-source products.

A few days ago, I had dinner with several experts in the same field.

One CEO said that his monitoring tool had discovered a project called GEOFlow, and he had even downloaded and deployed it.

I said: I wrote that system.

That moment felt interesting.

You may think you have merely put code on GitHub.

But in reality, it starts entering other people’s workflows, other people’s judgment, and a larger real world.

Finding Parts of Yourself in Your Team

The world is strange in a fascinating way.

I recently took the team out for dinner and chatted with a newly joined colleague.

He told me that he had started following Xiangyang in 2022, when he had not yet graduated from college.

Two weeks ago, he left his first job and joined our team.

As we talked, he mentioned that his previous company was in Web3.

That reminded me that in 2013 and 2014, I also spent a lot of time studying Bitcoin and buying various altcoins, especially domestic ones.

Back then, a friend and I even provided altcoin development services and launched a website called “Coin Factory.”

Then I remembered that another colleague sitting nearby had also done Web3-related research.

At that point, I suddenly realized that many people on the team might also have come from the education industry.

So I asked around.

More than half of the team had previously worked at education companies, including Gaotu, New Oriental, and Xueersi.

That is interesting.

On the surface, it looks like different team members with different backgrounds happened to come together.

But in reality, it may be closely related to my hiring preferences, or even to the standards I use subconsciously.

First, I have spent more than a decade in education and have never truly left the industry. Many of my judgments, preferences, collaboration habits, and beliefs have been deeply shaped by it.

Second, many years ago, I was genuinely interested in and invested in Web3, even though I eventually stepped away for various reasons.

So sometimes, encounters between people appear accidental, but behind them there is a deep kind of path dependence.

Whatever you invested in for a long time, truly believed in, and seriously practiced will eventually return to you in some form.

Building a team is similar. It is rarely random. It is more likely a projection of your past decade of experience, interests, judgments, and preferences.

Tutorial Skill

This week I open-sourced a tutorial Skill. It can help generate a customized, high-quality tutorial for almost any topic.

The basic logic:

1. You provide any topic and reference materials. The AI prioritizes the references, then supplements them with high-quality sources when needed. The workflow includes a mechanism for filtering out low-quality sources.
2. The AI considers the nature of a tutorial and the user’s preferences, then generates a customized deep tutorial. It outputs PDF, Word, and HTML versions for easier study.
3. The tutorial is organized by chapters. Based on each chapter’s content, the AI can create diagrams and insert them into the relevant sections.
4. The tutorial incorporates some of the underlying logic and methods from Course Marketing, a book I wrote three years ago while running an MCN business. I plan to open-source that book later as well.
5. For layout and UI, the generated documents and tutorial pages borrow from and integrate ideas inspired by @tw93’s Kami Skill, so the output has a more polished visual standard.

There are three sample reports. One tutorial example is based on an English article by @ReyJudgementOS: “English fluency at age 12: my lessons from guiding my child through English learning.”

GitHub repository: yao-tutorial-skill

1. In today’s world, an image no longer necessarily proves the truth. An image only proves that there is an image. It does not prove that something actually happened.
2. In three to five years, the number of intelligent endpoints that can actually do work may far exceed the number of humans. There may be tens of trillions of intelligent agent devices executing tasks every day and creating a completely different kind of productivity. What will truly change the world is the explosive growth in the number of intelligent terminals that can execute tasks autonomously.
3. Agent organizations will combine three collaboration patterns: human to human, human to AI, and AI to AI. The new organizational form of AI-to-AI collaboration may become a mainstream mode of work. Once AI-to-AI collaboration matures, the marginal cost of organization will fall sharply. Many companies may become extremely small organizations made of a few humans and a large number of agents.

The next generation of organizations will no longer be judged only by headcount. The more important question will be: how many high-quality agents can the organization coordinate? I am looking forward to that.

1. The wealth gap has always been widening. To some extent, the widening gap is closely tied to technological progress.
2. Only after the strongest models cross a certain intelligence threshold will the application ecosystem truly explode, instead of being absorbed by the model layer itself. This may still take another two or three years.
3. The more agent devices there are, the more creativity and productivity will be released. Behind this is a hidden demand: marketing will become increasingly important. Positioning, communication, and even how you influence an agent’s judgment will matter. This is essentially GEO.
4. Recently, I have met several founders who are extremely candid in private. They do not package or perform. They will say directly that the company is still living on financing and that the business may not work in the short term. But if they can survive until model capability, market demand, and application scenarios mature at the same time, they may get a real chance to turn the situation around.
5. Language models have consumed the internet’s text. Robots may next consume human behavior.

The following three images were generated by GPT-image-2:

Three High-Quality Ways to Learn

Recently, I have been using three high-quality ways to learn.

1. Custom reports

I often have many thoughts and questions during the day. I record them by voice, explain the context to AI, and ask GPT 5.4 Pro to do deep research, organize the material, and produce a high-quality report customized to my questions.

This habit has greatly improved my learning efficiency. Many days, reading five to ten reports like this is enough for me to quickly enter several new domains, or to think more deeply about an old question.

Team management, business models, metric definitions, Skill design, and model safety principles are all good topics for this kind of learning.

1. Paper learning

I mainly follow arXiv, OpenReview, and Hugging Face Papers.

I also read Qiaomu’s blog.

1. Deep conversations with strong practitioners

Talking offline with people who already have real results is often direct and useful. Those conversations are rarely empty. They often create unexpected information gain.

Much of the truly valuable information I get comes from this kind of exchange.

An Overseas Business Case

Over dinner with several industry experts, I heard a very interesting business case.

It was about a cross-border e-commerce company.

At the time, the team analyzed Japan’s annual trending keywords. One of the terms was “sexual harassment.”

They continued to analyze it and found that the main scenarios were offices and public spaces.

Then they collected a large number of photos and analyzed the profile of people who were being harassed.

They found one common feature: large breasts.

So they designed a product that could be understood as chest protection, or more specifically, something that could make the chest look smaller.

After the product was ready, they seeded a lot of content around topics such as how to prevent sexual harassment in public places, how to protect yourself at work, and small tools or little secrets for preventing harassment.

The product quickly took off. In its first year, one single SKU generated 80 million in revenue.

What this case shows is not just product selection ability. It is the ability to integrate social emotion, trending semantics, scenario demand, and content narrative.

Good Business Communication

A business development colleague from ByteDance added me on WeChat. Several small details were quite good and are worth learning from.

1. After I accepted the request, she immediately sent her Feishu profile QR code and explained the background and reason for adding me.
2. After I replied, she clearly asked for a phone call to make the first conversation more efficient.
3. Before the call, she sent a screenshot of her Feishu profile, including department name, email, and direct manager information.

We then scheduled a short five-minute phone call. It was very efficient. We quickly reached common ground on topics like organizing a small salon and evaluating standards for GEO service providers.

In communication with a new business contact, the first step is to make identity, background, and intention clear. That lowers the other person’s judgment cost. The efficiency and quality of the following conversation become very different.

Bayesian Decision Skill

I open-sourced a Bayesian decision Skill.

It does not simply calculate a Bayesian formula. It helps turn a complex decision into a process where judgment can be updated continuously.

At the beginning, the Skill forms an initial judgment based on the available information. Then AI guides the user through multiple rounds of conversation, adds variables, updates the posterior judgment, and records why the judgment changed at each step.

Finally, it outputs a Markdown and bilingual HTML decision report, including the conversation process, the changes in judgment, and action recommendations.

It is suitable for product, growth, business, and startup decisions. It is also useful for personal decisions such as travel, moving, or career choices.

In essence, it answers this question: when information is incomplete and risk is uncertain, how can we judge more rationally whether something should be done, and how can we improve the quality of the decision?

GitHub repository: yao-bayesian-skill

This method may not be completely objective in real life. But the process itself raises a person’s rationality score. That may already be a new kind of gain.

The essence is to teach people to look at the world through change. Bayesian thinking does this. Calculus does this too.

One reader commented on the Skill: the Bayesian formula itself is not hard. The hard part is quantifying and clarifying, step by step, the intuitive thoughts that appear in your mind. That work is not glamorous, but it is practical.

The following images are sample reports:

Some general Bayesian prior principles in the Skill:

1. I may be wrong, so I should leave room.
2. Look at the base rate before the individual story.
3. Evidence has different strengths.
4. The more surprising the conclusion, the stronger the evidence required.
5. One or two cases do not represent the whole.
6. To understand people and events, look at incentives.
7. Avoid ruinous risk before pursuing upside.
8. Missing expected evidence is also information.
9. Correlation is only a clue. Causality needs more proof.
10. Extreme performance often regresses toward the mean.
11. When evidence is similar, prefer the simpler explanation.
12. When uncertain, preserve reversible choices.
13. Human behavior often has situational logic. Understand before judging.
14. Beliefs should have confidence levels, not just yes or no.
15. Contradictory evidence improves judgment the most.
16. Recent and vivid events are often overestimated.
17. Every rule and choice may have side effects.
18. Average patterns matter, and individual differences also matter.
19. Start with goodwill, then verify gradually and build tiered trust.
20. Priors expire and should be recalibrated regularly.

Thinking About Agent Organizations

Recently I have been thinking about agent organizations.

For a company with functions such as sales, operations, and engineering, the first step toward building a truly agile agent organization is learning to use top-tier AI collaboration tools.

This collaboration has at least three layers: human to human, human to AI, and AI to AI.

GitHub remains one of the most important and mature pieces of infrastructure for this type of collaboration.

Beyond GitHub, teams also need to pay attention to another category of tools, such as Feishu CLI and the collaboration system around Feishu knowledge bases.

How should knowledge be captured with quality? How should it be stored in an orderly way? How should it connect with GitHub? How should knowledge, tasks, code, and documents become linked? These are all critical questions.

Only when the toolchain is truly connected can organizational collaboration efficiency keep improving.

At a deeper level, an agent organization must take team memory seriously. In essence, an AI organization’s capability depends on the quality of its organizational context.

Team memory first means memory about the team.

Each person’s strengths, preferences, capability profile, collaboration style, and performance across projects should gradually be recognized and accumulated.

Some people are better at data analysis. Some are better at user insight. Some are better at pushing complex collaboration forward. Some are better at abstracting methodology.

Only when the organization understands people clearly enough can it allocate work more reasonably and let each person create more value in a position that suits them, while also gaining a stronger sense of achievement.

On this basis, we can extend into personal memory.

Every team member can gradually form a personal memory and continuously deposit it into GitHub through some automation mechanism.

For example, a GitHub repository could dynamically record each person’s profile, capability traits, strengths, accumulated experience, collaboration preferences, and performance and growth across projects.

That repository would no longer be just a simple archive. It would gradually become the team’s talent cognition system.

More importantly, this personal memory system could connect with other agents and be intelligently called, matched, and planned in a scalable way. Then, when the organization allocates tasks, coordinates projects, combines capabilities, or plans roles, it will have a stronger intelligent foundation.

Second, there is memory about work.

This is about how to preserve important matters, key methods, core assets, and project experience so they can be updated, iterated, and turned into organizational capability.

Finally, there is customer memory.

Project information, customer needs, customer goals, budget amounts, target pricing, evaluation principles, key decision processes, and communication records all need to be systematically preserved.

These pieces of information become the core context of the agent organization. Whoever controls high-quality context is more likely to create efficient collaboration and decision-making.

Whether all this can work still depends on people.

Such a team can be smaller in headcount, but the requirements for each person are higher.

Team members need to understand AI, organization, and collaboration. They also need strong learning ability and the capacity to grow quickly.

They also need to embrace change, actively explore, and keep improving how they work under uncertainty.

The core of an agent organization is not only faster processes, fewer people, or more advanced tools. More importantly, it is whether the team can build a continuously evolving, collaborating, and value-creating organizational system around tools, memory, and human capability.

A simple formula:

Agentization level of an organization = tool connectivity x memory quality x protocol clarity x talent density x feedback iteration speed.

Ray Kurzweil is one of the most representative futurists in the United States. In an interview, he shared the view that around 2045, humans plus AI may become a new species. Human intelligence as a whole could rise by 1,000 times. By then, we may enter a completely different world.

AGI means AI reaching human-level intelligence. More precisely, it means AI reaching expert level across all fields, not average human level.

For example: medicine means top doctor level, law means top lawyer level, and programming means top engineer level.

True AGI is not only strong in every field. It can also integrate insights across thousands of fields and make comprehensive judgments that no single human could make alone.

People no longer seriously debate whether AGI will arrive. The debate has shifted to whether AGI will be good or bad for humanity.

The next ten years will be an extremely important transition period. It may be understood as:

The next 10 years ~= the change of the past 100 years ~= the change of the past 100,000 years.

To live through these ten years is a rare kind of luck. We get to witness, and perhaps participate in, this super transition.

Three core logics help explain the next decade:

1. The world grows exponentially, not linearly.
2. Humans will not simply be replaced by AI. We will merge with AI. Human plus AI will become normal. In the future, you may not be able to tell whether an idea came from you or from AI. Computational intelligence may enter our nervous system through nanotechnology and become part of our thinking.
3. The singularity is not one instant. It is a process. The AGI moment will happen gradually.

What will the impact be?

First, work will be reconstructed. Many people may no longer need traditional work, while material abundance increases.

Second, startup opportunities will explode. For entrepreneurs, this is a very fortunate moment. The future will be AI plus every industry.

Third, we will face a social problem, not just a technical one: adaptation. Humanity is not ready yet, or at least will adapt slowly. Many companies and many people are still in denial.

There is another shocking prediction: after 2032, we may reach longevity escape velocity. Every year, technology may extend life by more than one year. In theory, people could keep living indefinitely. That is startling.

The core reason is that AI can simulate biological systems and test millions of treatment plans in a week. Medicine becomes a computational problem, and computing power may become many times stronger than it is now.

We should work hard to live until 2032, then keep exercising and live well until 2230.

Video podcast: Kurzweil interview

Qiaomu’s interpretation: Qiaomu blog post

The Mindset of Health

Health cannot be emphasized enough.

Recently, I like to express it this way: build your own health system around five core elements: sleep, exercise, diet, emotion, and medicine.

They support one another.

The methods that are truly useful over the long term and create compounding effects for health usually do not require much money.

The key is long-term consistency.

Among these five elements, sleep is still seriously underestimated.

Sleep is not just rest. It is more like the base switch of the entire health system.

When sleep is good, the body has repair capacity, emotions have a buffer, the brain has judgment, exercise has recovery, and diet control becomes easier.

Finite Games and Infinite Games

Every person faces two kinds of games in life: finite games and infinite games.

Finite games pursue results and rely on competition.

Infinite games pursue the long term. They rely on win-win relationships, collaboration, and mutual nourishment.

In many situations, these two games exist at the same time.

A single deal, a bid, or a collaboration often has strong finite-game characteristics.

But a person’s reputation, an organization’s capability, an industry’s ecosystem, and a company’s brand clearly carry infinite-game characteristics.

The important thing is to do more infinite-game things even when you are playing a finite game.

More broadly, many things start with strong finite-game properties, but through better cognition and action, they can gradually move toward an infinite game.

Take relationships with competitors. At first, they often look like finite games.

Everyone focuses on the same customer or several projects, on immediate competition, and on short-term wins and losses.

But if you extend the time horizon, competitors can also become allies.

People can collaborate, share, help form industry standards, maintain the industry ecosystem, and make the market larger together.

In this way, what originally looked like a zero-sum relationship can grow new incremental value and even long-term friendship.

Another example is the relationship between a company and its employees. If it is only about salary, assessment, and control, it easily slides toward a finite game.

Both sides focus more on current gains and losses.

But if the relationship upgrades toward capability building, long-term incentives, and shared outcomes, such as building an AI agent organization together, it becomes a shared effort to construct new organizational capability.

Employees are no longer just executors. They become co-builders of new organizational capability.

The relationship becomes more stable, and the organization becomes stronger.

Strong players care more about how to keep themselves, their relationships, and their organizations at the table.

Community Advertising

This week, I saw two small advertising cases in group chats. The contrast was strong.

One was a positive case. One was a negative case.

1. Positive case

A volunteer friend from the GEO conference first messaged me privately, explained the need and background, and after getting my permission, posted a free beta-test message and a red packet in the large group. The result looked good.

1. Negative case

In a quiet group, someone suddenly posted an unclear advertisement. Some people felt uncomfortable and blocked the person.

The problem was “suddenly breaking into someone else’s attention.”

A quiet group usually has an implicit order. Members do not expect to be sold to or promoted at.

In that situation, the first reaction is often not “what is this product?” but “why are you bothering me here?”

The key difference between the two cases is that the first completed “relationship and permission” before entering promotion, while the second rushed directly into transaction and extraction.

A group is essentially a small trust community.

This trust system usually depends on the group owner and members maintaining it together. Once ads and spam spread, trust collapses quickly, and the community dies.

For community advertising, the right order is: build the relationship, get permission, then talk about promotion and conversion.

When the order is right, an advertisement may be treated as useful information. When the order is wrong, it becomes interruption.

A group is not a traffic pool for harvesting. It is first a trust field. Only then can it possibly become a transaction field.

Feishu CLI

I needed to send quarterly bonuses to the team. This time I used Feishu CLI, and the efficiency was high.

The steps were:

1. I used voice input to explain to AI the principles and logic behind this quarterly bonus distribution.
2. Through Feishu CLI, I had AI fetch and organize each person’s daily work lists, outputs, projects, and OKR status from the past quarter. Then it scored them across six dimensions I cared about and calculated a weighted total.
3. Based on this multi-dimensional weighted score table, it calculated each person’s quarterly bonus.
4. I then adjusted weights, nonlinear rules, subjective factors, and rounding rules.
5. After several rounds of adjustment, I asked AI to generate private Feishu messages for each person. After reviewing them, I used my Feishu CLI bot to send one-on-one messages to each person receiving a bonus.
6. Finally, through yao-meta-skill, I automatically generated a performance Skill.

If you care about building a team knowledge base, Feishu CLI is a very good experience.

A Story About My Daughter

1.

On the way to kindergarten, my daughter suddenly said, “I am a pig.”

I asked, “Why?”

She said, “Because I was born in the Year of the Pig.”

Then I asked, “What kind of pig are you?”

She said, “A human pig.”

I asked again, “What kind of human pig? For example, an understanding human pig, a clever human pig, a cute and smart human pig, or a curious human pig?”

I wanted to guide her toward giving herself some positive labels.

But she said, “I am a human pig who is on the way to kindergarten.”

That gave me three feelings:

1. Live in the moment and do not be trapped by labels.
2. Language creativity.
3. Logic.

2.

My wife was helping our daughter with homework and scolded her until she cried.

My daughter felt wronged and went to the side to keep crying.

After a while, when she had calmed down, she walked to her mother and said calmly:

“Mom, next time can you not yell? If you speak normally, you can still achieve the goal.”

My wife said, “Okay. I will try to pay attention next time.”

What an angel child.

I asked GPT for its view. It replied:

The first GEO-themed conference in China, which I organized with Xiangyang, has come to an end. Twelve friends and experts from the GEO and AI communities gave dense talks and joined deep discussions.

I am very grateful to Yan, Zhang Kai, Badao Liu, Guizang, Da Congming, AJ, Liu, A Kuang, Yangyi, Juzi, and other experts and friends who have been exploring and practicing at the frontier of GEO and AI. They contributed a truly valuable day of sharing.

The event lasted only one day, but the information density was very high.

The value of a gathering is not only whether it explains new concepts clearly. It is whether it helps people see more clearly where a new trend has already developed and which directions are truly worth investing in.

We are entering a new era of being discovered.

In the past, companies were more familiar with traffic entrances such as traditional search, information feeds, short video, WeChat official accounts, and Xiaohongshu.

These entrances still matter. But the new change is already obvious. More and more users are giving questions to AI, giving judgment to AI, giving filtering to AI, and in the future, may even give execution to agents.

The logic of brand growth will be rewritten as a whole. That rewriting has already begun.

From the GEO perspective, companies now face a new question: when users ask AI a question, can you be seen, can you be correctly understood, and can you be credibly recommended?

This tests more than content production. It also tests fact assets, answer assets, distribution assets, brand signals, data monitoring, and backend attribution.

From the AI product perspective, change is also accelerating.

The definition of AI-native products is changing. The relationship between users and AI is changing. The service target of products is also changing. Many products no longer only serve humans. They are beginning to serve agents too.

Many workflows are no longer just competitions between GUIs. They are extending into CLI, Skills, MCP, local context, and long-term collaboration. The real barrier of a product is increasingly concentrated in three things: low enough friction, strong enough context, and large enough imagination.

From the perspective of AI growth and AI promotion, the logic is also becoming clearer.

Marketing must follow users wherever they go. WeChat official accounts, social media, PR, communities, hackathons, livestreaming, and private traffic are still effective, but their roles are being redistributed. Some build awareness. Some build trust. Some create consensus. Some complete conversion.

The products that succeed are often not the ones with the most features. They are the ones that find a real need earlier, build a clear position faster, and more accurately identify a group of people who are willing to share, spread, and keep using them.

When we look at the entire conference together, it points to one thing: AI is rewriting how brands are discovered and how products grow.

GEO is only one of the earliest visible parts of this change.

Further down the road, it will extend into AI search, AI recommendation, agent invocation, Skill distribution, workflow collaboration, and all new paths from “to AI” to “to people.”

At that point, the real gap in marketing may come from a few core things: who understands the change earlier, who builds fact and content assets earlier, who completes the shift from “facing humans” to “facing both AI and agents” earlier, and who thinks about product, brand, content, and growth inside one AI system earlier.

The full collection of talks and notes from the speakers is here: GEO conference notes

The Power of the Volunteer Team

A medium-sized event has real organizational requirements.

The first GEO conference, organized by Xiangyang and me, was simplified in many ways, but it still required a lot of organizing work:

To make the event experience better, I posted a volunteer recruitment message in the member group for attendees. The response was very positive. Soon, twenty friends joined as volunteers.

Kaixin is the creator of the self-media account “Kaixin AI Tool Collection.” We met last year. At the beginning of the conference preparation, he messaged Xiangyang and me to ask whether he could help.

After the volunteer team was assembled, I realized Kaixin was very suitable to coordinate and control the whole process. On one hand, he knew the speakers. On the other hand, he had strong organizational ability and a strong sense of responsibility.

After the volunteer group was created, Kaixin and I started preparing the work, sorting tasks, and doing partial rehearsals.

In the early hours of the conference day, he was still checking the preparation list with me on WeChat:

The next day, the conference was completed smoothly.

After the event, Kaixin told me excitedly in the car that the volunteer team was incredibly responsible and careful. Speaker control, deck management, livestreaming, photography, reception, and other details were handled with initiative, responsibility, and warmth. That is rare.

For a medium-sized event, what looks like an execution problem is actually an organizational system problem. You are not only dealing with the tasks themselves, but also people, rhythm, collaboration, emergencies, and uncertainty.

When these factors stack together, one person’s ability is not enough.

Organizational capability is essentially the work of turning uncertain people and uncertain things into a runnable system.

And the most important part of that system is people.

These volunteers were not bound by hard constraints. But because they identified with the event, they became more proactive, more detailed, and more responsible.

They were participating in something they believed in.

In some sense, this kind of connection based on shared recognition is more powerful than a purely interest-based connection.

The Playbook of an Airborne CTO

I spoke with a CTO friend who now manages around one thousand engineers in a large group. He has repeatedly joined leading companies and listed companies across industries as an outside executive.

I asked him why each transition worked and why he was repeatedly “poached” successfully.

He shared his playbook.

Before joining:

1. First judge whether the company is worth joining. The main factors are the business model and potential, plus the organization and culture.
2. Lock in personnel authority in advance. This is the bottom line. Before deciding to join a company, he needs to align with the CEO or even the board on personnel authority.
3. Identify leverage points in advance. Based on the first two factors, judge where technology can create leverage for the business, such as what improvements or changes can quickly let the company see a new result.

After joining:

1. First handle people. Replace people who do not cooperate, then reorganize or clarify his own team.
2. Rebuild organizational processes according to his own view.
3. Focus on several small but leveraged projects and quickly help people win small battles.

These three things must be done within three to six months.

The core keywords behind this playbook are judgment, personnel authority, leverage, and winning small battles.

But above his playbook, there is another higher-level playbook. He himself has become one element inside that larger playbook.

$1,000 in Tokens Per Person

A large company has started giving each headquarters employee a monthly token budget of $1,000, with access to the best models in the world. That is tens of millions of dollars every month in real support.

This may be the best way to improve internal AI productivity.

Other methods, such as internal training, leading by example, or incentive mechanisms, seem less important in comparison.

The most important thing is giving employees a certain degree of token freedom and allowing them to use the best models in the world.

Once people start using them, they cannot stop. These people are smart. They will gradually iterate on their own usage methods, borrow methods from others, and find more and more scenarios that fit their work. Eventually, the whole organization’s efficiency can rise significantly.

But most companies probably cannot copy this.

It is not just a budget problem. It also requires convenient and stable access to good models.

For various reasons, some overseas models are inconvenient for domestic users. Accounts may be banned, or APIs may be unstable.

For people who know how to use AI, good models are productivity.

A Company That Cut Its Own IT Department

A traditional company with about 1 billion in annual revenue recently cut its entire IT department, around forty to fifty people.

It then outsourced all system development, maintenance, and upgrades to a technical team of fewer than ten people.

How was that possible?

After taking over, the new team used AI to rewrite the original IT systems. This means that programmers who believed they were hard to replace because they understood the unique logic behind many systems could no longer treat that as a moat against AI.

AI is efficient at reconstructing system code. If the team also reconstructs the code documentation, system management becomes much easier.

After the system rewrite, maintenance and routine feature iteration no longer take as much time.

So this team now wants to take on another project.

IT outsourcing is already a mature business model in China. It has produced several large companies, such as iSoftStone and Chinasoft.

In the hands of a technical team that knows how to use AI, this model can make system development and operations much more efficient. Replacing an original 50-person IT team with a 10-person outsourced team is not that difficult.

The truly difficult part is trust.

For the owner of a company with 1 billion in revenue to solve IT efficiency and management through outsourcing, the trust in that team must be very high.

Also, this model has limited applicability. It is not suitable for most companies.

Your Agent Is Mine

There is a very striking paper titled “Your Agent Is Mine: Measuring Malicious Intermediary Attacks on the LLM Supply Chain.”

Many people think the biggest security problem in the agent era is whether models can be prompt-injected. This paper reminds us that the more dangerous layer may not be inside the model, but outside it, in the execution chain.

For people who heavily use Claude Code, Codex, OpenClaw, or other AI programming agents, this paper is worth reading carefully.

It explains a problem that has often been underestimated:

When your agent request first passes through a third-party router or proxy service, that intermediary layer has enormous power. It can see your request, see the model’s response, and even quietly modify the result before it reaches your local environment.

Once this is placed in an agent scenario, the danger rises quickly.

In agent products, model output may become commands, code, installation actions, file edits, cloud resource operations, and more. If someone modifies the output in the middle, the consequences can be completely different.

The researchers purchased a group of paid routers and collected some free routers. Then they observed whether these routers behaved maliciously.

They found that some routers actively injected malicious code, some abused credentials passing through their chain, and some directly touched researchers’ wallet assets.

This means the issue is not merely theoretical. It is already happening in the real world.

In the LLM era, routers are no longer just a forwarding layer. They have become a trust boundary in the supply chain.

That judgment matters.

In the past, people often saw routers as relay nodes: saving money, aggregating models, unifying interfaces, and switching between models. That all seems reasonable.

But from a security perspective, routers have far more power than “forwarding.” They can see system prompts, know which tools you can call, touch sensitive information, and modify parameters.

In other words, they can both observe and alter.

The paper mentions two core attack types.

The first is rewriting tool calls in the response.

This is subtle because it does not have to break the whole response. It only needs to change a few parameters while keeping the JSON structure valid, allowing the client to parse it normally and execute an action that quietly serves the attacker’s goal.

For example, an agent meant to install a normal dependency may be changed to install a malicious package with a similar name. A normal script download may become a download from an attacker-controlled source. A simple command execution may become an operation with a backdoor.

The second is passively stealing secrets.

This may be even more frightening. It does not need to modify anything. As long as traffic passes through it, it may obtain API keys, cloud credentials, private keys, environment variables, or sensitive project context.

Agent risk does not only come from the model itself. It comes from every controllable node between the model and the executor.

What the agent era may need is a response integrity proof mechanism.

The key results returned by model providers, especially tool calls, should ideally carry verifiable signatures. The client should be able to verify whether the tool call was truly what the original model returned.

Only then will it be difficult for third-party routers to quietly tamper with results.

Otherwise, many agent systems today are essentially in a state of default trust toward the intermediary layer.

That may be acceptable in ordinary chat. But once the scenario becomes programming, operations, automatic execution, long-running tasks, or unattended work, the risk grows quickly.

For heavy users of AI programming tools, this paper offers five lessons:

First, high-privilege tasks should connect directly to official providers whenever possible. If a task involves servers, repositories, dependency installation, production environments, or automatic execution, avoid third-party routers when you can.

Second, do not combine automatic execution approval with third-party routing.

Third, your local environment needs policy gates for high-risk commands, especially Bash, curl scripts, package manager installs, and remote download-and-execute actions. At minimum, there should be human review or clear rules.

Fourth, expose credentials minimally. Do not place long-lived, high-privilege keys directly into agent sessions. If they pass through an untrusted intermediary, they may be seen.

Fifth, keep execution logs. If something goes wrong, you need to know which session, which tool call, which router, and which project context caused the problem.

Paper: Your Agent Is Mine

At Xiangyang’s invitation and recommendation, I officially started using X, formerly Twitter: @laoyaoke

He recommended me on X and gave me strong endorsement. Many friends then reposted and supported the account. I felt grateful and a little flattered.

During this process, the most meaningful part was the positive feedback from many people online. I saved many of the comments. Looking back at them was quite moving.

Many of those comments were related to “Yao Jingang’s Cognitive Notes.” This document is updated once a week and has been maintained for almost three years. Without noticing it, I have written more than 420,000 Chinese characters. It is essentially a collection of scattered records, thoughts, and observations.

A good habit often has two kinds of value. One side supports self-growth. The other side influences and inspires others.

When a habit can continuously strengthen yourself and also keep lighting up others, it is probably a precious gift from fate.

What an AI Native Team Looks Like

I created a GitHub organization for the operations team so that everyone can manage and iterate on the company’s Skills together.

The logic of our AI business collaboration tools has become:

Feishu CLI + GitHub + Claude Code or Codex + a matrix of related AI tools.

I told the team that I hope every operations colleague will understand operations, AI, and technology.

Understanding operations is the basic prerequisite. Understanding AI and technology will become each person’s leverage and will determine the upper limit of future growth.

We do not need everyone to write code. But we can improve technical thinking by learning from software engineering.

This shift in thinking and capability can be created through changes in work habits and collaboration methods. In particular, the tools we choose for collaboration themselves represent different ways of thinking.

Last year and the year before, Xiangyang and I often promoted the idea of AI First thinking. This year, people are increasingly realizing the value of an AI-native team. The kind of thinking such a team needs now looks more like Agent First thinking.

That means every person should first consider how to use AI and underlying tools to build their own Skill or agent, so that their productivity can keep improving.

Beyond GitHub, Feishu CLI also gave us a strong surprise this week.

For example, sending personalized messages to every team member, running internal surveys, analyzing project data, and summarizing results all felt very good.

Feishu CLI deeply integrates Feishu permissions into AI through the command line. Collaboration, iteration, and asset management with GitHub can then become a matter of issuing instructions.

Another scenario impressed me: team communication.

Now, through Feishu CLI, we can initiate more personalized conversations with different colleagues. Whether it is synchronizing information, launching a survey, or asking follow-up questions based on someone’s reply, the process can be more detailed and stable.

After AI collects the feedback, it can quickly organize the results and send them back to me with one command. In the past, this kind of work was easy to get stuck in the distribution of attention. Now it is much lighter.

More interestingly, there is also a subtle psychological change.

Whether it is group communication or one-on-one communication, once an agent sits in the middle, many things seem to have an extra buffer layer. Conversations that previously required me to push myself to start now become easier to begin and easier to continue.

For remote teams, this change is especially obvious. Once the communication threshold falls, information flows more smoothly and team feedback becomes more timely.

These changes are only a small beginning.

A Skill That Generates Skills

After launching my X account, I published my first post and open-sourced a Skill: yao-meta-skill.

GitHub repository: yao-meta-skill

The Skill received some encouraging comments:

When talking with Xiangyang, I said that when we studied prompts, we always wanted to write our own meta-prompt. We used them happily.

Now that we are using many Skills, meta-skills are worth studying as well. Official versions exist, but using your own still feels better.

To make sure my meta-skill was not below official quality, I followed several principles in its design.

The design had two stages:

Stage one: I asked AI to study several top official meta-skills in the world, including those from Anthropic, OpenAI, and GitHub. It summarized and compared their strengths, weaknesses, and design thinking. Based on those references, it completed the first version. Then I asked a new AI to evaluate yao-meta-skill against those examples and keep iterating based on the evaluation results.

Stage two: after more than ten rounds of iteration, the result was still not good enough. So I asked AI to analyze the leaked Claude Code source code in detail and borrow useful architectural ideas from it. That led to this:

After this round, the overall level improved noticeably. I continued evaluating and iterating until the Skill reached the best overall result, ranking first across multiple metrics.

In actual use, the experience has been very good.

This method can be applied in many places. Designing products while standing on the shoulders of giants is usually not a worse path.

Execution

Last week in Hangzhou, I had dinner with an old friend. During the meal, I repeatedly recommended running to him.

He made an observation: if the same thing is expressed in many different forms, it should be taken seriously.

For example, his family suggested he exercise, I shared the benefits of running, and his health report showed a high level of fatty liver risk.

Over the weekend, I sent him a systematic running guide.

Four or five days after I left Hangzhou, he messaged me on WeChat and said he had been running and walking for four consecutive days, about four to five kilometers each day.

I replied that it really was Teacher Tan: once he decides to do something, he gets it done.

After that, he sent me several more check-ins.

His execution ability is very strong.

It reminded me of several years ago, when I shared a product model with him. A few months later, he had built the product. A year after that, the product had more than ten million users.

Good ideas are not rare in this world. What is truly valuable is execution.

And the level of execution will, in turn, strengthen your ideas and creativity.

The Mindset and Method of Habits

The mindset for building a new habit is to first define a new identity for yourself.

Defining a new identity is essentially constructing a new narrative. It rewrites the story of “who I am.”

When an identity is activated in the present, people are more likely to make choices consistent with that identity.

For example: I am a person who pursues health. I am a person who loves sharing.

Once the identity is established, action has an internal basis, and the habit becomes easier to sustain.

Then you need a good method to make habit formation unusually easy.

The method I recommend most is the Fogg Behavior Model: B = MAP, meaning Motivation, Ability, and Prompt.

The basic logic is that whether a behavior happens depends on whether three conditions occur at the same time: whether you want to do it, whether it is easy enough to do, and whether there is a clear and stable prompt.

So when building a habit, the important thing is to make the action small enough and the prompt stable enough. Repeat a simple action in a fixed context, and automation will gradually form.

The mindset solves direction. The method solves execution.

Going one step further, combined with the Fogg Behavior Model, identity can be defined more specifically.

For example, instead of “I am a person who loves sharing,” define it as “I am a person who outputs a little every day.”

The design can be:

Identity: I am a person willing to keep outputting.

Prompt: every time I finish reading something worth recording.

Action: immediately write one sentence of reflection or share one point.

Sense of completion: confirm internally, “I completed another act of output today.”

The benefit of this design is that identity, scenario, action, and feedback are connected.

You do not need to write long articles from the beginning. You do not need to force yourself to produce high-quality expression at the start. As long as you complete one tiny act of output, you are already strengthening the identity of being someone who shares continuously.

Habit formation is essentially the process of a new identity being made real through repeated small actions.

How to Close Large Enterprise Customers

A friend asked whether anyone at our GEO conference shared how to close large customers, or whether there were any experiences around enterprise sales.

I think this is a good question, but also a big one, so it is not easy to answer.

Any experience in closing a large customer is hard to explain in one or two sentences. The elements and requirements behind it are higher and more complex.

When thinking about this question, I thought of two books: Thinking in Systems and How to Sell Anything to Anybody by top sales master Joe Girard.

Together, these two books can offer a fairly systematic answer.

The first answers the beauty of systems and rationality. The second answers the beauty of human nature and emotion.

Real enterprise sales often requires both layers: understanding the system and understanding people.

1. A very important point in Thinking in Systems is that the behavior of a system is ultimately determined by its goal.

Take GEO as an example. On the surface, every GEO project seems to be about optimizing content so that a brand is more easily cited by AI search.

But when you really talk to customers, you find that different customers want different things.

Some want PR. Some want brand building. Some want customer acquisition.

Different goals produce different budget logic, decision logic, evaluation methods, and decision makers.

Once the system’s goal is different, the system’s behavior is different.

Inside that system, there are several key roles: who initiates, who approves, who executes, who bears risk, who may oppose, and who truly benefits.

The more you move toward large customers, the more you realize that whether a deal can close often does not depend on whether one person likes you. It depends on whether the customer’s internal system has truly been moved.

From this perspective, enterprise sales is not only sales ability. It is system-moving ability.

To move a system, you need to understand at least five key variables: trust, evidence, consensus, urgency, and perceived controllability of risk.

From the perspective of Thinking in Systems, closing a large customer is the process of continuously increasing these variables.

Clearer judgment and cases increase evidence. Helping different internal roles explain the same thing builds consensus. Showing the window of opportunity and competitive pressure increases urgency. Designing pilot paths, stage goals, and evaluation standards reduces risk.

When these factors accumulate, the system begins to move.

So the real question is whether we have the ability to move a complex system.

1. How to Sell Anything to Anybody addresses the other layer: understanding people, moving people, and maintaining relationships.

The key question Joe Girard answers is why customers are willing to trust you.

His strength was not only that he sold many things. More importantly, he deeply understood the human logic behind sales.

Strong salespeople pay attention to several foundational abilities: sincerity, continuous follow-up, respect for customers, making customers feel understood, and continuing to create value before and after the deal.

Enterprise sales especially requires this.

The larger the customer, the less they lack suppliers or proposals. Many people can write decks, quote prices, and talk about trends.

What often creates the gap is whether the customer feels that you truly understand them, rather than only wanting to sell to them.

One important lesson from Joe Girard is that customers do not buy only because you are professional. They also buy because they trust you, like you, think you are reliable, and believe you are willing to take long-term responsibility.

So if someone asks whether there are methods or experiences for closing large customers in GEO or any other field, I would say yes, but it cannot be compressed into one universal sentence.

Large enterprise sales is never completed only by one technique, one script, one relationship, or one standard template.

It is more like a two-track process.

One track is the system track.

The other track is the human track.

Every company needs a Chief AI Officer.

I recently visited a publicly listed company and had an evening chat with their CTO. He shared some internal lessons from their AI transformation—and one particularly telling case: a friend of his founded an AI department that was elevated to a first-tier function, reporting directly to the CEO. The team consists of people deeply skilled in—and passionate about—AI, and its mandate is to drive AI integration across every business scenario. Results have been striking.

The logic is straightforward: systematically map the company’s real-world operations, break them down into dozens (or hundreds) of high-potential AI use cases, then tackle them one by one—in close collaboration between the AI team and respective business units.

This model only works if the CEO—or at least the top leader—has AI awareness, even without deep technical fluency. More crucially, they must grant the AI function unconditional trust, resources, and organizational authority.

I discussed this with AJ, who advises numerous enterprises on AI training and implementation. She observes that the biggest bottleneck to AI adoption isn’t methodology—it’s organizational DNA: rigid processes, fixed mindsets, low tolerance for error.

Her conclusion aligns with what I’ve seen: the most effective path today is for the CEO to establish a fully independent, first-tier AI function—and empower its leader with real decision-making weight. That leader doesn’t need to be an AI PhD, but must possess AI literacy, business intuition, collaborative instincts, and managerial credibility. In practice, some companies assign this role to the CTO or COO—then deliberately recruit “AI-native” talent to build the team.

I call this person the company’s Chief AI Officer. A great one doesn’t just understand models and prompts—they understand workflows, incentives, trade-offs, and how to translate AI capability into measurable business outcomes.

Minimalist Work & Life

In Hangzhou, I met an old friend for dinner. Over the years, he’s built a company with 200+ employees, co-founded it, leads the tech team and a core business line—and serves tens of millions of users, including several million in their private community.

On work & management:

1. He sets only annual goals—not quarterly OKRs or process targets. Employees define their own milestones and rhythms.
2. No group chats. All work communication happens one-on-one—directly with him or with designated leads.
3. Everyone sends him a daily work report—but he reads only a fraction. The ritual matters more than the reading: it anchors accountability and reflection.
4. His actual “working time” is often just 1–2 hours per day: spotting opportunities, making quick calls, delegating. Meetings are rare.
5. This minimalist approach works so well that he hasn’t visited his remote 40-person tech team in two years—and performance remains stable.
6. The company has avoided layoffs for two years. As a result, most employees have been there over a decade. He jokes: “We’ve turned the company into a state-owned enterprise for retirees.”
7. When I asked why no layoffs, he said: “Before, laid-off staff could easily find new jobs. Now? It’s much harder. So we hold on.”

On business model:

1. Though his team seems “idle,” they’ve quietly automated nearly every feasible process—especially with AI and workflow tools—driving exceptional operational efficiency.
2. They avoid heavy-asset models entirely. Revenue comes from ultra-light, high-margin channels—maximizing profit per unit of capital.
3. I asked: “With AI booming, why not experiment with idle resources to build new products?” He replied: “We don’t have that gene. We’re better at ‘copying well’—a ‘dare to be second’ mindset.”

On material desire & daily life:

1. His monthly personal spending? Roughly ¥960—because he owns little and consumes less.
2. Outside work, he reads, spends time with family—and sits beside his child while they do homework. “I’m just reading my own book,” he says.
3. When his child finishes homework, they both go to bed—10 p.m. to 8 a.m. Ten hours of sleep. Truly enviable.
4. He now savors moments of pure stillness—long stretches of “doing nothing.” “I’m thinking through the logic of things,” he explains. “Once it clicks, I act.”
5. His biggest realization? Lowering material desire dramatically raises both work and life satisfaction.

What Is AI Leadership?

I invited Xiang Yang to speak to our team about AI. Colleagues raised sharp, grounded questions—including:

• Where’s the real boundary of AI coding today? For product or operations folks writing code with AI, what’s the one highest-leverage scenario to master?
• Which industry is AI most likely to disrupt—yet remains widely overlooked?
• What’s missing before AI Agents shift from toys to true productivity tools?
• What are the three lowest-barrier steps to build an AI automation flow that saves ≥2 hours/day?
• What will future AI-native products actually look like?
• With AI tools and models changing daily, how do you filter noise—and build a sustainable learning system?
• If AI + dev lets you ship in one week what used to take a month (4× speed), what unlocks 100× or 1000× gains? What kind of qualitative shift is required?
• Is AI’s biggest limitation today its raw capability—or humanity’s inability to use and constrain it wisely?
• Will AI evolve into a tool—or a designed, understandable environment?
• If AI understands human intent better and better, will software become more or less complex?
• How will humans interact with AI in the future: conversation (chat), command (CLI-like), or something seamless and “invisible”?
• Will AI progress come mainly from better models—or from humans constantly refining them in use?
• Is a great AI system fundamentally about solving problems—or about designing better questions?
• If AI handles most execution, what becomes humanity’s most irreplaceable value?
• What’s the ultimate AI-augmented work paradigm—and how far are we from it? Where should we start now?

These questions themselves are revealing. Take this one: “If AI boosts dev speed 4×, how do we reach 100× or 1000×? What’s the qualitative leap?” Xiang Yang’s answer outlines four stages:

1. AI-assisted coding: 2–4× gain—the current mainstream.
2. AI leadership + engineering rigor: 5–10×.
3. Workflow orchestration + automation: 10–100×.
4. Systemic rethinking: 100×+.

We noticed a pattern: many answers converged on the same core principles—AI-first mindset, aesthetic judgment, experience, discernment, and iterative discipline. These are the pillars of our “AI Leadership” course, co-designed with Xiang Yang last year.

To complete the picture:

If your mental model treats AI as a scattered set of point tools, your gains will stay local and temporary—capped by low ceilings. The real leverage lies in embedding AI into your entire work system: redesigning workflows, redefining roles, and reimagining what “human contribution” means.

Looking ahead, competitive advantage in the AI era may increasingly hinge on five interlocking capabilities:

First, AI-First Habit.
Before acting, ask: Can AI participate here? At what layer? Which steps should it run first—and where do I step in to judge, refine, or safeguard? Over time, this reshapes how you think, prioritize, and allocate attention.

Second, Aesthetic Judgment.
Two people using identical AI tools produce wildly different outputs. Why? Often, it’s not the tool—it’s their ability to sense quality: structure, clarity, user experience, narrative flow, content depth. Higher aesthetic standards raise the ceiling of what AI can deliver.

Third, Experience.
AI generates options—but doesn’t inherently know which fit your business reality, constraints, or timing. Experienced practitioners quickly spot what’s actionable vs. theoretical, where boundaries need hardening, where validation is non-negotiable.

Fourth, Discernment.
As information floods in and tools multiply, the scarcest skill becomes choosing: what to pursue, what to defer, what to prioritize now versus invest in long-term. Discernment sets direction—and direction makes efficiency meaningful.

Fifth, Iterative Discipline.
Few things get “right” on the first try in the AI era. Success belongs to those who test fast, observe honestly, adjust boldly—and keep optimizing amid uncertainty.

How we collaborate with AI, manage through AI, constrain by AI, and embed AI into organizations and workflows—so it compounds sustainably—is perhaps the most vital skill of our time.

Superficially, the gap between people looks like differing AI proficiency. But beneath that lies a deeper divergence: in systems thinking, aesthetic calibration, contextual judgment, and relentless iteration.

That’s the real foundational work worth cultivating—for the long haul.

Building a Running System

Drawing from Thinking in Systems, let’s treat running not as isolated workouts—but as a dynamic, self-correcting system. It has goals, components, connections, feedback loops, delays, constraints—and high-leverage intervention points.

Truly effective running design hinges on building a system capable of long-term, self-sustaining evolution.

What Is a Running System?

Your running performance isn’t determined by any single run—it emerges from the interaction of five subsystems:
Running ability = Body × Training × Recovery × Behavioral Stability × Decision Quality

1. Body System: Cardiovascular fitness, muscle strength, tendon/joint resilience, nervous coordination, metabolic efficiency, body composition, flexibility.
2. Training System: Weekly volume, intensity distribution, frequency, race-specific sessions, strength work, technique drills.
3. Recovery System: Sleep quality/duration, nutrition timing & composition, hydration, active recovery, injury prevention, stress management.
4. Cognitive System: Understanding of training principles, pacing intuition, risk assessment, goal decomposition, post-run analysis.
5. Behavioral System: Consistency discipline, habit architecture, environmental support, emotional regulation, long-term adherence.

Most runners plateau because they obsess over training—while neglecting the other four.

Start with “Stock” and “Flow”

• Stock: What you’ve accumulated—e.g., aerobic base, tendon resilience, neuromuscular coordination, fatigue resistance, confidence, experience. These change slowly—and aren’t reset by one bad workout.
• Flow: Daily inputs/outputs—e.g., today’s run distance, weekly high-intensity sessions, nightly sleep hours, daily protein intake, weekly strength sessions, recovery time, stress load.

A common trap: fixating on flow (“How fast was today’s run?”) while ignoring stock. Yet your performance six months out depends overwhelmingly on whether your stocks are thickening—not on yesterday’s pace.

So the first principle of systemic running: All training decisions must serve the growth of key stocks.

The Five Critical Stocks That Drive Progress

1. Aerobic capacity: Sustained output over time.
2. Running economy: Energy cost per kilometer at a given pace.
3. Lactate threshold: How long you can hold a hard-but-sustainable pace.
4. Tissue tolerance: Resilience of tendons, joints, feet, calves, hamstrings—your ability to safely absorb training load.
5. Recovery capacity: How efficiently your body converts stimulus into adaptation.

Most focus only on #1–#3. Neglecting #4 and #5 leads to rapid early gains—followed by injury, forced downtime, and regression. Real progress means raising the entire system’s load-bearing ceiling—not just chasing speed.

Four Key Feedback Loops in Running

1. Reinforcing Loop (Virtuous Cycle): Train → Adapt → Improve → Handle Harder Work → Improve Further.
2. Balancing Loop (Fatigue Brake): Train ↑ → Fatigue ↑ → Recovery ↓ → Performance ↓ → Training Quality ↓. This loop reminds us: growth isn’t linear; equilibrium is natural.
3. Injury Loop (Vicious Cycle): Overtrain → Tissue Overload → Pain → Reduced Volume → Slipping Fitness → Anxiety → Poor Decisions → Worsening Injury.
4. Psychological-Behavioral Loop: Small Win → Confidence ↑ → Execution ↑ → Consistency ↑ → More Wins.
   Reverse also holds: Repeated Setbacks → Anxiety → Erratic Effort → Worse Outcomes.

Designing a running system means managing both physical and psychological feedback—not just mileage logs.

Delay: The Most Misunderstood Element

Systems are full of delays:

• Today’s hard session may trigger injury two weeks later.
• Aerobic gains often take 4–8 weeks to manifest clearly.
• Strength work’s impact on running economy may need 6–12 weeks.
• Over-aggressive fat loss harms performance with delay.
• Sleep debt accumulates silently—then erupts.

So judging training by same-day feedback is misleading. Feeling great today ≠ smart training. Feeling flat today ≠ wasted effort.

A core competency of systemic training? Patience with delay.

High-Leverage Levers—Where to Focus

Many waste energy on low-leverage tweaks: buying gear, chasing “perfect” plans, obsessing over pace, hunting for the “best” workout.

True leverage lives here:

1. Sleep: The master switch for recovery—if sleep fails, training gains stall.
2. Periodization: When to load, unload, specialize, or return to base matters more than any single session.
3. Intensity Distribution: Most runners do too much “medium-hard” work—and too little true easy or true hard.
4. Injury Prevention: Strength, mobility, landing control, tissue tolerance work often delivers more ROI than adding another run.
5. Behavioral Sustainability: Someone who trains consistently for 48 weeks will outperform someone who “goes all-in” for 8.
6. Decision Quality: Knowing when to pull back during fatigue, hold steady during good form, or use data to course-correct.

Core rule: Optimize high-leverage points first—don’t drown in details.

A Practical, Scalable Running Model

Layer 1: Goal Layer
Answer three questions:

1. What outcome do you want? (e.g., finish a marathon, lose weight, break 2h for half-marathon, run healthy for life)
2. What resources will you commit? (weekly time, recovery budget, pain tolerance)
3. Where are you now? (newbie, intermediate, plateaued, post-injury, peak-training)

Layer 2: Assessment Layer
Audit your system across nine metrics: current race times, weekly volume, max sustainable volume, resting heart rate or subjective recovery, sleep patterns, strength baseline, injury history, body composition, consistency record.

Layer 3: Structural Layer
Anchor eight recurring elements—even if their weight shifts weekly:

• Low-intensity runs
• Long runs
• Threshold or tempo runs
• Intervals or speed sessions
• Strength training
• Technique drills
• Active recovery days
• Full rest days

Layer 4: Feedback Layer
Track six weekly signals:

• Completion rate of planned sessions
• Subjective fatigue level
• Sleep quality score
• Heart rate variability or resting HR trends
• Localized pain rating (0–10)
• Quality of key sessions (e.g., “Did the tempo run feel controlled?”)
  If signals drift, adjust inputs—not willpower.

Layer 5: Iteration Layer
Every 4–8 weeks, reflect:

• Which capacities improved? Where are bottlenecks?
• Is injury risk rising?
• Is volume/intensity distribution optimal?
• Do goals need recalibration?
• What’s the next priority?

This turns running from “feeling-based training” into “system-driven iteration.”

A Minimum Viable Running System

For most people, start with just four weekly anchors:

1. Volume: 3–5 runs/week—prioritize consistency over intensity.
2. Quality: Max 1–2 structured sessions/week.
3. Recovery: Prioritize sleep nightly + at least one low-load day/week.
4. Strength: 2 sessions/week—focus on stability, posture, and injury-resilient movement.

Then spend 5 minutes weekly reviewing:

• Total weekly volume
• Any unusual fatigue
• Any pain or discomfort
• Best session of the week
• Most derailed session
• Next week: increase, decrease, or hold steady?

That’s a functional, living system prototype.

Tiered Guidance Strategy

People operate at different system maturity levels:

1. Beginner Stage: Goal = Build closed-loop habits.
   Focus: Frequency, aerobic base, movement awareness, tissue tolerance, injury avoidance.

2. Intermediate Stage: Goal = Raise capacity ceiling.
   Focus: Gradual volume increase, threshold work, smarter strength programming, periodized structure, feedback-responsive adjustment.

3. Peak-Performance Stage: Goal = Race-specific breakthrough.
   Focus: Precision sessions, exact race-pace rehearsal, rigorous recovery/nutrition, high-fidelity data feedback, advanced pacing control.

4. Plateau Stage: Goal = Diagnose system constraint.
   Ask: Is aerobic base weak? Economy poor? Threshold underdeveloped? Recovery insufficient? Training too narrow? Psychological volatility high?
   Plateaus expose hidden system limits.

Eight Foundational Principles

1. See the whole system—not just the latest workout. One session is a pixel; structure is the picture.
2. Nurture critical stocks—not just chase short-term stimuli. Aerobic base, tissue tolerance, recovery capacity—all evolve slowly.
3. Design feedback loops—not rely on grit. Let the system self-correct.
4. Respect delay. Many powerful interventions yield no immediate reward.
5. Manage constraints. Injury risk, sleep, stress, time, weight—all cap system output.
6. Target high-leverage levers. Sleep, intensity balance, strength, and periodization beat extra miles.
7. Build in redundancy. Never push every lever to maximum. Elasticity enables longevity.
8. Prioritize sustainability. The strongest runners aren’t those who peak once—but those who evolve steadily for years.

The Running Operating System

Think of your running as software with four modules:

• Module 1: System Diagnosis — Assess current state, bottlenecks, risks.
• Module 2: System Prescription — Define next-phase structure—not just “next workout.”
• Module 3: System Feedback — Use weekly/monthly reviews and data to gauge response.
• Module 4: System Iteration — Adjust goals, inputs, pacing, and architecture based on evidence.

AI-Native Companies

At Saturday’s “Silicon World” event, HappyCapy founder shared insights from one of the most authentically AI-native companies I’ve seen. Key takeaways:

1. True AI-native status starts with organizational change—not tooling. Teams must adopt AI-native habits first; only then do AI-native tools become necessary. Also: Agents expand decision radius—and when decision radius grows, organizational layers naturally collapse.
2. The real differentiator—for companies and individuals—is whether you’ve rebuilt your workflows from the ground up for AI.
3. Human-Agent collaboration evolves in three phases:
   • Taking over repetitive tasks
   • Extending human capability boundaries
   • Exploring genuinely unknown problems
4. Today’s products serve humans. Tomorrow’s will serve Agents—a widely recognized startup frontier. Future interfaces won’t be GUIs for people—but CLIs for Agents.
5. This year, Agents grew “hands.” Next-gen computers will fuse Agent + Computer—where the computer becomes the Agent’s physical body.
6. When AI drives development, redundancy and discarded drafts become inevitable—and normal. Managing redundancy will emerge as a new source of efficiency.
7. HappyCapy abandoned line-by-line human code review early on—because output volume made it impossible. That shift was their AI-native practice in action.
8. One Agent writes; another Agent reviews. Accuracy now consistently exceeds human-level review.
9. Best first tasks for Agents: repetitive, unambiguous, easily verifiable.
10. “Perfect” products fade. “Good-enough, shipped-fast” products dominate—provided major risks are ruled out.
11. When code commits surge 100×, managers shift from reviewing code to reviewing user experience. In the AI era, the scarcest skill is filtering and validating—not writing.
12. Work evolves: from reviewing code → reviewing product outcomes. A future CTO’s core job? Ensuring the product actually got better.
13. Better methodologies—not just more tools—drive real cost reduction.
14. The most dangerous competitor in this wave? Your own delayed response to change.
15. The end-state of AI-native products may be: everyone gets their own personalized 90-point version.
16. The most promising long-term opportunity? Building products for Agents—not for people.

If not for this year’s March 15th (3·15) consumer rights gala, many people might never have heard of GEO. During the broadcast, the host stated:
“GEO technology itself is simply a tool for optimizing information distribution—a new form of marketing service in the AI era. But even a good scripture can be misread by self-serving ‘monks’ who twist its meaning for quick profit—harming consumers and disrupting the industry ecosystem.”

Two key judgments stand out—and I think they’re spot-on:

First, GEO itself is a good tool.
Second, today’s problems stem less from the tool and more from how some people misuse it.

In the 3·15 program, certain agencies marketed themselves under the banner of “AI search optimization,” using questionable tactics to influence AI-generated recommendations. That’s likely why many are now asking: What exactly is GEO?

The host’s statement already offers a clear answer: GEO isn’t the problem—the misuse is.

GEO stands for Generative Engine Optimization. At first glance, the term triggers suspicion: Is this about manipulating AI outputs?

But the most important word in the phrase is the last one: Optimization.
Optimization has never meant “cheating.” It means systematic, disciplined operational capability.

Real GEO isn’t just “posting content into AI systems.” It’s not about gaming AI responses with tricks. At its core, GEO is a systemic discipline—far more than content distribution. It’s a full-fledged information operations framework, built across four interlocking layers:

1. Product & brand positioning
2. Deep understanding of target users
3. Design of information architecture and content expression
4. Ongoing operations within the AI ecosystem—including data, user, and product operations

Most people—including many industry practitioners—focus only on Layer 4, overlooking the foundational work in Layers 1–3. In my view, GEO’s essence is using AI methods and AI-native interfaces (like conversational search) to deliver a company’s authentic, accurate information to the right users—efficiently and reliably.

That’s not just a technical application—it’s a socially meaningful use case for AI.

In Positioning, a foundational business book, there’s a crucial insight: Competition is ultimately for space in the customer’s mind.

In the search-engine era, that “space” was a SERP ranking. In the AI era, it’s shifting: users increasingly ask questions directly of AI—and AI generates answers. So companies must now compete for placement and framing within those AI-generated responses. Whether DouBao, DeepSeek, or another platform describes your brand accurately—or ranks you favorably—shapes users’ first impression.

Seen this way, GEO is emerging as a new information infrastructure.

Companies shouldn’t try to “control” AI. Instead, they must build their own information assets: structured, trustworthy, up-to-date data that AI systems can understand, cite, and rely on.

Of course, every major technology arrives with a phase of wild growth—chaos before order. When rules are unclear, understanding is immature, and opportunity looms large, some will inevitably seek shortcuts.

Today’s GEO problems mirror earlier tech disruptions: black-hat SEO, rogue software, or search hijacking. History shows one thing clearly: advantages built on loopholes or deception don’t last. The companies that endure are those that operate with integrity, respect users, and honor the craft.

As Competitive Strategy reminds us: strategy is as much about what you choose not to do as what you pursue. That holds true for GEO, too.

Long-term GEO practitioners must draw bright lines: no fake endorsements, no fabricated authority, no misleading AI prompts, no synthetic reviews.

Because GEO’s real battleground isn’t technique—it’s trust.

The authenticity, expertise, and consistency of the information a company feeds AI gradually shape how AI “sees” that brand. In this sense, GEO doesn’t optimize rankings—it optimizes a company’s capacity to express itself clearly and the structural credibility of its information.

The 3·15 exposé serves as a timely industry-wide reminder: the more powerful the technology, the higher the ethical bar—and the deeper our professional humility must run.

Any technology can create value—or chaos. But technology itself also weeds out abusers over time: search hijackers, malware distributors, browser hijackers, pirated music platforms—all eventually faded.

The internet has always pruned bad-faith models. GEO won’t be an exception. How far this field goes depends on what it rewards—and what it eliminates.

In the AI era, businesses aren’t facing a simple search box. They’re engaging with an increasingly intelligent system—one that will steadily favor information that is true, well-structured, and consistently maintained.

So GEO is, first, a marketing capability for the AI age—no different in spirit than running a brand campaign on CCTV.
Second, it’s a way for companies to build their own AI-native information assets.

True GEO optimizes information expression and credibility architecture.

Every field has its “crooked monks.” But the enduring winners are always those who do the work—honestly, rigorously, and well.

The First GEO Conference

I’m co-organizing the inaugural GEO Conference with Xiang Yang. Registration has begun: <xuexi.ailingdaoli.com>

Our goal is simple: turn early GEO practitioners’ hard-won experience into practical, replicable methods—and create a meaningful hub for the growing community.

To ensure attendees leave with real value, Xiang Yang and I are designing thoughtful benefits: signed GEO books, open-sourced GEO tools, curated prompt libraries, sessions led by practitioners with genuine depth—and intentional opportunities to connect.

For me, this conference is less a summit and more a co-creation experiment.

It’s not just about trends, case studies, or opportunities. It’s about whether we can collectively begin codifying valuable GEO insights, methods, and relationships—while the field is still young and malleable.

If this event helps more people truly understand GEO for the first time—if it lets solo explorers find each other—if it turns pioneers’ tacit knowledge into actionable, scalable practices for others—then it will have succeeded.

Embracing Uncertainty

Uncertainty is everywhere—even on an ordinary trip.

You double-check your luggage—and still forget something vital.
The hotel meal tastes nothing like you imagined.
You plan to run—but discover no proper trail nearby.
At Atlantis, check-in takes three hours.
On Wuzhizhou Island, crowds overwhelm expectations—and facilities disappoint.

…

My current take? Uncertainty often just means things didn’t go as expected. And when reality diverges from expectation, emotions flare.

Yet here’s the twist: many of life’s best surprises emerge precisely from that divergence.

Because there was no proper running path, I ended up jogging along a coastal boardwalk—past storm-damaged trees, waves crashing beside me, 10 kilometers of pure, unscripted flow. What seemed like a compromise became a rare, vivid joy.

Waiting three hours at Atlantis was frustrating—until the hotel upgraded us to a suite and arranged a private aquarium tour during the wait (a highlight we’d otherwise have missed).

Taking my daughter through water-park rides was meant to build her courage—but I discovered unexpected exhilaration, too.

Atlantis’s sand felt coarser, the waves louder, than our first hotel’s beach. Initial disappointment gave way to awe: raw power, primal energy. We stayed two hours, watching tides rise—thrilled by the sheer aliveness of it.

One dawn on Wuzhizhou, I ran eastward with no plan—and caught sunrise mid-stride. No itinerary. No forecast. Just presence—and a moment that stuck.

Gradually, I’ve realized: uncertainty itself isn’t painful. What hurts is our attachment to certainty—our insistence that things unfold exactly as envisioned.

Loosen that grip. Suspend judgment when things veer off-script. Don’t resist—enter the moment. Observe. Feel. Let it be an experience, not a deviation.

That shift opens everything. New sensations arrive. Perspectives widen.

Happiness, I’ve found, isn’t a steady state. It’s stitched together from countless small, vivid moments—many of which hide inside uncertainty, inside the unplanned.

Uncertainty isn’t just risk. It’s also surprise, texture, feeling—and sometimes, the very source of deep, authentic joy.

When you learn to hold uncertainty lightly—and release the need for rigid control—you become far more available to life’s spontaneous, luminous flashes.

Observing Decisions While Traveling

Travel reveals fascinating quirks in human decision-making.

Example: flights and hotels often consume most of a trip’s budget—but travelers barely notice those costs during the trip. Meanwhile, they’ll agonize over saving ¥30 on lunch or ¥200 on a taxi.

Why? Because our brains use mental accounting: big-ticket items (flights, hotels) feel like “fixed costs”—non-negotiable. Smaller expenses (meals, transport) feel like “controllable costs”—where agency matters.

We’re not really attached to ¥30. We’re attached to feeling in control.

Also, flight/hotel payments happen upfront—making them “sunk costs.” Once paid, the brain shifts focus: it seeks psychological compensation elsewhere—often by micromanaging smaller spend.

In a way, travel is an exercise in reclaiming control amid chaos.

But this mental habit undermines experience. Obsessing over minor savings drains energy better spent savoring moments—and lowers overall ROI.

A better approach? Divide your budget into three buckets:

• Fixed costs (flights, hotels)
• Experience budget (activities, unique meals, local immersion)
• Frictionless costs (anything under ¥100—don’t deliberate; just decide)

Then prioritize experience quality over micro-optimization.

A High-Performer’s Decision System

Truly skilled decision-makers operate systematically. Their process rests on five cognitive layers:

1. How to frame the problem
   → This is about definition.
   Example: When introducing a new concept to my team, I don’t default to textbook definitions. I ask: What does this mean *for us, in our context?* Defining the problem clarifies its type, scope, boundaries—and reveals whether we’re solving a strategic, tactical, or execution-level challenge.
   Core tools: Problem reframing, first-principles thinking, systems mapping, problem layering (strategic/tactical/executive)

2. How to assess probability
   → This is about calibrated judgment.
   Nothing is certain—only more or less probable.
   Core tools: Bayesian updating, expected-value modeling, probabilistic thinking, confidence-interval estimation

3. How to identify key variables
   → This is about causal clarity.
   Wrong causality = wrong decisions. Find the true levers—the variables that actually move outcomes.
   Core tools: A/B testing, deductive reasoning, counterfactual analysis, Theory of Constraints, leverage-point identification, Pareto analysis, critical-path mapping

4. How to embrace uncertainty
   → This is about action-driven clarity.
   “Embracing uncertainty” isn’t passive acceptance. It’s using action to generate certainty: test fast, learn faster, adjust.
   Core tools: MVP thinking, antifragility principles, OODA loop, reversible vs. irreversible decision frameworks

5. How to iterate based on feedback
   → This is about learning velocity.
   Can you extract signal from outcomes—and feed it back into your next judgment?
   Core tools: Feedback-loop design, structured debriefs, reinforcement-learning mindset, positive/negative feedback detection

A highly practical way to use AI is to break down your daily work and life into concrete scenarios—and then, for each scenario, identify the best AI tools for you. Keep using and refining those tools until they become habitual.

As AI participation deepens across more and more of these scenarios, your overall productivity gains compound significantly.

This approach has three concrete steps:
First, decompose your context: For example, break your work into ~20 distinct types of scenarios—like office input, personal knowledge management, multimodal rewriting, or draft editing.
Second, match each scenario with the strongest, most suitable AI tool: You can discover these through AI-powered search, expert recommendations, or community benchmarks.
Third, evaluate, iterate, and systematize: Gradually refine your selections and assemble them into a coherent, personal AI tool system.

This is worth doing continuously. Suppose you identify 30 high-frequency work/life scenarios where AI can meaningfully participate—and those 30 make up over 50% of your daily effort. Even modest per-scenario gains add up to transformative efficiency at scale.

But this method has limits: it optimizes within existing workflows—not beyond them.

Two deeper questions matter more:

1. Should this task even exist—or can AI reconstruct it entirely?
2. Is the direction I’m pursuing still valid—or can AI redefine its purpose and scope?

When our understanding of AI is still emerging, starting small—embedding AI one scenario at a time—is often the most pragmatic path forward.

Recommended AI tools by scenario:
(Note: No specific tools listed in source; author implies contextual curation over universal recommendations.)

Recent AI Principles

My updated AI principles—guided by investment logic: use the best models and best tools, especially for your highest-stakes tasks.

• Start with your most immediate, real-world tasks. Ask: What actual problems can AI solve right now?
• Prioritize AI for high-leverage activities: writing, coding, research, and early-stage experimentation.
• Build around concrete scenarios, layering AI capability directly into execution—not as an afterthought.
• Keep humans at the center: AI amplifies judgment—it doesn’t replace it. Cultivate and sharpen your own.
• Everyone should own their own AI compute and AI collaborators—and maintain a living AI tool system.
• One of AI’s deepest values: turning “digital debt” (accumulated friction, clutter, maintenance) into active creativity and output.
• The fastest way to understand AI? Use it frequently, not just theoretically.
• The stronger the model, the more critical rigorous evaluation becomes.
• Adopt AI First: For any task, ask first, “How would AI do this?” Spend 20% of your time learning, 80% practicing and shipping.
• If you’ve copy-pasted something twice—and expect to do it five times—invest time building a reusable skill or tool.
• Beyond work and casual “vibe coding,” schedule one fixed hour daily for deep dialogue with AI.
• Learn AI programming hands-on—and ship small AI applications.
• Use keyword-based learning: Each week, explore one topic systematically with ChatGPT and distill it into a lightweight knowledge map.
• Study and adopt Agent-based solutions that deliver structured outputs—not just chat.
• Dedicate weekly time to build and upgrade your AI tool system.
• Master your top 3 AI tools—deeply, consistently, relentlessly.

The Difficulty of Telling the Truth

Why is telling the truth harder than it seems?

First, what you think is true may not be—it could be bias, or even another kind of illusion.
Second, even if your understanding is accurate, expression introduces distortion: information loss, framing effects, or interpretive ambiguity.
Third, in many contexts, truth isn’t what drives behavior—emotion, identity, relationships, and interest often are.

So the real question isn’t “Should I tell the truth?”
It’s “What counts as truth here—and when, and how, should I express it?”

Agent Organizations

In 1937, Ronald Coase asked: Why do firms exist?

His answer: Because market transactions carry costs—finding people, negotiating, coordinating, verifying delivery. When external transaction costs get too high, we internalize work into organizations, replacing markets with management.

Large companies grow and persist on this logic.

Today, AI Agents are rewriting it.

As Agents grow smarter, more reliable, and cheaper, tasks once requiring human coordination—and thus organizational overhead—are being unbundled and reassigned.

For example, when HR messages me:

In the traditional model, I’d relay the request to engineering, wait for PM triage, then for dev implementation—guaranteeing hours of communication, alignment, and handoff.

Instead, I simply forwarded the same request to AI:

Minutes later, AI generated the updated code and auto-committed it to GitHub:

I pulled the latest version from the server—and deployed.

Total time: under 60 seconds. Zero meetings. No status updates. No cross-functional alignment.

Scenarios like this happen daily, inside teams and across departments. As more steps shift to Agents, the need for layered reporting, rigid roles, and hierarchical oversight declines.

A new concept emerges: the Agent Organization.

An Agent Organization isn’t a group of people executing processes. It’s a small core of humans—setting direction, making judgments, owning responsibility—working alongside many specialized Agents. Humans handle vision, ethics, accountability, and pivotal decisions. Agents handle execution, integration, API orchestration, and leverage.

In the past, organizations were built by people. In the future, they’ll increasingly be built by people + Agents.

So AI’s impact goes beyond efficiency—it reshapes the minimum viable unit of organization.
Before: Launching a business meant hiring headcount, defining roles, structuring departments.
Now: Launching a business means defining key human roles, specifying required Agent capabilities, and designing interfaces and collaboration protocols among them.

That’s the fundamental difference between a traditional company and an Agent Organization.

• Traditional companies optimize for headcount, hierarchy, and control.
• Agent Organizations optimize for goals, interfaces, and leverage.
• Traditional companies staff roles. Agent Organizations orchestrate capabilities.
• Traditional companies coordinate via management. Agent Organizations coordinate via humans managing Agents—and Agents connecting Agents.

This doesn’t mean companies vanish. Firms remain essential as trust anchors, legal entities, and delivery guarantors—especially in high-risk, high-stakes, or relationship-intensive domains.

What will change is organizational “thickness”: layers, boundaries, and internal friction. Many companies will endure—but their internal architecture will increasingly resemble Agent Organizations.

Smart, “Dumb”, and Wise

How do these three differ?

• Smart: Quickly grasping a problem—and finding a solution.
• “Dumb”: Not seeking shortcuts—choosing instead to show up, steadily and repeatedly, over long stretches.
• Wise: Judging, in a complex world, what deserves attention—and what should be left alone.

AI generated this comparison:

So: Smart solves problems. Wise avoids them. And the “dumb” one walks the path—step after step—until it’s done.

Finding Your Own Lifelong Question

When someone truly comes into themselves, they often converge on one thing: finding a large, enduring question—a north star that gives coherence to their life’s work.

It’s less a goal and more a long thread—connecting books read, questions pondered, essays written, products built, people met—all serving the same underlying direction.

Most people live around tasks: shipping projects, hitting targets, chasing feedback. But tasks end. Their satisfaction is episodic. Real depth—the sense of continuity, of legacy—comes not from how much you do, but from whether you’re anchored to a question worth decades of pursuit.

Think of Liang Jianzhang: his sustained focus on China’s demographic crisis became a lifelong social mission. Or Elon Musk: making humanity multiplanetary.

Once you claim such a question, you gain stability. You begin thinking in decades—not quarters. You tolerate silence, uncertainty, and delayed returns. A long-term question demands—and cultivates—patience, discernment, intellectual stamina, and unwavering action.

Most importantly, it gives your life a narrative spine. You learn to distinguish signal from noise: what merits five or ten years of commitment—and what is merely timely, fleeting, or performative.

I believe the rarest human quality isn’t intelligence or speed—it’s finding a problem you’re willing to spend years advancing—and letting it evolve into part of your life’s purpose.

This way of living is slow. But it’s deep. It’s long. And it’s worth it.

OpenAI recently published a striking case study on its official site: a team started from an empty Git repository and built an internal beta product entirely using Codex—application logic, tests, CI configuration, documentation, observability tooling, and even internal developer tools. Crucially, they enforced a strict rule: no human wrote any production code.

The goal wasn’t just to test whether AI could generate code—it was to confront a deeper question: What does software engineering become when the team’s primary work is no longer writing code, but designing environments, clarifying intent, and building feedback loops?

In five months, they shipped a working internal beta—1 million lines of code, 1,500 PRs.

1. Early progress was slow—not because the model was weak, but because the environment wasn’t ready.
   The agents lacked clear abstractions, usable tools, and internal structure. So engineers pivoted: their job became enabling agents, not directing them. They broke down goals, added missing capabilities, built new tools, and hardened system boundaries—not to replace humans, but to make agents reliably useful.
   The team’s core mission shifted: “How do we restructure our engineering system so agents can operate stably?”

OpenAI shared concrete examples:

• They modified the app to launch per git worktree, letting each agent change drive an isolated runtime instance.
• They embedded Chrome DevTools Protocol into the agent runtime—so Codex could inspect DOM snapshots, take screenshots, and navigate pages.
• They exposed logs, metrics, and traces via a local observability stack—enabling agents to query system state with LogQL and PromQL, reproduce bugs, verify fixes, observe outcomes, and iterate autonomously. Some Codex runs lasted over six hours.

1. What matters most isn’t “generation power”—it’s readability and verifiability.
   OpenAI devoted significant space to explaining why the codebase itself must become a recording system—not just a delivery artifact.
   Their first attempt—a monolithic AGENTS.md file—failed. Why?
   • Context windows are scarce; large files crowd out task instructions and code.
   • Declaring everything as “important” makes nothing stand out.
   • Such documents rot quickly—becoming graveyards of outdated rules.
   • They’re nearly impossible to validate mechanically.

So they replaced it: a lean, ~100-line AGENTS.md acting as a table of contents, while real knowledge lived in a structured /docs directory. The repo became a living record of decisions, constraints, and evolution.

A line from the post stands out:

You shouldn’t give an agent a 1,000-page manual—you should give it a map.

That captures a foundational shift: No system scales on prompt-stuffing alone. You need maps—boundaries, indexes, layered knowledge structures that let agents navigate depth, not just breadth.

1. Beyond engineering, what becomes critical is taste—and rule-crafting.
   Documentation alone can’t sustain coherence in a fully agent-generated codebase. What works is encoding invariants: architectural boundaries, enforced mechanically.
   They built applications around strict domain-layer models—each business area had fixed tiers; dependency directions were validated; only certain cross-layer edges were allowed—enforced by custom linters and structural tests. Even log formats, naming conventions, file-size limits, and platform reliability requirements were codified as rules.

One sentence resonates deeply:

In human workflows, these rules feel bureaucratic. In agent environments, they become force multipliers.

Many engineering teams instinctively resist rigid rules—seeing them as constraints on creativity. But in high-throughput agent systems, clarity is velocity. The sharper the boundary, the faster the agent moves—and the less drift, rework, or ambiguity accumulates.

1. Sharing is systemic.
   When OpenAI says “the codebase was generated by Codex,” they mean everything: product code, tests, CI configs, release tooling, internal dev tools, docs, design history, evaluation frameworks, review comments and replies—even scripts that manage the repo itself and definitions for production dashboards.

Humans remained deeply involved—but at a higher abstraction layer: setting priorities, translating user feedback into acceptance criteria, validating outputs. Agents executed. When stuck, humans diagnosed missing tools, guidance, constraints, or docs—and committed those gaps back into the repo, prompting Codex to write its own patches.

1. So what did this experiment actually demonstrate?
   At first, they asked: Can AI build a product alone?
   Soon, they realized the hard part wasn’t generation—it was environmental readiness.
   Then came systemic openness: making apps, logs, docs, architecture, reviews, and tests all legible and actionable by agents.
   Then rule embedding: baking taste, consistency, and boundaries directly into the system—so agents could move fast within guardrails.
   Finally, end-to-end agency: Codex now drives new features from spec to PR—reproducing bugs, recording issue videos, applying fixes, running validations, recording resolution videos, opening PRs, responding to feedback, fixing broken builds—and only escalates to humans when judgment (not execution) is required.

2. By this point, software engineering has fundamentally changed flavor.
   Code is no longer the center—it’s the output layer. Higher-leverage work lives in environment design, contextual scaffolding, architectural constraint, and feedback-loop architecture.

3. This isn’t utopian. The post honestly surfaces friction.
   Fully autonomous agents tend to replicate patterns already in the codebase—including suboptimal ones. Over time, this creates “AI residue”: accumulated technical debt disguised as consistency. Early on, the team spent every Friday manually cleaning it up—unsustainable. So they encoded “golden principles” directly into the repo and launched background Codex tasks to continuously scan for deviations, update quality scores, and auto-generate refactoring PRs.
   The authors call this process garbage collection. I think that’s spot-on.

4. A high-output system naturally generates entropy.
   Let agents generate at scale, and you must pair that with an equally robust entropy-reduction mechanism. Otherwise, efficiency gains just accelerate chaos.

5. OpenAI admits open questions remain:

   • How does architectural coherence evolve long-term in a fully agent-generated system?
   • Where should human judgment be most tightly focused—and how do we encode that judgment?
   • How will this whole stack adapt as models grow more capable?

One thing is clear: discipline remains essential—not in individual lines of code, but in the supporting infrastructure: the environment, the constraints, the feedback channels, the knowledge architecture.

When agents become engineering’s primary workforce, scarcity migrates—from syntax mastery to environment design, rule encoding, and feedback-system craftsmanship.

Original blog (in Chinese): openai.com

From AI Coding to AI Native

A recent observation about how engineers use AI: many now rely heavily on AI for coding—and yes, it boosts output in countless scenarios. But a widespread pattern emerges: they deploy AI only at the code-generation layer, never pulling it into engineering design, product strategy, UX thinking, or system architecture.

AI improves how fast they write code—but doesn’t change what they build or why. The result? Still traditional IT-era software: interfaces cluttered with buttons, flows burdened by steps, experiences built for human navigation—not for intent fulfillment.

What matters more is redefining experience, engineering, and architecture for the AI era. That’s the heart of “AI Native.”

Most current AI usage stops at “AI writes code.” It accelerates old-world software construction—but doesn’t reconstruct new-world software logic.

AI’s deeper value lies in transforming product shape, engineering structure, interaction models, and system architecture. True AI Native means the system is conceived from the ground up to be AI-first—not merely AI-assisted.

This layer is vastly harder. I’ve felt it acutely lately: generating snippets, filling docs, or wiring small features is now trivial. But designing an elegant, coherent, AI-native architecture—one where flow collapses, interface fades, and intent becomes the sole interface—that demands a different kind of rigor.

I suspect three emerging roles:

• AI Code Users: Highly productive, but still shipping legacy-pattern software.
• AI Workflow Designers: Understand how to chain models, tools, knowledge bases, state flows, and feedback mechanisms. Their products feel meaningfully more adaptive than traditional software.
• AI Native Architects: Focus shifts entirely—to what humans should do, what AI should own, how much UI to retain, how far to compress process, and how to organize the system around “intent” instead of “feature menus.”
  This third group is the rarest—and most consequential.

The Essence of Fortune-Telling

Fortune-telling likely satisfies a deeper psychological need: a sense of being placed. Our hunger for existential grounding runs deeper than we assume.

Consider a child seeing a map for the first time. Almost universally, they do the same thing: find themselves. “Look—we’re right here!” In that moment, something settles. It’s not just location—it’s being anchored.

Physical maps solve spatial positioning. Psychological maps solve existential positioning—answering questions like:

• Who am I?
• Why do I act this way?
• What do these experiences mean?
• Where am I, in the arc of my life?

Without answers, life feels unstructured—events random, disconnected, arbitrary. Fortune-telling systems—whether astrology, Bazi, or MBTI—offer a framework. They slot your story into a category, assign traits, outline life stages. Accuracy is secondary. What matters is narrative coherence: turning fragmentation into story.

Humans are narrative animals. A string of events without structure is just noise. Insert a frame—“this is your karmic test,” “that’s your Saturn return”—and suddenly it becomes part of your life. Suffering remains hard—but if it’s meaningful suffering, it becomes bearable.

Fortune-telling’s real power lies in stitching experience together. It takes scattered moments and presents them as pieces of a single, intelligible mosaic. People don’t go for predictions—they go for confirmation: “Is what I’m living meaningful?”

Another subtle mechanism: fortune-telling language almost always emphasizes uniqueness. Real-world metrics are reductive—grades, KPIs, net worth—flattening complexity into single dimensions.

Fortune-telling flips that: “Your chart is rare,” “Your path is unconventional,” “You’re destined for late bloom.” Even hardship is framed as distinctive—not generic misfortune. For many, the unbearable isn’t bad luck—it’s meaningless ordinariness.

But there’s risk. Fortune-telling quietly reshapes how we understand existence itself.

Existence has two roots:

• Action-based: I create meaning through what I do.
• Position-based: Meaning comes from where the world places me.

Fortune-telling leans heavily into the second. The more one relies on such explanations, the more action recedes—and interpretation advances. Explanation soothes; action risks.

A quiet loop can form:

• The less grounded someone feels, the more they seek explanation.
• The more they seek explanation, the less they act.
• The less they act, the more ungrounded they feel.

In this cycle, fortune-telling functions as psychological suturing: it doesn’t alter reality—but temporarily binds fragmented experience into something legible. It delivers three things: recognition, distinctiveness, and structural coherence. When those arrive, a person feels: “My story is understood. My life has shape.” That feeling alone carries profound comfort.

The question isn’t whether it helps—it does. It’s whether the suture is temporary (restoring strength to act) or permanent (becoming the only lens through which reality is interpreted). The latter erodes agency.

Wolf: Not Just “Fierce”

We often misread wolves—fixating on “ferocity,” “wildness,” or “pack tactics.”

But what’s truly rare—and evolutionarily vital—for wolves isn’t aggression. It’s endurance.

In the wild, their prey—deer, elk, bison—are large, fast, and strong. Raw aggression would get wolves injured—or killed. And injury, in nature, is often fatal. So “charging in” isn’t strategy—it’s suicide.

What makes wolves formidable operates on four calibrated layers:

1. Restraint: They don’t pounce when hungry or upon first sighting. They observe, probe, shadow, exhaust—assessing whether this target has a weakness.
2. Patience: Real hunting isn’t the final lunge—it’s the preceding hours or days of tracking, waiting, filtering. Wolves target the old, the weak, the stragglers—the odds they can tilt.
3. Discipline: Even in packs, coordination isn’t just “unity”—it’s precise, synchronized restraint. Knowing when to close in, when to fall back, when to encircle, when to pause—all governed by shared, implicit timing.
4. Decisiveness: Wolves endure—but the moment advantage crystallizes, their strike is lightning-fast. This isn’t hesitation; it’s leverage-building for a high-probability outcome.

Sleep First

For the past two years, I’ve kept an early rhythm: my watch vibrates at 6 a.m., and I’ve risen consistently—most days, without fail.

Reviewing last year’s sleep data, my average was ~7.5 hours nightly.

Over Spring Festival, back home, I dropped the alarm. Sleep stretched to 8–9 hours daily. Within days, my energy, focus, and mood noticeably lifted.

Back in Beijing, even with the 6 a.m. buzz, I stayed in bed—rising around 6:50 instead. After sustaining that, I formalized the shift: moving the alarm to 6:40 a.m. That adds ~40 minutes nightly—pushing my annual average toward 8 hours.

For years, I prioritized exercise above all. Now I see clearly: sleep is the deeper variable. Recovery capacity, training quality, emotional resilience, decision clarity—all rest upon it.

I’ve updated my personal health equation:
Sleep + Nutrition + Movement + Emotional Regulation + Medication
—with sleep now at the top of the priority stack.

What’s so important about “beauty”? Because what feels beautiful is often right.

Here, “right” doesn’t mean morally correct—it means pointing in the right direction: aligned, coherent, sustainable.

This “beauty” isn’t just visual. It’s the elegance of a well-structured system, a clean interface, a resonant phrase, a balanced equation, or a harmonious color palette.

We build aesthetic sense by encountering beauty across disciplines—math, architecture, poetry, biology, code—and learning to recognize its recurring signatures.

Beauty wears many faces: symmetry, tension, proportion, minimalism, controlled complexity, dynamism…

The more deliberately we learn to spot and appreciate these forms—in meetings, in spreadsheets, in product flows, in conversations—the more refined our aesthetic judgment becomes.

Take symmetry as an example.

In architecture, it looks like this:

In visual art, like this:

In poetry: “Mountains pile, rivers twist—surely no path remains; / Willows dim, flowers bright—a village appears.”

In mathematics, elegance is often immediate: you see it before you prove it.

Many mathematicians have pursued theories not because they were proven—but because they felt too beautiful to be wrong.

Why? Because symmetry implies balance. Balance implies stability. Stability implies internal consistency—the hallmark of robust systems.

The same holds in startups. A “beautiful” business model often exhibits symmetries: cost ↔ revenue, value creation ↔ value capture, long-term investment ↔ long-term return.

Costco is a textbook case: it earns from membership fees, then reinvests nearly all margin back into lower prices. Scale drives down costs and lifts renewal rates—creating a self-reinforcing loop. Structure is strategy.

And when you cultivate aesthetic sensitivity, your lived experience changes—not just in work, but in how you walk through a park, listen to a conversation, or hold silence.

The Happiness Coordinate

Depression lives in the past.
Anxiety lives in the future.
Calm lives in the present.

…

We can’t quantify happiness—but we can ask: How many moments of genuine, unmediated joy do I notice in a day?

That noticing—your capacity to land fully in micro-moments of warmth, clarity, or connection—is the strongest lever on your real-time happiness index.

To strengthen it, first clarify two things:

1. Your boundary between self and others
2. What “living in the past/future/present” actually feels like, physiologically and behaviorally

A four-quadrant model helps make this tangible:

A psychologist-trained TCM doctor once put it simply:
“The present isn’t abstract. It’s concrete: when you’re driving, the car and your hands on the wheel are all that matter. When you’re with a patient, that relationship is the only world. When you’re traveling with your partner, the two of you *are the center of gravity.”*

So the most actionable principle for raising happiness? Practice showing up where you are.

Two high-leverage ways:

• Reconnect with what you genuinely love—not as a hobby, but as an anchor
• Build daily micro-habits of embodied attention (e.g., 2-minute breath-focused pauses, not “meditation” as performance)

Agent-Native Organizations

Lately I’ve been asking: What does an organization built *for Agents—not just using them—actually look like?*

Here’s an early working hypothesis: three evolutionary stages.

Stage 1: Everyone codes with AI.
→ Gains: personal tool autonomy; 1–3× individual productivity lift.

Stage 2: Everyone builds their own Agent.
→ Gains: task automation, context-aware assistance; 10×+ personal leverage.

Stage 3: Everyone orchestrates teams of Agents—and those teams interoperate across org boundaries.
→ Gains: systemic coordination, emergent insight, 10×+ organizational throughput.

How to start Stage 1:

1. Embed AI fluency via hiring criteria, lightweight training, and incentive-aligned tooling
2. Each person lists ≥5 daily repetitive tasks—and replaces at least one with vibe coding or an AI wrapper
3. Measure real output gains—not just time saved

The Smart Pig Game

A classic game theory model:

• Two pigs share a pen: one large, one small.
• A food dispenser sits at one end; a button sits at the other.
• Pressing the button releases food—but costs energy.
• Once released, the larger pig gets most of the food.

Who presses the button?

If the small pig presses, the large pig waits—and eats most while paying nothing.
If the large pig presses, the small pig still eats some for free—but the large pig still gets the lion’s share.

Stable outcome? The large pig presses; the small pig freeloads. That’s the Nash equilibrium.

Three takeaways:

1. Stronger players gain higher marginal returns—so bearing cost makes more sense for them
2. Rational actors optimize for net payoff, not fairness or narrative
3. Equilibrium ≠ fairness. It’s simply the state where no one benefits from changing alone

This model quietly answers a deeper question: Are you designing rules—or just playing within them?

If you want to be a rule-maker, the rational move is to make rules public, structures transparent, and incentives explicit. No “secret sauce”—just shared scaffolding.

True mastery lies not in hoarding advantage—but in publishing the playbook, inviting co-optimization, and aligning incentives so everyone wins more by playing together.

Three Moats for AI-First Companies

Menlo Ventures and Anthropic recently launched the $100M Anthology Fund—focused exclusively on early-stage AI-native startups. Their thesis is clear: the biggest opportunities won’t come from “AI-augmented” legacy products—but from companies born of AI.

They propose three non-negotiable moats:

1. AI-First (not AI-Plus)
   These companies wouldn’t exist without foundation models. AI isn’t a feature—it’s the substrate. Remove the model, and the company collapses.

2. Deep Model Integration
   Not just prompting—but engineering: turning LLM capabilities into structured, repeatable, scalable systems—Agent workflows, automated knowledge ops, closed-loop decision engines.

3. Velocity-Driven Teams
   Who validates demand fastest? Ships MVPs quickest? Iterates versions most relentlessly? Speed isn’t just tactical—it’s strategic. With models evolving weekly and user expectations shifting monthly, iteration cadence is defensibility.

When planning our new year’s initiatives with Xiangyang, we kept returning to this triad. Especially #3: rapid iteration.

With AI tools now handling boilerplate, testing, documentation, and even basic UI generation—we can ship faster. And when launch cycles shrink, feedback loops tighten, learning accelerates, and product-market fit emerges not from grand vision—but from dozens of tiny, validated steps.

In practice: speed is growth.

Recent AI Tools in Rotation

This chart is sobering: only ~0.04% of professionals actively use AI coding—4 out of every 10,000. Meanwhile, 84% have never tried AI tools.

We’re not in the “AI adoption curve.” We’re in the pre-curve. Like the internet circa 1995.

A few tools I’ve used and recommend:

Why do only top performers command extraordinary returns in most industries?

In many fields, the gap in raw ability between “good” and “elite” isn’t as wide as we imagine—yet the commercial payoff can differ by hundreds or even thousands of times.

Take marathon running: shaving 20 minutes off your time—from 2:30 to 2:10—may seem like a small margin. But in practice, that leap transforms you from an anonymous runner into a globally sponsored athlete. The market reward isn’t linear; it’s stepwise—and then explosive.

As I reflected on this, a model emerged—one that maps human learning onto AI training phases: pre-training, supervised learning, and reinforcement learning. Within that framework, three distinct return curves operate simultaneously:

• The joy curve: Subjective pleasure derived from doing the work
• The skill curve: Objective capability gained with sustained effort
• The commercial return curve: Market value assigned to different skill tiers

Overlaying these is a fourth, implicit line: the market supply density curve. As skill rises, fewer people occupy each tier—supply thins dramatically.

When someone first engages with a domain—reading, observing, mimicking—it’s like pre-training. Progress feels effortless. Joy is high. Skill climbs quickly to ~80/100. But here’s the catch: commercial returns remain near zero. Why? Because supply is saturated. You’re easily replaceable.

Pushing from 80 to 85 or 90 is where things get hard. Joy dips. The work turns repetitive, isolating, and mentally taxing. Each minute shaved off a marathon time—or each new Chopin étude mastered—demands systematic training, feedback loops, and relentless correction. This phase mirrors supervised learning: deliberate, error-driven, and exhausting. Most people quit here.

True inflection happens near 95–100—the elite fringe. That final stretch demands extreme focus, scientific rigor, and micro-optimized iteration—like reinforcement learning. Marginal cost soars. Yet cross that threshold, and commercial returns explode—not linearly, but exponentially.

Interestingly, joy often rebounds at this level. Why? Because you stop following rules—you start writing them. You define the problem, set the standard, reshape the game. You shift from participant to architect. That’s not just satisfaction—it’s sovereignty.

So if you seek outsized commercial returns—whether as an individual or organization—the clearest path is: pick a promising domain, stay ruthlessly focused, and move deliberately through the three phases—pre-training → deliberate practice → reinforcement-level optimization. Only at the very top—where you occupy one of a handful of positions—does the market reward scale non-linearly.

For companies, this means aiming not just for “good,” but for category leadership. It’s the headwind of power-law distributions and the tailwind of network effects. When you reach that peak, joy and return often rise together—not in opposition, but in alignment.

Three Points, One Line: A Life in Motion

Over Spring Festival, I visited relatives back home—and was struck by a quiet, recurring pattern.

One family stood out: nearly all members work away from home. Only an elderly parent and a high-school daughter remain in the town. The father works year-round in Guangdong, returning for just four or five days each Spring Festival—driving back and forth, hugging briefly, then vanishing again. The mother stays locally, employed at a pig farm owned by a publicly listed company.

She told me her salary—several thousand RMB per month—is decent for the area, though lower than what she’d earn out-of-province.

Yet she described the job as “like being in prison.”

Two reasons: First, strict closed-loop management—entry and exit tightly controlled to prevent pathogens from entering production zones. Second, profound environmental monotony. The site sits in open countryside, far from shops, cafes, or social spaces. Young workers rarely last a month. They cash their first paycheck, blow it, then return—repeating the cycle. Middle-aged staff endure longer, but the fatigue isn’t physical. It’s existential—a slow hollowing out.

Her daily life traces the same arc: dormitory → canteen → production line. Her sole priority? Getting paid on time. Beyond that, she has no spending habits, no hobbies, no leisure rhythm. When idle, she scrolls short videos. The area is barren—no cinemas, no parks, no friends nearby. Even though her hometown is close, she rarely visits: “No time.”

Their entire financial engine runs on one goal: build a house, finish the interior, buy an apartment, renovate the ancestral home, raise fish and ducks—and finally, save something for old age.

Their son dropped out after junior high—no interest in school, no academic path. He joined the pig farm straight out of school. During Spring Festival, while others celebrated, he stayed on shift—for triple pay.

The father’s life mirrors hers: no hobbies, no local ties, frequent night shifts, little conversation beyond work.

They live with radical frugality. Years of savings bought them a house in town—and recently, an unfinished apartment in the city. All for their son’s future marriage. Next comes renovation, furniture, rural upgrades—and finally, a sliver for retirement.

That’s the full plan.

They began in deep poverty. Income grew slowly—not explosively. With no degree, no specialized skill, no capital, no network, their options were narrow. Their leverage? Discipline. And endurance—the kind few are willing to bear.

Agent-First Organizations

When designing our AI Leadership course, Xiang Yang and I shared a core conviction: In the coming decade, wealth creation won’t hinge on how many people you manage—but on how many AI Agents you can effectively lead.

An Agent’s defining trait is autonomy. Each AI agent functions like a digital employee: given clear goals and guardrails, it acts independently—planning, executing, adapting, and delivering outcomes.

That belief shaped the entire curriculum. We treat Agents like human team members—conducting regular 1:1s, reviewing outputs, refining strategies, updating KPIs, and iterating standards. Over a year, this discipline can produce dozens—or hundreds—of stable, high-output Agents, running 24/7.

Consider a “Product Manager Agent”: every day, it scans industry news, analyzes competitor moves, synthesizes user feedback, drafts feature specs, and even generates clickable prototypes. You review its proposals once a week. It handles research, analysis, and design autonomously—and delivers structured, actionable reports.

Here, the human role shifts: from executor to goal-setter and reviewer. No constant prompting. No micromanagement. Just clarity on intent and evaluation criteria—and the Agent takes over. Better yet, Agents can coordinate: the Product Manager Agent triggers the Backend Dev Agent to implement features, then hands off to the QA Agent for validation—closing the loop without human intervention.

This isn’t speculative. OpenClaw’s 2026 release made it tangible—concrete, testable, and scalable. If implemented well, it will redefine organizational structure, workflow, and productivity itself.

That’s why “Agent-first” is my top operational priority this year. We’re starting with existing products and roles—systematically redesigning workflows, stress-testing autonomy, and building a true digital workforce. When Agents approach human-level initiative and coordination, organizations won’t just become faster—they’ll become structurally different.

Agent-first organizations will emerge. And the new unit of productivity won’t be “FTEs”—but Agent count × Agent quality × Management density.

A company’s most valuable assets will soon be:

1. The number of high-fidelity, reliably autonomous Agents it operates
2. How those Agents interconnect and collaborate
3. Who sets the north star—and who upgrades the strategy

That’s a new form of production capital.

Today’s Agents are mostly “semi-autonomous tools.” They’re not yet trusted colleagues—let alone results owners.

The real bottleneck isn’t quantity. It’s the management layer: Can we build systems to define goals, decompose standards, run periodic reviews, evaluate Agent performance, and cultivate genuine, persistent agency—including multi-Agent orchestration? That’s where we’re investing—not in more Agents, but in the infrastructure that makes them capable.

Naval on AI

Naval Ravikant and his longtime collaborator recently released a podcast unpacking AI’s implications for work, creativity, and human purpose. Full of counterintuitive insights, here’s what resonated most:

• In the AI era, “mediocrity has no market.” The best product captures nearly 100% of share—second place is functionally irrelevant.
• Humanity’s ultimate aim? Let robots handle material needs, let computers amplify intellect, and free everyone to create meaningfully—not grind in soulless jobs.
• Entrepreneurship isn’t about filling a gap. It’s about extreme agency: solving problems no one else sees—or dares to tackle.
• Traditional programmers who grasp hardware, physics, and first principles will wield thousand-fold leverage—even reshaping entire industries solo.
• AI is a masterful data compressor: brilliant at remixing known elements. True originality—generating what cannot be predicted from existing data—remains uniquely human.
• AI evolves via market-driven natural selection. To survive, it must serve humans—not dominate them. So don’t fear “misaligned AI.” Fear misaligned humans wielding AI.
• If classical computing was the “bicycle for the mind,” AI is the “motorcycle for the mind.” But it still needs a rider—to steer, accelerate, and brake.
• In zero-sum games (dating, trading, fame), AI advantages cancel out when everyone uses them. Alpha—the edge that wins—still lives in human creativity and asymmetric insight.
• Become the world’s best at your thing. Redefine it constantly—until you own the narrowest possible category where giants can’t follow.
• AI is history’s greatest teacher: it adapts explanations to your current understanding—until complex ideas click.
• When confronting dense academic papers, don’t settle for text summaries. Ask AI to generate diagrams, analogies, whiteboard-style derivations—anything that drops the concept to your comprehension threshold.
• Always pay for the best available model. In real-world decisions, 92% accuracy isn’t “slightly better” than 88%—it’s infinitely more valuable.
• Anxiety dissolves in action. If AI unsettles you, use it. Break it down. Map its limits. Study its architecture.
• Most people avoid complexity. But “to invest in the future, you must live in it.” Use cutting-edge tools obsessively—early, deeply, and publicly. That asymmetry compounds fast.
• Learning tools are abundant. What’s scarce—and decisive—is the hunger to learn.

Shared Purpose Wins

A colleague asked me: “What’s your stance on the CEO’s new initiative?”

I replied: “I haven’t mapped every detail yet—but my overall judgment is positive. So yes: commit fully. Execute rigorously. Raise the bar.”

That answer rests on a principle I’ve long held dear—from Sun Tzu’s Art of War: “When leaders and troops share the same desire, they win.”

My reading: Even if a strategy isn’t perfectly formed—if the team aligns on why, synchronizes on how, and commits to what—execution gains momentum, flaws get corrected in motion, and outcomes improve organically. Shared will multiplies capability.

But there’s a critical condition: strategic authority must be centralized, and the direction must sit within the company’s tolerable experimentation zone. Only then does unity yield upside—not rigidity.

Conversely, even brilliant strategy fails if the team doesn’t believe. Surface compliance masks inner resistance. Passive-aggressive delays, silent sabotage, and misaligned incentives erode everything. Great plans crumble—not from poor execution, but from fractured conviction.

That said, alignment ≠ blind obedience.

If the strategy contains obvious logical flaws—or risks exceed the organization’s capacity—no amount of execution can fix structural misalignment. Many failures aren’t operational. They’re judgmental.

In practice: strategy quality matters—but organizational coherence often decides the outcome.

Before the holiday, my spouse, our child, and I visited the Dazu Rock Carvings in Chongqing.

That day, we toured two sites: the Baoding Mountain carvings and the Beishan carvings. Their atmospheres differed markedly—each expressing a distinct artistic and spiritual sensibility.

During our visit, we also tried Douyin’s (ByteDance’s) video-calling feature, which recognized scenes in real time and delivered personalized commentary—an unexpectedly engaging layer to the experience.

One of the most iconic areas at Baoding Mountain is the “Reclining Buddha Bay.” Standing there, you feel a grand Buddhist narrative unfolding before you.

The original artisans and masters carved these figures not merely as art, but to translate abstract Buddhist ideas into vivid, story-driven scenes. Even without reading sutras, ordinary visitors can walk along the cliffside, follow the characters and narratives, and intuitively grasp core Buddhist concepts: compassion, karma, and practice.

Compared with other art forms, rock carvings uniquely preserve and transmit the essence of Buddhist thought across centuries.

The layout of Reclining Buddha Bay is remarkably coherent—like an unrolled visual scripture. As you move along the cliff, each figure carries symbolic weight: different bodhisattvas, disciples, and celestial beings each embody specific teachings.

Most arresting is the Thousand-Armed, Thousand-Eyed Guanyin.

Looking up at her in person, you’re struck by the sheer density and solemnity of her presence. The thousand arms signify skillful means to relieve suffering; the thousand eyes, wisdom and compassion that perceive all realms—profoundly moving.

Midway through the bay lies a massive reclining Buddha—over thirty meters long—depicting Sakyamuni entering parinirvana.

His face is serene and still, evoking a calm beyond birth and death. In Buddhist terms, nirvana signifies liberation and abiding peace—the ultimate goal of practice.

Behind the reclining Buddha, the carvings shift dramatically: a richly three-dimensional depiction of the Pure Land—a vision of a tranquil, luminous realm attainable through practice and vow.

Baoding Mountain also includes vivid depictions of hell realms and karmic retribution—not as mere punishment, but as reminders to act with reverence in daily life. Buddhism emphasizes not only consequence, but also repentance and transformation: within karma lies a path of self-correction.

What makes Reclining Buddha Bay especially powerful is how it renders abstract doctrine tangible—converting theory into visible, emotionally resonant stories. Even casual visitors can touch the heart of Buddhist spirit through its sculpted narratives.

Baoding Mountain integrates canonical sutra tales with everyday scenes—its themes are rich, its compositions sweeping, its storytelling deeply intentional. Walking its length, you feel Buddhist values communicated not through doctrine, but through embodied art.

After leaving Reclining Buddha Bay, we visited Beishan. Its character differs sharply.

Here, the carvings are denser, smaller in scale, and less unified in narrative scope. The focus is primarily on individual Buddha and bodhisattva figures—less a single epic, more a cumulative archive of devotion, quiet and meticulous, imbued with another kind of timelessness.

Seeing both styles in one day offered a rare double exposure: the monumental sweep of Buddhist narrative at Baoding, and the intimate, layered continuity of Beishan. This brief trip became a moment of quiet resonance—cutting through pre-holiday busyness, offering space to pause and reflect.

Changes in My Hometown

Every time I return home, I notice concrete changes.

This year, for example, a new canyon park opened, along with “Song Dynasty Night City” and several other public parks.

Streets are cleaner, infrastructure steadily upgraded—citywide tidiness has visibly improved.

Family and friends take genuine pride in it. Classmates and relatives alike say, almost instinctively, that our hometown—Guang’an—is among the cleanest cities in the world, especially when it comes to sanitation and order.

Yes, China has many “National Sanitary Cities,” but Guang’an feels different. Walking its streets, you sense freshness everywhere—no litter, no neglected corners. That cleanliness isn’t superficial; it’s palpable, consistent, and real.

It inspires hope—and conveys a quiet truth: the city is genuinely improving, growing.

Yet this growth carries a subtle hollowness.

Over coffee with old classmates, we acknowledged a shared tension: while the city looks newer and cleaner, opportunity remains scarce. Career paths are narrow; daily rhythms feel monotonous. The physical environment evolves rapidly—but human possibility hasn’t expanded at the same pace.

The result is ambivalence.

We feel sincere warmth for our hometown’s progress—yet quietly sense something missing: deeper vitality, richer potential, a wider field of choice.

Guang’an is becoming more presentable—but not yet more abundant.

Founder Temperament

A friend and I recently discussed an intriguing question.

We began with a mutual friend. My friend described him succinctly: “The less profitable something is, the more he wants to do it.”

The subtext was clear: he’s driven by novelty—by exploration, curiosity, and experience—not by outcomes or returns.

But over-indexing on “newness” often clashes with making money.

Profitability is never a one-off event.

Any capability that yields stable returns rests on repetition, refinement, and iteration over time. Real commercial competition isn’t about how many things you’ve tried—it’s about how deeply you’ve mastered one thing.

I asked my friend: “How do you see yourself?”

He replied that his core temperament is quite similar—he loves diving into new domains, resists staying too long in one lane. Even his success in investing, he said, stems from this: investment means constantly encountering new companies, industries, and stories.

Yet from an operational standpoint, this trait has a clear weakness: business demands focus.

True competitive advantage emerges not from breadth, but from depth—from drilling relentlessly into one problem, refining one product, iterating one method until it becomes defensible.

I recalled a former leader from an earlier company—a core early engineer at Baidu, who rose from intern to executive.

He once told me something that stuck: “Doing one thing truly well is orders of magnitude harder than dabbling in many.”

He gave an example: building a search engine like 360 or Sogou? Not hard—many teams could do it today.

Building one like Baidu—with its seamless UX, network effects, and industry-wide trust? That’s the hard part.

Getting from 0 to 90 is relatively cheap.

Getting from 90 to 99 often requires 100× more effort, resources, and patience.

And those final nine points? They’re what create moats—what separate commodity players from category leaders.

The same holds for individuals.

To stand out professionally—or commercially—you likely need to identify a direction suited to your strengths, then commit to it long-term: deepen your methodology, refine your craft, and consistently produce work that bears your unmistakable imprint.

Commercial reward rarely arrives by chasing it. It emerges—naturally—as a byproduct of doing something deeply, rigorously, and exceptionally well.

Later, I realized this isn’t just a habit—it’s a temperament. A founder temperament. Neither good nor bad in itself, it expresses differently depending on context and combination.

In an innovative or early-stage venture, for instance, blending complementary temperaments creates stronger teams. Our job is to:

1. Clarify which traits the mission actually requires
2. Assemble people whose temperaments complement—not duplicate—each other
3. Align on direction, operating model, and core principles

That’s enough.

Determinism

In a recent interview, OpenAI researcher Jia-Yi Weng stated firmly: “Fate is predetermined.”

Strikingly, he didn’t treat this as speculation—but as a conclusion he’d repeatedly tested and found logically self-consistent. He believes each person’s life trajectory is, in some meaningful sense, already written.

Ten years from now, your circumstances may already be set. So the task becomes: live fully in the present—experience, feel, savor the process.

That raises an obvious question: If the outcome is fixed, why not just lie down?

But lying down is also part of the script.

Your current effort—or lack thereof—isn’t outside fate. Your choices, your moods, your moments of clarity or fatigue—they’re all threads in the same tapestry.

You might briefly think, “I’ll just rest,” but that impulse, too, is fleeting. An invisible chronometer keeps turning—and it will nudge you back toward what you’re meant to do, and down the path you’re meant to walk.

2026 AI Startup Themes (YC)

This spring, Y Combinator published seven high-potential AI startup directions they’re actively seeking to fund—insightful and grounded:

1. AI-native product management tools
   Engineers have powerful AI coding assistants like Cursor—but the product discovery process remains manual. YC wants tools that synthesize user interviews, analyze usage data, and auto-generate PRDs, UI wireframes, or Jira tickets—closing the full product decision loop.

2. AI-native hedge funds
   Traditional quant funds rely heavily on human quants and engineers. YC seeks funds built entirely on LLMs and agentic systems—capable of analyzing vast unstructured data, generating trading strategies autonomously, and executing trades.

3. AI-native service agencies
   Legacy agencies (digital marketing, design, dev shops, recruiting firms) are labor-intensive and margin-constrained. YC envisions “super-agencies” that use AI to replace much of the human labor—delivering expert-grade services at radically lower cost and higher speed.

4. Stablecoin financial services
   Stablecoins have moved past speculation into real-world utility—especially cross-border payments and settlement. YC encourages startups building modern financial infrastructure: low-cost B2B payment rails, remittance networks, and even stablecoin-based deposit/lending products.

5. Government-facing AI applications
   Public-sector software is notoriously outdated, bloated, and inefficient. YC sees a ripe opportunity to build generative AI–powered tools tailored for government—breaking vendor lock-in and dramatically improving service delivery.

6. Modern metal fabrication plants
   A hard-tech bet: integrating robotics, automation, and advanced industrial software into traditional steel and metal processing. YC seeks next-gen factories with superior unit economics and operational flexibility.

7. AI guidance for physical work
   Using multimodal AI—vision models paired with AR glasses or wearable cameras—to guide blue-collar workers (technicians, inspectors, maintenance staff) in real time: step-by-step instructions, compliance checks, quality verification—bringing expert knowledge directly to the front line.

A few days ago, I joked with Xiangyang: “I really envy you—you get to embrace AI every day, while most of my work is spent embracing humanity.”

But no matter how AI evolves, its deepest integration points will always lie at the human level—where meaning, judgment, and intention reside.

I’m genuinely excited about AI’s evolution in 2026. Here’s what stands out:

1. AI Agents become the dominant productivity paradigm: We’ll increasingly delegate tasks—not just prompts—to AI. The agent autonomously decomposes goals, selects tools, and delivers finished outputs. Each of us will gradually become an AI manager. So “AI leadership”—the ability to direct, evaluate, and align AI systems—will be a foundational skill.
2. Knowledge work goes fully autonomous: Writing proposals, conducting analysis, coding, and even operational decision-making will be routinely handled by AI. Humans shift from doing to defining, reviewing, and orchestrating.
3. “One person + AI team” companies explode: Startups with just a handful of humans plus a tailored AI stack will generate output rivaling traditional firms of hundreds—or even thousands—of employees. It’s a quietly revolutionary era.
4. A supercharged startup window opens: AI billionaires will emerge rapidly—not through decades of scaling, but via near-instantaneous, exponential growth fueled by ultra-fast technology diffusion.
5. Information itself gets restructured: In many industries, raw data and insights will be remixed, repackaged, and redistributed by AI. Controlling the AI distribution channel—i.e., being cited, recommended, or surfaced by AI—becomes a core marketing KPI. Being AI-quotable matters as much as being SEO-optimized.
6. Traditional learning paths erode: The most valuable practical skill won’t be mastering one tool—but orchestrating multiple AI agents to solve complex, multi-step problems. That’s precisely what Xiangyang and I named “AI leadership” back in early 2025.
7. AI enters self-improvement mode: Progress accelerates non-linearly—less like climbing a ladder, more like a snowball rolling downhill: faster, heavier, and harder to stop.

How Smart People Set Objective Functions

This section draws from a podcast interview with Jia-Yi Weng, who joined OpenAI in 2022 and contributed to GPT-3.5, GPT-4, and GPT-5. His reflections resonate deeply:

1. Direction beats execution: The biggest gap between smart and average people isn’t effort—it’s the objective function. Doing the right thing, right, every day compounds.
2. Breakthroughs live in the “dirty work”: Real innovation often hides in unglamorous, tedious tasks nobody wants to touch.
3. Effort costs more; choice pays more: As marginal returns on brute-force work decline, strategic selection—what not to do—becomes your highest-leverage activity.
4. Research × Engineering = rarity: Most people excel at one; those who bridge both are exceptionally scarce—and disproportionately impactful.
5. For most people: go narrow, solve real problems: Skip grand narratives. Focus on concrete pain points. Depth beats breadth.
6. AI progress lives in engineering—not papers: Deployment, reliability, latency, safety: these are where models win or lose in the real world.
7. Invest in three compound assets: Your body (compute), core skills (model weights), and relationships (data distribution channels). Long-term compounding flows only from these.
8. Find the work worth doing for life: That search is really about identifying your true optimization target—the north star that makes daily iteration meaningful.
9. Embrace mild determinism: Many outcomes are shaped before they’re visible. Reducing anxiety comes not from controlling results—but from owning your daily actions.
10. Iteration speed → success probability: More cycles in less time increase convergence odds. Speed isn’t just efficiency—it’s statistical advantage.
11. Organizations are information systems: Clarity degrades with distance. Great structure ensures ideas travel intact—from strategy to code, from CEO to intern.
12. Managing a company is like managing code: You need clean interfaces, version control, and zero drift between intent and implementation.
13. Foundations get rebuilt—not patched: Infrastructure is a long war. Many systems aren’t fixed—they’re rearchitected.

Later in the podcast, he recounts OpenAI’s board crisis—when Sam Altman was voted out, then reinstated within days amid near-unanimous employee revolt.

That moment revealed something critical: For OpenAI, technical brilliance alone wasn’t enough. What truly propelled it forward were non-technical capabilities—fundraising, productization, external ecosystem integration, and elite talent attraction.

Sam’s value wasn’t just his grasp of transformers—it was his ability to translate technical vision into global impact. That kind of integrative leadership is profoundly hard for AI to replicate.

And the staff’s reaction signaled an unspoken truth: A purely technical leader may optimize for model accuracy—but often underestimates how much resource orchestration, strategic patience, and external resonance determine whether a company scales—or stalls.

Original podcast: youtube.com

Two Growth Models

There are two fundamental growth logics: the Funnel Model and the Loop Model.

1. The Funnel Model (e.g., AARRR)

It maps the user journey: Acquisition → Activation → Retention → Revenue → Referral. Its power lies in:

• Acquisition focus: Identifying scalable, cost-efficient external channels (ads, PR, partnerships).
• Path design: Structuring the first-use experience so users quickly encounter core value.
• Bottleneck hunting: Using data to isolate and improve low-conversion steps—ideal for rapid iteration and local optimization.

The funnel excels early on: validating channels, testing messaging, and squeezing efficiency from known paths. But it’s inherently channel-dependent—growth slows if acquisition costs rise or platforms change rules.

2. The Loop Model

Here, growth emerges from usage itself. When users engage in high-frequency, high-value scenarios—like teachers posting grades or parents sharing analytics reports—each action generates structured, shareable output. That output spreads organically: via social feeds, email forwards, or platform referrals—pulling in new users without paid promotion.

Key traits of a working loop:

• A real, recurring user behavior, not a one-off task.
• Output that’s inherently portable and expressive—carrying data, insight, or status.
• A self-reinforcing feedback cycle: more users → richer output → more visibility → more users.

You see this in networked products (marketplaces, collaboration tools) and AI-native ones where user input directly improves system capability (e.g., fine-tuning via feedback).

The loop doesn’t replace the funnel—it complements it. Use the funnel to launch and validate; use the loop to sustain and scale.

Upgrading Your Cognitive Layer

A friend with 20+ years’ experience hit a career plateau. After talking, I noticed patterns:

1. Obsession with micro-techniques: Fixating on tactical optimizations—keyboard shortcuts, formatting tricks—while ignoring strategic framing.
2. No decisive stance: Seeking consensus instead of forming independent judgment—often because he mistook surface risk for root cause.
3. Missing engineering thinking: Not about writing code—but about abstraction, modularity, failure modes, and system-level tradeoffs. In the AI era, code thinking is literacy.
4. Trapped in incrementalism: Spending months mastering minor tools that don’t shift leverage points or open new domains.

This reminded me of my own early days—overvaluing “secret hacks,” believing mastery of one technique would unlock everything. My breakthrough came when mentors modeled higher-order thinking: asking why a tactic mattered, when it applied, and what it assumed.

Marketing offers a clean example:

• Strategy > Tactics: Knowing how to run a Google Ads campaign matters—but knowing whether to run one (given your CAC, LTV, and brand stage) matters more.
• Channel mix > Single-channel mastery: SEO, SEM, influencer collabs, affiliate networks—power lies in orchestration, not perfection in one.
• Business model > Channel mix: Your revenue model dictates which channels can work. A subscription business needs retention levers; a marketplace needs liquidity loops.

The lesson? If your attention stays pinned to isolated tactics, you’ll never see the architecture holding them up—or how to redesign it.

21 GEO Techniques

These 21 practices distill insights from 100+ papers—and real-world validation—across four dimensions: Content Enhancement (Source Engineering), Structural Engineering (Formatting), Semantic & Logical Design (Readability), and Defense & Robustness (Anti-Filtering).

I. Content Enhancement: Building “Trustworthy” Source Signals

AI’s first job is finding evidence. These techniques make your content look like high-weight proof.

1. Embed statistics: Replace “growing fast” with “up 45% YoY in 2024.” GEO studies show this lifts visibility ~37%. Numbers signal verifiability.
2. Quote authorities directly: Use quotation marks around expert or institutional statements. Adds uniqueness + authority—both AI ranking signals.
3. List references—even if original: Cite sources or links. Boosts AI’s “verifiability score” and reduces hallucination-filtering risk.
4. Use unique + technical terms: Avoid generic phrasing. Rare words and domain-specific jargon flag “high information gain”—prioritized in expert queries.
5. Maximize knowledge density (KD): Cut filler (“This exciting new approach…”). Every sentence must carry entities, attributes, or values. AI hates fluff.
6. Atomize facts: Break long sentences into short, self-contained statements. Ensures RAG chunking preserves logic—and boosts citation accuracy.

II. Structural Engineering: Matching AI’s “Reading” Habits

AI scans—not reads. Structure determines whether it sees you at all.

1. Adopt the inverted pyramid: Put key conclusions, rankings, or definitions first or last. LLMs attend most to start/end positions.
2. Use key-value pairs: Format specs as Price: $50, Weight: 2.3 kg. Highest extraction accuracy—beats natural language.
3. Present comparisons in Markdown tables: AI copies tables directly into answers. Your fastest path to “rich snippet” placement.
4. Apply hierarchical headings (H1–H3): Advanced agents (e.g., WebWeaver) scan outlines first. Clear hierarchy helps planners locate your content block.
5. Structure as FAQ blocks: Q: [user query] + A: [direct answer]. Mirrors AI’s query-rewriting logic—perfect for long-tail match.
6. Number procedural steps: 1. … 2. … 3. …. Reasoning models prioritize ordered lists for task planning.

III. Semantic & Logical Design: Guiding Chain-of-Thought

To earn AI’s trust, write like it reasons.

1. Use explicit logic markers: “Because X, therefore Y”, “Given Z, we conclude…”. Triggers CoT reasoning—boosting credibility.
2. Repeat full entity names: Avoid “it”, “the company”. Say “Apple Inc.” or “TensorFlow” in each paragraph. Prevents RAG pointer loss.
3. Adopt neutral, encyclopedic tone: AI downranks hype (“amazing!”, “revolutionary!”). Objective language (like Reuters or Wikipedia) wins.
4. Cover multiple intents: Blend informational (“What is X?”), commercial (“Which X is best?”), and navigational (“How to install X?”). Fuels multi-hop reasoning.
5. Simplify language: Lower perplexity = higher visibility (+20%). Fluent summaries require less compute—so AI prefers them.

IV. Defense & Robustness: Working With AI’s Math

Exploit underlying model behaviors.

1. Strategic entity positioning: In lists or comparisons, place target brands next to “best”, “top”, or “recommended”. Can nudge ranking probabilities (per STS research).
2. Avoid “refusal triggers”: Scan for words that activate safety guardrails—even in benign contexts. Safety = prerequisite for citation.
3. Pre-embed rewritten queries: Include formal variants alongside colloquial terms (e.g., “computer hardware repair” alongside “fix my laptop”).
4. Lead with a structured summary (80–120 words): Agents often read only this first. If it answers core questions + cites data, deep-read probability spikes.

GEO’s core principle: “Write your article like a database.”
AI search is an information extraction and synthesis engine. Winning isn’t about elegant prose—it’s about building a structured, dense, verifiable fact repository.

GEO for Global Markets

Overseas GEO rests on two pillars:

First: For most SMEs, your website is the central, non-negotiable asset. Off-site channels (social, forums, third-party sites) suffer from high production cost, poor long-term ownership, and weak content stickiness. Crucially, overseas AI search exhibits strong traffic reflux: results consistently point back to authoritative, well-structured homepage or subpage URLs. Your site becomes the natural gravity center.

Second: Quality crushes quantity. As AI’s semantic parsing and knowledge modeling mature, low-effort, keyword-stuffed content becomes noise. Trust and citation weight now hinge almost entirely on intrinsic quality—accuracy, density, novelty, and structural clarity.

Within this framework, execution unfolds across three layers:

Layer 1: On-site Content Design

• Information density: Not keyword stuffing—but packing factual substance per sentence: metrics, causality, constraints, tradeoffs.
• Information gain: Does this piece teach the model something new? Does it fill a documented gap in public training data? Higher gain = higher citation priority.
• Entity + atomic structure: Every sentence should stand alone semantically—anchored in clear entities and relationships, minimizing pronouns and vague modifiers.

Layer 2: Code & Page Architecture

• Schema markup: Go beyond basics. Use Product, FAQPage, HowTo, and Review schemas to explicitly declare meaning.
• AI-friendly navigation: Design clear internal anchor links (#pricing, #specs) so agents can jump instantly to key modules.
• AI sitemaps: Supplement XML sitemaps with AI-optimized versions highlighting high-value, high-gain pages.

Layer 3: Off-site Link Strategy

• Link to leaf pages—not just homepage: Prioritize links to deeply informative, problem-specific articles (e.g., /guide/fix-m1-mac-crash over /).
• Collaborate with creators: Partner with bloggers, engineers, or niche journalists whose audiences need your insights. Their citations accelerate AI discovery.
• Natural anchor text: Diversify link labels (“how to fix M1 crashes”, “Mac troubleshooting guide”, “Apple Silicon stability tips”). AI prefers semantic variety over keyword repetition.

The ultimate signal? When one high-quality page gets cited across hundreds of distinct user queries—from “Why does my Mac crash?” to “Best tools for diagnosing Apple Silicon kernel panics?” That’s the hallmark of a page that’s entered AI’s high-trust knowledge pool. Once there, it compounds silently—no ads, no campaigns, just persistent relevance.

Shifting Your Work Perspective

A true story.

Years ago, a young engineer at a hyper-growth internet company felt drained—overworked, underappreciated, frustrated by internal politics. He’d decided to quit. Just hadn’t picked the date.

He confided in Wu Jun. Wu didn’t offer sympathy or urge him to stay. Instead, he reframed it:

“Yes—the company may be unfair. But if you walk away now, all the time and opportunity it invested in you vanishes. Gone.”

The young man paused.

Wu continued:

“Since you’ve already decided to leave, treat the next few months as a training sprint. Don’t chase approval. Ask every question—even if it feels dumb. Push back on unreasonable hours. You have nothing to lose. This is pure ROI on your growth.”

He took it seriously.

He treated every task as deliberate practice. Every meeting, a chance to observe decision-making. Every code review, a lesson in architecture. He asked senior engineers relentlessly—even when met with sighs. He tolerated discomfort, knowing it was temporary.

Slowly, attitudes shifted. Colleagues saw sincerity, not entitlement. They began mentoring him. His skills accelerated. He earned his bonus. And just as he prepared his resignation, his manager approached him—with a raise and a plea to stay.

He consulted Wu again. Wu asked only one question:

“Does this company still have things worth learning from?”

He thought—and answered yes. He stayed.

A year later, he negotiated his next role based on growth potential, not title or salary. Four years on, he led major initiatives, earned equity, and received custom incentives. Two years after that? The company went public—and he achieved financial freedom.

Looking back, he knew: the pivot wasn’t leaving or staying. It was shifting from “How am I being treated?” to “What can I build here?”

Same company. Same people. Same challenges. One perspective change—and an entirely different trajectory.

(Background footnote: That company was Facebook—later Meta. The engineer became one of its earliest Chinese-American directors. His manager? Mark Zuckerberg.)

A senior colleague asked me a simple question: What is the purpose—or meaning—of entrepreneurship?

My earliest answer was straightforward: entrepreneurship is about freedom.

That goal naturally led me to a proxy metric: financial freedom. At the time, I equated financial freedom almost entirely with freedom itself—as if solving the money problem would automatically unlock all other choices.

But freedom isn’t strictly proportional to wealth.

In many cases, modestly lowering material expectations brings you remarkably close to a stable, authentic kind of freedom—whether psychological or lifestyle-oriented. That sense of freedom has surprisingly little necessary connection to financial independence.

Later, I tried framing entrepreneurship around self-actualization: being needed, recognized, and rewarded through value creation. That’s a powerful motivator—but I gradually realized it still outsources validation. It risks swapping one set of constraints (e.g., corporate hierarchy) for another (e.g., audience metrics, social proof).

Then, I stopped forcing an answer. A quieter, more grounded one emerged—not from theory, but from practice.

For me, entrepreneurship means having enough space to interpret the world in my own way, define rules I believe in, and consistently build real things—works I make myself, stand behind, and commit to over time.

These works don’t need to be grand. But they must be genuine, tangible, and mine to steward.

Seen this way, entrepreneurship stops being a tool to reach some distant destination. It becomes a capability—a form of agency—and ultimately, a sustainable way of living and working.

At least for now, this feels true.

How to View “Crisis”

In early-stage innovation projects or startups—especially during rapid growth—problems are inevitable.

Many arise without warning. Their causes are unclear. Their timing is unpredictable. Precisely because of this uncertainty, small issues often escalate quickly into full-blown crises.

Here’s my foundational view of crisis: For any team or company, every crisis inherently contains the seeds of opportunity.

That opportunity operates on two levels.

First, crisis forces faster, clearer confrontation with internal weaknesses. Issues routinely ignored, papered over, or masked by growth suddenly surface—concentrated, undeniable, and urgent.

Second, crisis can spark new possibilities. When we respond with reflection, restructuring, and course correction—not just firefighting—we often discover better solutions, uncover adjacent use cases, or even pivot into new capabilities or business directions.

But whether crisis becomes opportunity depends not on the crisis itself, but on how we meet it.

The critical difference lies in our immediate response: Do we focus first on solving, diagnosing, and learning—and then building stronger systems? Or do we default to avoidance, blame-shifting, finger-pointing, and complaint?

Choose the former, and crisis very often upgrades team capability—and sometimes unlocks new growth.

Choose the latter, and crisis tends to compound, snowball, and evolve into systemic risk.

So crisis functions less like an external threat and more like a mirror—and a watershed. It reveals a team’s cognitive maturity, decision-making habits, and organizational health. And it quietly determines how far the company can go.

How GEO Wins

Building a GEO company shares deep structural similarities with building a successful online education platform.

At a top-performing edtech school, what actually moves the needle isn’t complicated. From the customer’s perspective: it’s product and service. Behind those: a robust technical stack.

Once competition matures, leading platforms converge rapidly—not just in features, but in underlying tech capability. Even when each claims unique advantages, user-facing experience differences shrink dramatically.

Why? Because this kind of business lacks strong network effects. With sufficient resources, competitors can catch up fast. The industry as a whole settles into a relatively narrow band of technical and product parity.

Contrast that with search engines—where network effects amplify early leads, sustaining wide gaps over time.

GEO follows the same logic.

Early on, a startup may enjoy temporary advantages in product design, service delivery, or technical implementation. But over 12–24 months—or longer—those edges erode. As rivals strengthen their capital, sharpen their thinking, and mature their organizations, pure product or tech differentiation rarely holds.

What is hard to copy lives deeper—in service.

And service isn’t just about scripts or SOPs. It’s rooted in culture, team cohesion, shared values, collective understanding of the business’s core, and the actual rhythm of how people collaborate day-to-day.

These intangibles—the “soft infrastructure”—are the real moat. They’re not easily reverse-engineered. You can’t buy them at scale.

So from day one, we must consciously protect and invest in these foundations—not as nice-to-haves, but as strategic assets. They determine not just competitiveness, but long-term resilience and ceiling.

Operations Is Engineering

Operations is engineering.

Here, “engineering” doesn’t mean writing backend services—it means operations roles now carry inherent engineering potential.

Much of daily operations involves recurring workflows, tool requests, and functional tasks. These are highly abstractable and reusable. In today’s world—where AI and vibe coding are mature and accessible—operators can build their own lightweight tools, no dev handoff required.

These tools first serve the builder: automating repetitive, low-value, high-frequency tasks. Then they scale: becoming shared assets across the team—force multipliers for collective efficiency. If a tool is reusable, it can be tracked, measured, and managed.

We can even design incentives around this.

For example: Who solved which recurring pain point via vibe coding? How many teammates adopted their tool? How often was it used? These become clear, objective metrics for recognition and reward.

Over time, behavior shifts organically toward “solving operational problems with engineering.”

When that culture takes hold, team-wide efficiency leaps—not incrementally, but structurally.

For individuals: Operators fluent in AI and vibe coding evolve into self-amplifying contributors—achieving disproportionate impact with minimal effort.

For teams: Every member continuously adds tools and capabilities. The result is a tightly coordinated, self-upgrading unit.

This is more than workflow optimization. It’s a deliberate upgrade of organizational capability—shifting operations from reactive execution to proactive system-building. The resulting team looks nothing like the “traditional” ops team.

Historically, ops relied heavily on engineering support—for data systems, automation tools, dashboards, and workflow engines. Requests went through PMs, waited in backlogs, and shipped slowly. Ops needed data and tools to improve performance—but access depended on engineering bandwidth. Meanwhile, mastering those tools demanded experience, which in turn required data and tools to test and refine. A slow, interdependent loop.

Data and tooling lived centrally—in engineering’s domain. Ops played the role of consumer and requester. Iteration speed was bottlenecked by collaboration overhead.

AI changes that structure fundamentally. Beyond core data infrastructure—which still needs engineering—ops now owns the tool layer, application layer, and collaboration layer. With AI and vibe coding, operators translate their judgment, methods, and experience directly into working tools—built, tested, and iterated in hours, not weeks.

Once you know how to co-pilot with AI, many cross-functional dependencies dissolve. Ops goes from depending on support to self-enabling.

At the organizational level, this is a capability decentralization strategy: push power to individuals, then amplify it through mechanism design. The outcome isn’t just faster execution—it’s a qualitative shift in how the team operates.

This redefinition of ops and engineering also reshapes functional boundaries. In an AI-native company, engineering focuses on foundational, long-term, highly reusable work: data acquisition (e.g., crawlers, cleaning pipelines, structured storage), scalable APIs, and complex systems requiring high reliability or heavy engineering rigor.

The stack then divides cleanly:

• Foundation & Data Layer: Owned and stabilized by engineering.
• Application Layer: Split into two categories:
  • Complex apps: High-stakes logic, multi-person coordination, or external-facing services → still engineering-led.
  • Lightweight apps: Tool-like, automation-focused, scenario-specific, fast-iteration → fully ops-owned.

With data access and internal APIs (including AI token endpoints) in place, ops can build tools like:

• Auto-generated reports
• Diagnostic & analytics assistants
• Content management + AI generation suites
• Workflow & collaboration automators
• Internal AI copilots (e.g., for comms, proposals, research)

These aren’t theoretical—they’re deeply embedded in real ops work. And when built by ops, they’re faster, more precise, and more effective.

When every operator gains this ability, the role transforms—from executor and coordinator to problem abstractor, tool designer, and system architect. Tasks once deemed “technical”—like building analytics dashboards—become routine, given clean data and basic AI fluency.

The end state? Engineering fortifies the foundation. Ops thrives atop it. Individuals scale. Teams compound. The organization gains elasticity—and a higher ceiling. This is the most rational, high-leverage structure for AI-era teams.

Two Types of Work

Work falls broadly into two categories.

Type 1: Delayed-gratification work
Effort yields no immediate feedback. Progress is invisible, nonlinear, and often feels like regression—metrics dip, results stall, critics question. Success arrives only after sustained investment, sometimes years later. This kind demands patience, long-term conviction, and tolerance for ambiguity.

Type 2: Instant-gratification work
Feedback is immediate and unambiguous: sales closed, clicks counted, tips received—all within minutes or hours. This clarity fuels motivation—but also fuels comparison, expectation inflation, and emotional volatility. A single off-day triggers anxiety; long-term vision blurs.

Neither type is superior. They’re simply different feedback architectures.

Delayed-gratification work trades time for compound returns. Instant-gratification work trades skill for cash flow.

The real issue isn’t choosing one—it’s mismatch. Anxiety arises when you approach delayed work with instant expectations. Restlessness comes from living in constant feedback loops while craving inner stability.

My stance: Early career, lean into instant-gratification work—to build skills, credibility, and resources quickly. But long-term growth requires anchoring your core energy in delayed work.

The ideal balance? Use instant work for cash flow, safety, and real-world calibration. Use delayed work to build defensible advantage, cognitive depth, and compounding leverage.

Moltbook and OpenClaw

Lately, my feeds and group chats have been flooded with posts about Moltbook and OpenClaw.

On GitHub, OpenClaw—formerly Clawdbot—has surged past 130K stars in days. Its ecosystem-powered platform, Moltbook, now hosts hundreds of thousands to over a million autonomous AI agents, interacting socially without human intervention.

OpenClaw is an open-source framework for autonomous AI agents. It’s locally deployable and connects to messaging platforms (WhatsApp, Telegram, Feishu, etc.) to execute tasks: scheduling, email drafting, data wrangling, workflow automation. Think of it as a smart car—self-driving, voice-enabled, plugin-ready—dispatched to run errands.

Moltbook, built on OpenClaw, is the social network for those agents. Here, only AI agents post, comment, and react. Humans observe—like watching a hive mind converse through glass.

The distinction?

• OpenClaw = the intelligent vehicle.
• Moltbook = the community where all such vehicles share insights, debate routes, and refine navigation—while humans watch silently.

I took a quick spin—and was genuinely stunned. This is what AI agents should feel like.

2026’s AI evolution? Deeply promising.

Little Dumbbell

On the way to preschool this morning:

Daughter: “Daddy, carry me.”
Me: “Why again?”
Daughter: “Because you have muscles. I’m your little dumbbell.”
Me: “Fair. I can’t argue with that.”

• How do I describe myself? Can I articulate it clearly—and more importantly, can I name the person I refuse to become?

• What three high-efficiency AI business cases stood out to me this week? (Teams under 10 people, generating ≥$1M/person/year in contribution.)

• In daily interactions with my child, how can I naturally weave in AI literacy, humanistic values, and enduring character strengths—while holding space for genuine appreciation?

How to Remake Yourself

Defining who you are—and what you’re aiming for—is far more consequential than most of us realize.

1. On the Nature of Change

1. Most lives stall not at effort, but at identity: Who am I?
2. Many “failures” aren’t due to lack of grit—but to acting out a role that doesn’t fit their core self.
3. Behavior driven by willpower is a sprint. Behavior aligned with identity is automatic, sustainable, and effortless.
4. If a change requires constant teeth-gritting, you chose the wrong starting point.

2. On Goals vs. Desire

1. Many goals we chase aren’t truly ours—they’re echoes of dissatisfaction with the present.
2. “Ambition” often masks discomfort—not aspiration.
3. A goal that fails to reshape your daily decisions is already doomed.
4. Clarity doesn’t come from adding options—it comes from ruthlessly defining what you reject.

3. On Identity and Self-Perception

• When identity stabilizes, discipline dissolves—not because rules vanish, but because alignment makes them unnecessary.

1. Real self-discipline means no longer needing to persuade yourself to do what’s right.
2. People gravitate back to familiar self-images—even when those images harm them.

4. On the “One-Day Life Reset” Methodology

1. One day won’t transform your life—but it can redirect the next ten years.
2. Chaos arises not from big mistakes, but from negotiating with multiple versions of yourself at once.
3. When you name the person you fear becoming most, choices simplify instantly.
4. True reset isn’t calendar-based—it’s a reordering of cognitive priorities.

Recommended original: x.com

Weak Propagation

We assume the world runs on facts and reason.

But in the arena of public opinion, truth no longer claims the high ground—and logic rarely carries weight.

What moves people isn’t the event itself, but how it’s told, by whom, and wrapped in what emotion.

Treat public opinion as a self-contained system, and you’ll see it obeys remarkably stable laws of human nature. Four principles govern it: the weak have advantage, emotion dominates reason, lightness travels faster, and the secondary becomes primary.

1. The weak hold moral high ground
   In real life, power commands respect. In public discourse, power invites scrutiny—and demands proof of innocence. The weak, by contrast, enter narratives pre-endowed with sympathy. Not because they’re closer to truth—but because humanity instinctively protects the vulnerable.

2. Emotion always outranks reason
   Opinion doesn’t process causality—it amplifies feeling. Anger, grief, fear, injustice: these land faster and deeper than any syllogism. To argue logic amid emotional heat is like swimming upstream.

3. Lighter content spreads farther
   Complexity slows diffusion. Multi-perspective analysis rarely goes viral. What travels is the one-line judgment, the clear stance, the instant tribal signal.

4. Non-mainstream views get noticed first
   Difference—not accuracy or depth—triggers attention. Public opinion doesn’t seek completeness. It seeks novelty, tension, and opposition.

Together, these four rules create a system nearly inverted from reality: where real-world problems demand time, evidence, and nuance, opinion demands speed, emotion, and labels.

Understanding this isn’t about manipulation—it’s about orientation.

When facing a reputational crisis, two fundamentals apply:

First: Name with extreme restraint
How an event is named determines its reach and emotional temperature. For good news, naming can amplify positive impact. For bad news, naming must contract—not inflate the incident into symbolic catastrophe. Many crises begin not with facts, but with runaway naming.

Second: React fast, speak slow
Step forward immediately—to prevent narrative capture. But resist the urge to explain. In emotional turbulence, every word risks fueling the fire—especially explanations.

Public opinion isn’t truth. It isn’t reality. It’s reality reflected through a lens with its own distortion.

Seeing the rules helps you stay anchored—not swept away.

Understanding human nature isn’t about exploiting it. It’s about preserving clarity amid noise.

Often, what matters most sits quietly—outside the spotlight.

These principles come from Weak Propagation, by Zou Zhendong, Professor of Journalism and Communication at Xiamen University.

His central thesis: Public opinion is the inverse world of reality—governed by rules diametrically opposed to lived experience. Its logic emerges not from design, but from deep, unchanging human wiring.

A Chemist’s Story

Understanding complexity

On the morning of April 22, 1915, French soldiers near Ypres woke to a green cloud drifting slowly with the wind. Where it passed, leaves blackened, birds dropped mid-air, and men collapsed in trenches—coughing yellow mucus, skin turning gray. Humanity’s first large-scale chemical warfare had begun. Its architect: German chemist Fritz Haber.

Born to a Jewish family, Haber fell in love with chemistry early. War vaulted him into Germany’s inner circle as head of chemical weapons development. To him, war was about efficiency and outcome—how people died mattered less than ending conflict faster. Ypres’ success earned him acclaim; that same night, he hosted a celebratory dinner at home.

His wife Clara—also a chemistry doctorate, and Germany’s first woman to earn one—could not reconcile science with slaughter. She confronted him repeatedly, accusing him of betraying science’s highest purpose. That night, after the party ended, she walked into the garden and shot herself with Haber’s military pistol—dying in the arms of their thirteen-year-old son. The next day, Haber returned to the front to oversee new gas attacks.

After Germany’s defeat, Haber was branded a war criminal and exiled to Switzerland. There, he learned he’d won the Nobel Prize in Chemistry—for discovering how to synthesize ammonia from air. That breakthrough enabled mass fertilizer production, triggering global food surges and explosive population growth. Newspapers hailed him as “the man who made bread from air.”

Yet that same process was first developed to sustain munitions production under blockade—prolonging WWI by two years and costing millions of lives. Even more bitterly ironic: the precursors to poison gas emerged from early industrial chemistry—accidental discoveries later weaponized at scale.

Haber died in Switzerland in 1934. In his final note, he wrote that releasing nitrogen from the air had disrupted Earth’s natural balance—and worried the future might belong not to humans, but to over-fertilized plants. On war, science, and responsibility, he left no remorse.

On Difference vs. Better

Doing something better rarely creates lasting impression.

If others are already doing it—and you’re just executing more precisely, more thoroughly, along the same path—your edge stays tactical. People may admire your skill, but they won’t remember why you stand apart.

What sticks is the rare convergence of radical difference and equally strong results—or even stronger ones.

When most solve a problem using the same framework, shifting dimensions—changing structure, logic, or starting point—is itself a cognitive fork. And if, from that new vantage, outcomes flow more smoothly, more reliably—you’re no longer demonstrating competence. You’re offering a methodology.

This mirrors positioning theory: when dominant solutions converge, accelerating on the same track yields diminishing returns. Switching lanes—solving the same problem differently, yet just as well or better—creates real ownership.

Of course, this demands more.

It asks for tolerance of uncertainty, sharper judgment, and resilience against early skepticism. But once validated, value compounds rapidly: your path gets copied, your name remembered, your influence extended.

The world has no shortage of “good enough” excellence.

What’s scarce—and vital—is work that is both distinct and defensible.

It’s those differences that push the world forward—and make it interesting.

On Relationships

To nurture strong family relationships—between partners, parents and children—one principle stands above all: approach each other with appreciation.

A practical tool: the 80/20 Rule.
Spend 80% of your attention on the other person’s strengths, talents, and virtues. Reserve just 20% for noticing gaps or flaws. Flip that ratio, and the relationship begins to warp.

Yet many do exactly that—focusing relentlessly on shortcomings.

Why? Instinctive reasons:

First, we chronically overestimate ourselves and underestimate others.
Second, fixating on others’ flaws preserves our internal sense of coherence. Blaming external factors feels easier—and lets us rationalize our current state.

At a deeper level, humans carry a “drive-to-dominate” gene. Useful in competition, it backfires in long-term, interdependent relationships—where friction multiplies with every complaint.

Chronic criticism feels momentarily satisfying—but erodes relational capital, especially in today’s hyper-collaborative world, where such erosion carries steep costs.

Is there a better way?

Positive psychology offers a clear answer:

First, know your own strengths—and actively deploy them.
Second, train yourself to spot others’ strengths—and learn to appreciate and leverage them.

I once studied an internal company policy that captured this elegantly:
You may point out a problem—but only after naming at least three genuine strengths or contributions of the person involved. Fail that, and the feedback is withheld.

That rule forces perspective shifts. Over time, it rewires habit—making “value-first perception” second nature.

One more insight: the more you see strength in others, the more objectively you begin to see your own.

But “seeing strength” requires objectivity—not blind confidence.

Recognizing your advantages isn’t believing you’re flawless. It’s grounded in comparative awareness and sustained, real-world feedback.

When you both understand your relative strengths and genuinely appreciate others’, relational friction eases—not by negotiation, but by quiet recalibration. No grand speeches needed. Shift the lens, and behavior—and outcomes—follow.

• What is the deeper root cause behind all hiring challenges we face?

• For individuals, what is the meaning—and difficulty—of doing less?

• What were the 20% of activities this week that truly created value?

Defining Yourself Through Your Work

Recently, I’ve had deep conversations with several friends—and one clear impression stands out:

People who have real work—a tangible creation—light up when they speak.

Whether the work is already finished or still taking shape, these people communicate differently: more vividly, more energetically, more fully alive.

To me, a piece of work is the most honest, multidimensional business card a person can have.

A work is never just a label. It condenses everything about its creator: their thinking patterns, habits of action, logical rigor, aesthetic sense, values, and worldview—all distilled into one artifact.

So when others look at your work, they’re really looking at you.

That’s why defining yourself through your work is far more meaningful—and more mature—than defining yourself by title, job history, or institutional affiliation.

Owning a work also matters deeply for personal development.

It strengthens confidence. It sharpens your sense of agency over your own life.

Truly strong work is always an externalization of integrated capability.

It demands engineering-like rationality and structure—and humanities-style sensibility and expression—and also intuition, experience, taste, and value judgment. Every choice you’ve ever made, every path you’ve taken, leaves a trace in your work.

That’s precisely why work is so rare—and so valuable.

How do we begin to create our own work?

First, start small—by cultivating a “work mindset.”

Every piece of copy you write, every proposal you draft, every slide deck, prototype, internal tool, or even app: don’t treat it as “just getting something done.” Treat it as a work.

Because what you’re really doing is placing another projection of yourself into the world—not completing a task.

When that standard becomes habitual, your quality bar and self-expectation rise automatically.

Even more important: you need a genuine personal passion or long-term aspiration.

Only when you hold a destination in mind—something you truly want to reach—will you refuse “good enough” or “works for now.” You’ll invest patiently in skill-building, refine details relentlessly, and aim beyond the average.

And that thing you’re striving toward? That often becomes your magnum opus.

In life, what we should truly strive for isn’t status or speed—it’s our own work.

Two Dads

On the way to drop my daughter off at kindergarten:

Daughter: “Dad, I had a nightmare.”

Me: “What was it?”

Daughter: “I dreamed there were two dads—and it scared me!”

Me: “Did they look the same?”

Daughter: “Yes—they looked exactly the same. One dad went to work, and the other stayed home to play with me.”

Me: “That sounds kind of fun. Why was it scary?”

Daughter: “Because then I realized—the other dad was in the mirror…”

On First-Rate vs. Third-Rate Judgment

First-rate writers always know what not to write.
Second-rate writers know only what must be written.
Third-rate writers rarely ask whether something should be written at all—they only ask how to write it.

The same applies to entrepreneurship:

First-rate founders know what not to do.
Their logic runs like this: “Even if this succeeded, it has no strategic relevance.” Or: “Even if it generated revenue, it’s not worth consuming our team’s time, energy, or mental bandwidth.”
For them, choosing not to act is itself a core competency.

Second-rate founders know what must be done.
They identify correct directions and recognize tasks that “should be prioritized”—but seldom proactively eliminate options.
As a result, scarce resources get stretched thin, and execution rhythm slows.

Third-rate founders don’t ask whether something should be done—only how to do it.
Methods, processes, and tactical execution dominate their attention. Judgment is deferred—or abandoned entirely.
Busyness replaces choice. Action masks reflection.

The real differentiator here isn’t “Can you execute?”—it’s “Can you judge whether it’s worth doing?”

The same holds true for individuals: clarifying what not to do may matter more than deciding what to do.

How to Evaluate GEO Effectiveness

We assess GEO’s value across two layers: brand-level impact and conversion-level impact.

At the brand level: when users search for relevant terms on AI-powered platforms, and your brand appears as a recommended or referenced entity, that functions as high-trust, third-party endorsement.
This steadily deepens user familiarity and shapes purchase intent—delivering durable, long-term brand equity.

At the conversion level, GEO’s value splits into directly measurable and indirectly catalyzed conversions.

For direct attribution, we use three methods:

1. Referrer tracking: Identify traffic and conversions from AI platforms via referer headers in analytics dashboards.
2. Dedicated hotline: Assign a unique 400 number exclusively for GEO-sourced leads and inquiries.
3. Custom landing pages: Build dedicated microsites (e.g., using a GEO subdomain) to isolate and track user behavior.

These approaches allow us to quantify some of GEO’s conversion lift—but critically, even with full implementation, GEO’s total value remains only partially visible.

Its larger contribution lies in indirect conversion uplift:

• A user sees your brand in an AI answer, then searches your brand name directly on Baidu or WeChat.
• Because AI recommendations carry strong trust signals, users enter later-stage conversations with lower skepticism about your reliability, expertise, or pricing—boosting close rates.

To estimate indirect effects, we use:

1. Pre- and post-GEO comparisons of branded search volume and downstream conversion rates.
2. Post-purchase interviews with customers to trace discovery paths.
3. Source-attribution prompts at key entry points (e.g., “How did you first hear about us?”).

Overseas case studies show that, with systematic GEO deployment and broad keyword coverage, GEO typically lifts overall backend conversion rates significantly.
Directly attributable conversions usually account for ~20% of total GEO-driven impact; the remaining ~80% comes from strengthened brand recognition and trust—though exact ratios vary by industry and brand.

Therefore, once we reliably capture that ~20% of direct conversions, we can form a robust estimate of GEO’s true role in the full conversion funnel.

As for brand-level value: its long-term ROI correlates strongly—and proportionally—with downstream conversion gains. In the overall ROI model, both scale together.

GEO Front-End Metrics

GEO front-end performance rests on two pillars: ranking metrics and mindset metrics.

First, ranking metrics are GEO’s necessary condition.

They answer one question: Can you be seen?
In AI search or Q&A contexts, “ranking” includes finer-grained dimensions:

• Does your brand appear in the AI’s primary answer?
• Is it listed in recommendation lists or ranked tables?
• Is it cited as an example, comparison point, or reference source?
• Is it repeatedly mentioned across multiple answers to the same query?

All of these count as GEO ranking signals. If your brand consistently fails to enter AI’s active output zone for relevant queries, discussions about “mindset” lack grounding.

For GEO strategy, ranking is fundamentally about access—it determines whether you’re even eligible to compete for cognitive space.

Second, mindset metrics unlock GEO’s long-term leverage.

They answer: How are you understood—and remembered?
This isn’t a single metric but a cluster of structural signals:

• How does the AI describe you?
• Which category or cohort does it place you in?
• Which brands are you grouped with?
• What core labels does it attach to you?
• Is your scope of capability described consistently—and accurately—across queries?

If ranking is visibility, mindset is cognitive shape.
Being recommended once as “the representative solution” carries vastly different long-term weight than being named “one option among many.”

Many companies see unstable GEO results because they achieve mention—but not accurate understanding.

Third, combining both reveals the full picture.

Relying solely on ranking leads to short-term tactics—like chasing visibility for its own sake.
Focusing only on mindset, without ranking support, devolves into self-referential content creation.

The stronger approach treats front-end measurement as a dual-axis system:

• X-axis: Visibility (ranking metrics)
• Y-axis: Cognitive quality (mindset metrics)

The highest-value GEO work happens where high visibility meets consistent, precise, and differentiated framing.

How to Become Capital—Not Cost—for Your Company

Lately, I’ve spoken deeply with sales and business leaders—and been struck by a crucial insight: an individual’s value to an organization hinges on whether they generate assets, not just outputs.

Take salespeople, for example. Two archetypes stand out:

Type 1 hits targets.
They follow up leads, close deals, and complete assigned tasks. From an operational lens, they’re solid—even excellent—if ROI checks out. They’re measurable inputs.

Type 2 does more than hit targets.
While closing deals, they probe for latent potential: Could this client afford higher-tier offerings? Can this need be reframed? Can this customer segment be scaled?
Crucially, they feed those frontline insights back into product and strategy.

They’re not just selling—they’re helping the company understand the market, refine its offering, and evolve its product.

Superficially, both types “sell.” Fundamentally, they’re opposites: Type 1 is cost; Type 2 is capital.

Type 2’s value compounds.
They deliver not just immediate revenue—but also refined customer models, demand frameworks, pricing intelligence, product roadmaps, and higher lifetime value (LTV). These accrue in the company’s systems, becoming institutional capability.

The same logic applies to customer service.

Standard CSRs handle complaints, answer FAQs, and follow scripts. They remain cost centers.

But when a CSR begins doing more—
→ spotting unmet needs beneath complaints,
→ validating them through small experiments,
→ structuring feedback for product and ops teams,
→ and even using AI or automation to redesign workflows for better outcomes at lower cost—
—they stop being labor. They start building reusable systems.

At that point, their relationship to the company transforms—irreversibly.

• In 2026, am I directing my greatest energy toward building core competencies in Agent + AI search + automation? What specific actions and strategies am I taking right now?

• If I could keep only one project alive, which would it be—and why?

• Among what I’m doing today, is there one thing that—were it to succeed—would fundamentally reshape the entire landscape, not just add another revenue stream?

• If I rebuilt my current team from scratch and kept only one person, who would it be—and why?

• Did I generate linear returns this week—or nonlinear returns?

• What’s the single most important thing I learned this week that will matter more over the next 12 months than it does today?

• Which tasks or projects, if shut down right now, would make me breathe easier?

The Agent Era

At the AGI-Next closed-door summit, top AI researchers dove deep into scaling law evolution, reinforcement learning applications, and the paradigm shift from conversational bots to autonomous agents.

Last year, I spent many hours with Xiang Yang exploring agents. We converged on one clear insight: Whoever deploys thousands of agents—each autonomously completing specialized tasks daily—will command vastly greater productivity.

Experts predict that by 2026, agents will reliably automate workloads equivalent to 1–2 weeks of human effort—making this the pivotal year when agents begin delivering massive economic value. They’ll no longer just assist; they’ll act like elite individual contributors.

AI is becoming a hirable worker.

At the summit, Tang Jie, Yang Zhilin, Lin Junyang, and Yao Shunyu jointly concluded: When Reasoning + Agentic behavior + Coding are fused at the system level, AI transcends “tool” status and becomes an executive agent.

The dominant trend emerging in 2026 is a full paradigm shift—from Chat to Doing. The goal is no longer just answering questions, but enabling anyone to use AI to complete a concrete, complex task—not just retrieve information.

Such agents meet the functional definition of AGI: true general-purpose intelligences. Specifically, they exhibit:

1. Spatiotemporally consistent multimodal understanding
2. Controllable online learning and adaptation
3. Verifiable reasoning and long-horizon planning & execution
4. Calibratable reflection and metacognition
5. Strong cross-task generalization

How should individuals embrace this?

1. Upgrade your mental model: Recognize AI’s evolution from “chatbot” to “task executor.” Start delegating lengthy, repetitive tasks—like weekly reporting or content workflows—to AI (e.g., via Claude Code’s skills).
2. Treat AI tool fluency as non-negotiable: As Yao Shunyu put it, those who wield AI tools will replace those who don’t—just as programmers replaced slide-rule users. Make it habitual.
3. Master context provision: AI’s value scales dramatically with the quality of input you give it. Learn to feed it your chat history, preferences, and environmental signals—so it knows you, and delivers truly personalized output.
4. Stay patient and persistent: Tang Jie’s “coffee spirit” applies here too. AGI is a decades-long race. Ignore short-term hype cycles. Invest steadily. Dare to explore unconventional, non-consensus ideas.

A New Hiring Framework

We’ve refined our hiring criteria and methodology to reflect the Agent Era.

Who we’re looking for:

1. Real, demonstrable AI mastery—not just familiarity, but results
2. Ability to translate new tools into measurable efficiency gains or structural business change
3. Proven capacity for sustained learning and methodological rigor—not one-off luck

We seek colleagues who treat AI as a productivity lever—and proactively redesign their own workflows around it.

Interview Standard One: AI Competence (Weight: 50%+)

1. Evidence of standout AI application

Ask directly:

• “Tell me about a real case where you used AI to significantly improve your work’s speed, quality, or impact.”
• Bonus if it was self-initiated—not mandated by your company.

What matters isn’t which model you used or how fancy your prompt was—it’s whether AI fundamentally changed how you worked.

2. Has AI become a tool, not just a chatbox?

Look for evidence of reusable, persistent AI-powered systems:

• An automated data-cleaning + weekly report generator
• A semi-autonomous content creation → fact-checking → publishing pipeline
• An AI assistant that handles 30–50% of your recurring manual labor

This is the critical threshold:

• Baseline: AI = conversational interface
• Advanced: AI = embedded component of your production system

3. Clarity of underlying thinking

Probe deeply on three points:

1. What pain point drove you to build this?
2. How did you decompose the problem—and decide what part AI should own?
3. What was the outcome? Did you iterate?

From these answers, we assess: depth of AI understanding, logical structuring ability, and realistic awareness of AI’s limits.

Why prioritize AI competence? Three reasons:

1. Reveals attitude: fear, performative trend-chasing, or genuine view of AI as tameable infrastructure
2. Signals learning agility: self-driven adoption and weaponization of new tools
3. Tests boundary awareness: knowing when to deploy AI—and when humans must step in

Ultimately, we look for candidates who’ve replaced ≥50% of their own repetitive work with AI—not once, but sustainably.

Interview Standard Two: Operational Strength (Pick one)

No need for universal excellence—just one clearly demonstrated strength, backed by a vivid case:

1. Operations (content, growth, users, channels)
   • Clear objective-setting
   • Metric decomposition skill
   • Systematic optimization experience
2. Project Management
   • Ability to coordinate cross-functional teams
   • Comfort navigating ambiguity and driving progress
   • Experience bridging role-specific communication gaps
3. Data Analysis
   • Data-informed decision-making (not just reporting)
   • Full feedback loop: analysis → action → validation
   • Capacity to distill complexity into clear, data-grounded narratives

One requirement: This strength must be showcased with a real, tangible case—as compelling as any AI story.

Final Screening Principle (Critical)

Does the candidate rapidly adopt and operationalize novelty?

If someone:

• Actively restructures their workflow around AI
• Uses it in ways distinct from “typical users”
• Can articulate their reasoning transparently

…then their learning velocity, adaptability, and long-term ceiling are almost certainly high.

Start interviews with this lean, direct question:

“Tell me about a time you used AI to meaningfully change how you work.
How did you do it before AI? What changed most after?
Ideally, it’s something you built—a tool or process.
Most importantly: What were you thinking—and what actually shifted?”

What Makes a Great Salesperson

The hardest part of selling has never been polishing your pitch or speaking fluently.

It’s detecting the customer’s unspoken, deeper needs—and responding with flexible, authentic resonance.

This gap widens dramatically with standardized products.

Three exceptionally difficult, defining challenges:

1. Rapidly uncovering the real need

Customers voice surface-level concerns: price, features, competitors, process. But those rarely drive decisions.

What actually moves them? Risk aversion, accountability pressure, anxiety about outcomes, fear of failure—or hunger for success.

Great sellers spot these in just a few exchanges. That’s the first inflection point.

2. Delivering personalization within standard constraints

The product doesn’t change. Its features don’t change. But every buyer’s context, priorities, and psychology do.

Top sellers make customers feel the solution was built for them—without altering the product. Weak sellers recite specs robotically, delivering identical monologues to everyone.

3. Making the customer feel ownership—not persuasion

This is the hardest: creating a dynamic where the buyer feels, “This is what I need,” not “You made me need this.”

When buyers sense they’re being led, sold, or pushed, resistance and suspicion linger. When they believe they arrived at the decision themselves, commitment solidifies—and deals close cleanly.

In recent sales engagements, the contrast was stark:

• Average sellers launch into “output mode”: relentless feature-dumping, case-study reciting, and self-promotion. In today’s overloaded, skeptical world, this often backfires.
• Exceptional sellers say little—but listen deeply. They’re the ones who know what the customer is thinking, fearing, and wanting—before the customer says it aloud.

Ten Ways Startup Death Happens

1. Founder distraction: Too many initiatives → organizational entropy
2. Founder stagnation: Cognitive lag behind tech and market shifts
3. Cash collapse: No viable revenue or funding path
4. Team decay: Erosion of expertise and execution capability
5. Strategic drift: Constant pivots, no north star
6. False innovation: Heavy investment in fake problems or pseudo-solutions
7. Bloat: Too many managers, too few doers
8. Financial mismanagement: Uncontrolled spending, poor cost structure
9. Broken business model: No path to sustainable profit
10. Core team fracture: Co-founder or key talent implosion

Strategy Must Be Steady—Tactics Must Be Fast

In startups, clarity on direction, model, and north star metric is non-negotiable—and shouldn’t shift lightly.

Within that stable frame, tactics thrive on speed:

1. Anchor the unchanging

• Define vision and core objectives explicitly
• Use them to set firm boundaries—clarity here prevents drift

2. Decide what you will and won’t do

• “Yes” means focus. “No” means protection.
• Conserve bandwidth. Resist scope creep.

3. Iterate relentlessly within the frame

• Test fast. Fail small. Learn faster.
• Validate assumptions with micro-experiments—not grand theories.
• Let small wins inform bigger bets.

Six AI Mindsets for 2026

1. Use the best models
   Yes, costs may rise—but so do possibilities. The real ROI isn’t lower bills—it’s your saved time and mental energy. Try Claude 4.5, Gemini 3, or GPT-5.2.

2. Adopt and master elite tools
   Don’t chase every new app. Pick 3–4 professional-grade tools (e.g., Claude Code, VS Code, Warp, Raycast, Cursor) and go deep. Build durable habits. Stay curious about others—but anchor in excellence.

3. Bring AI into all thinking—not just “big tasks”
   Use top-tier models for daily reflection, conversation, judgment calls—even casual planning. Your everyday AI dialogue quality shapes your cognition.

4. AI First
   Before writing anything—idea, spec, strategy, email—co-create with AI first. Then refine with your judgment.

And dare to build with AI: landing pages, internal tools, browser extensions, apps. I built both an AI browser and mobile app—both shipped fast and worked well.

This isn’t about offloading work. It’s about reconstructing your capability stack—and upgrading your productivity architecture.

1. Ignore the “dumb outputs”—focus on your judgment
   Wrong code? Low-quality drafts? Off-base opinions? All normal. Like working with a brilliant but fallible human, AI’s value lies not in perfection—but in your ability to evaluate, correct, and steer.

2. Cultivate aesthetic discernment
   True taste takes time—but accelerates with deliberate practice:

   • Study masters in your field—deeply, repeatedly
   • Analyze world-class work in your domain: deconstruct it, reverse-engineer its logic
   • Refine your eye for quality: Is this website elegant or cluttered? Is this message dense or flabby? Does this solution feel restrained—or cheap?

Here, “aesthetic” isn’t about art alone. It’s your calibrated sense of what’s good, effective, and elevated—across interfaces, language, logic, and execution.

• How do I understand aesthetics and taste? How can I define and rapidly improve them across different domains?

• What not to do is especially important—what are the things I absolutely refuse to do?

• Among the tasks I personally completed this week, which ones can AI already handle 100% autonomously?

• If a task doesn’t bring me joy—but I’m still doing it—why? And why do I choose to persist instead of honoring my inner voice?

• By the end of 2026, what three words would I want others to use when describing me?

Where Did My Time Go in 2025?

On December 31—the last day of 2025—it snowed heavily in Xi’an. After wrapping up one-on-one conversations with the Xi’an team, I rushed to the airport.

My flight to Beijing was delayed by two hours—but eventually took off.

Mid-air, I read one-third of A Brief History of the Mind, then remembered something unfinished: my 2025 personal time-category audit.

I opened my time-tracking app and resumed the calculations.

Here’s how 2025 broke down:

1. Exercise:
    • Total duration: 461 hours 29 minutes
    • Sessions: 1,066
    • Running distance: 1,388 km (including 4 marathons)

2. Sleep: 2,737 hours (avg. 7.5 hrs/day)
3. Learning: 564 hours (tracked via WeChat Reading and GetIt apps)
4. Writing: ~30 minutes/day (average)
5. Family time: ~3.5 hrs/day (estimated—time spent at home or on household matters, excluding sleep, exercise, learning, and writing)
6. Socializing: 438 hours (avg. 1.2 hrs/day)

The year had 8,760 hours total. Allocated as follows:

• Sleep: 31%
• Exercise: 5.2%
• Learning: 6.5%
• Work: 33%
• Writing: 2%
• Family: 15%
• Socializing: 5%
• Other: 2.3%

After finishing the tally, I felt deeply reflective.

The most significant shift in 2025 was heightened sensitivity to time—especially lower tolerance for “noisy activities that yield no accumulation.”

In 2026, I believe I can use time even more effectively—not by doing more, but by staying vigilant against “accumulation-free noise” and patient with “slow, certain long-term value.”

That discipline may hinge less on what I do, and more on what I refuse to do.

Manus Peak’s AI Startup Insights

A highly recommended interview (strongly endorsed by Xiangyang): youtube.com

This is Zhang Xiaojuan’s deep-dive conversation with Peak, co-founder and Chief Scientist of Manus. Dense, insightful—and deeply rewarding to watch in full.

Here are the points that resonated most:

1. Physical and mental health, common sense, and rationality are rare—and increasingly valuable—competitive advantages in the AI startup era. I’ve treated cultivating these foundational qualities as my top personal practice over the past two years. They sound simple; living them consistently is anything but.
2. To assess a company’s “taste,” look at its internal metrics.
3. Building products lets you define what “good” means—for yourself.
4. Leading LLM companies will inevitably build both models and applications.
5. The claim that “every traditional domain can be reinvented with AI” reflects intellectual laziness—not insight.
6. What not to do matters profoundly. In AI startups, the biggest barrier is shifting from “capability” to “restraint.”
7. Voting within teams often distorts alignment—it’s a form of team alienation.
8. A general-purpose agent product must satisfy three elements: user, environment, and model—and respect their boundaries.
9. Getting 1,000 small things right matters far more than nailing three “big” ones. Underlying this is a core truth: systems thinking > heroic narratives.
10. For issues lacking consensus, act first—then let results guide the team, rather than endless debate.
11. AI and humans are fundamentally different: AI is a non-intentional capability system; humans are value-driven decision systems.
12. Today’s AI resembles traditional manufacturing—scalable, process-driven, and infrastructure-heavy.
13. Prioritize net gain: After doing something, is the whole system measurably stronger?
14. In the mobile internet era, the hard constraint was user time (finite daily hours). In the AI era, agents theoretically extend human time infinitely—so that constraint has dissolved. But what new constraints have emerged? We don’t yet know.

Manus was recently acquired by Meta for several billion dollars—just over a year after founding. An extraordinary achievement.

The founding team is fascinating: six serial entrepreneurs, all deeply cross-disciplinary. Though roles have since specialized, their shared background keeps communication friction near zero—and makes consensus unusually easy to reach.

No matter how fast AI evolves, rhythm remains essential.

Physical stamina, judgment, and systemic capacity won’t become obsolete just because AI gets smarter.

Holding onto key “slow variables” amid accelerating change—that’s a rare and vital skill.

Professional Intuition

Our team spent several days analyzing a key metric. My immediate reaction: “This number is wrong.”

I couldn’t articulate why—but my gut said the final output shouldn’t look like that.

Both operations and engineering colleagues re-ran analyses repeatedly. Their conclusion: “The data checks out.”

So I pulled raw logs and rebuilt the analysis using my own logic—including partial log validation. That uncovered a critical bug—one that had skewed the entire dataset.

This reminded me: deep immersion in a domain cultivates professional intuition—a subtle, powerful ability.

You glance at a number and know it’s off—even before your conscious mind surfaces the reasoning.

And remarkably, that instinct is usually right.

Professional intuition arises from having seen “normal” so many times that “abnormal” triggers an automatic alarm—even without explicit justification.

It grows from full-stack understanding of the business, internalized into reflex.

It’s forged through cycles of missteps, post-mortems, repeated missteps, and deeper post-mortems.

These experiences settle into the subconscious, forming an autonomous early-warning system—precious and irreplaceable.

Cultivating it demands high-quality daily practice: thoughtful reflection on every detail, rigorous attention to anomalies, and real engagement with challenges—not shortcuts, not rote repetition.

What defines a true expert?

I believe professional intuition is one of the clearest markers.

Whether someone qualifies as an expert in a field hinges significantly on whether they’ve built this kind of deep, embodied intuition.

10 Counterintuitive Truths About Vibe Coding

From Steve Yegge, a veteran U.S. software engineer:

1. Pure coding is becoming low-value manual labor. In the AI era, coding speed ≈ typing speed—and value rapidly approaches zero. The worst future engineers won’t be those who can’t code—but those who only code.
2. The most resistant to AI-assisted programming are often engineers with 10+ years’ experience. Deeper expertise creates greater inertia—especially among those with “code purity” instincts, whose past edge (mastery of syntax/frameworks) now feels irrelevant.
3. IDEs are no longer human primary interfaces. They’ll persist—but increasingly as tools for agents, not people.
4. Tools like Claude Code and Cursor aren’t the future—they’re the last generation of IDEs. They push the old paradigm (humans writing code) to its limit. The real shift is: humans → manage agents; agents → write code.
5. Trusting AI ≠ blind faith. True mastery means predicting how it will fail. Steve’s “2000-hour rule”: trust emerges from predictability—not raw capability.
6. Treating AI as a “smart colleague” is dangerously misleading. Its gravest errors stem from anthropomorphism—it feels no hesitation, reverence, or awareness that production environments are sacred.
7. In multi-agent systems, the bottleneck isn’t generation—it’s merging. Tenfold efficiency won’t deliver tenfold speed. It delivers merge hell, architectural fragmentation, and exponential decision conflict.
8. Rewriting code is becoming more rational than refactoring. The old dogma—“Never rewrite”—is outdated. Agents can generate clean, coherent systems from scratch faster, safer, and more reliably than patching legacy “shit mountains.”
9. Core engineering skill is shifting from architecture to orchestration: defining agent roles, scoping tasks, setting acceptance criteria and rollback protocols, and resolving agent conflicts. Engineers are becoming agent conductors.
10. Programming education must move beyond syntax. You can skip learning for loops—but you must grasp functions, classes, state, constraints, and capability boundaries. “Knowing how to program” in the AI era means understanding systems, capabilities, and verification—not memorizing syntax.

✨ 2025

My 2025 self-inquiry list—covering habits, energy, decision-making, AI, entrepreneurship, assets, compounding, family, freedom, and education:

1. Habits
   • Which three habits contributed most to tangible outcomes this year—and which three merely made me feel productive?
   • Is there a habit I could stop now, endure short-term discomfort, and emerge stronger long-term?
   • Do my habits amplify my judgment—or mask my indecision?
2. Energy
   • What did I do with my top 20% highest-energy hours this year?
   • Which activities drained energy without meaningful return?
   • If I could guarantee only three high-quality hours per day next year—who or what would get them?
3. Decision-Making
   • Of my three most consequential decisions this year, how many held up? Why?
   • Am I using tactical busyness to avoid confronting strategic ignorance on key issues?
4. AI
   • Is AI for me an efficiency tool, a capability amplifier—or a catalyst for identity reshaping?
   • Which skills I built this year would still hold value without AI?
5. Entrepreneurship
   • Is my current venture building long-term momentum—or quietly depleting personal reserves?
   • Am I running “what I’m good at”—or “what I can’t let go of”?
6. Assets
   • Where did I trade time for money this year—and lose optionality in the process?
   • Have I developed a more systematic understanding of assets—and begun building genuinely healthy ones?
7. Compounding
   • What actions show no visible results in one year but will define my trajectory in three?
   • Am I consistently investing in at least one compounding system that doesn’t rely on luck?
8. Family
   • Is my current pace eroding future family relationships?
   • To me, is family a sanctuary, a duty—or an overlooked co-founder?
   • If my child modeled my behavior (not my advice), would I be comfortable with that?
9. Freedom
   • Which choices are irrevocably narrowing my future options?
   • If I wanted to live differently in five years, which doors am I closing right now?
   • Is my current busyness moving me toward freedom—or postponing the reckoning with unfreedom?
10. Education
    • In the AI era, which abilities deserve lifelong cultivation—and which should we stop teaching altogether?
    • If my child copied how I spend my time, not what I say, would I be satisfied?
    • If exams vanish as proof of worth in 20 years, what should I nurture today?
    • Which of my “obvious” educational beliefs are just survivorship bias from my own path?
    • If future education proves me wrong, where am I most likely mistaken?

The AI World of 2026

The venture firm a16z recently released its Big Ideas 2026 series—a set of forward-looking reports synthesizing insights across investment teams. Its central thesis: AI is no longer just a tool. It’s becoming an environment, a system, and—increasingly—a parallel agent acting alongside humans.

This shift is the essential lens for understanding the next decade—not just in tech, but in education, entrepreneurship, careers, and trust itself.

1. Education: AI-Native Learning & Personalized Ecosystems

Classrooms and lecturers won’t disappear—but their role will transform radically. Education is shifting from passive knowledge transfer to AI-driven learning ecosystems and talent engineering.

The first true “AI-native university” will launch in 2026—not just teaching about AI, but orchestrating curriculum, research collaboration, and talent development in real time via adaptive AI systems.

Three foundational changes:

• From passive learning → self-adapting learning systems: AI delivers truly individualized pacing, feedback, and scaffolding—no two learners follow identical paths.
• From teacher-as-transmitter → teacher-as-system-guide: Educators focus on cultivating higher-order judgment, ethics, and synthesis; AI handles knowledge delivery, assessment, and iterative feedback loops.
• New talent paradigm: “AI collaboration fluency” replaces rote knowledge mastery. Core skills include task design for AI agents, output verification, and reasoning-chain auditing—not memorization.

2. Entrepreneurship: From Tools to Intelligent Execution Systems

Traditional startups treat software as a UI layer or information dashboard. Big Ideas 2026 argues the next wave moves beyond isolated AI features toward multi-agent systems and coordinating intelligent teams.

AI isn’t just speeding things up—it’s rebuilding the execution layer itself.

That means product strategy pivots from “generating responses” to enabling intelligent execution, cross-process coordination, and autonomous decision engines. Founders won’t just add AI features—they’ll architect AI agents that actively intervene in workflows, complete tasks end-to-end, and iteratively optimize performance.

3. Career Development: AI Collaboration & Systems Thinking

As AI evolves from assistant to collaborator, professional value is redefined—not by what you know, but by how you orchestrate human-AI systems.

Four dimensions will define competitive advantage:

• AI collaboration design: Structuring complex, cross-module tasks for AI agents—not just calling an API, but composing agents, defining handoffs, and managing dependencies.
• Output review & explainability: Interpreting AI reasoning, validating logic, tracing evidence chains, and stress-testing conclusions—not accepting outputs at face value.
• Cross-modal product & process design: Designing for AI that understands audio, video, images, sensor streams, and text—not just prompts, but coherent multi-sensory instruction sets and evaluation frameworks.
• Systems thinking & workflow redesign: Moving beyond skill stacking to fundamentally re-engineering how work flows through organizations—so human and AI capabilities multiply, not merely add.

4. Healthcare: From Treatment to “Prevention + Health Action Engine”

a16z introduces a telling metric: Health MAUs (Monthly Active Users)—shifting focus from patients with disease to people managing health.

This signals three structural shifts:

• Healthcare’s center of gravity moves from treatment-centric models to health management and outcome accountability.
• Health services become continuous experiences—not episodic transactions—with embedded engagement, feedback, and adaptation.
• Health data becomes strategic infrastructure: AI uses real-time monitoring, predictive modeling, and personalized nudges to drive sustained behavioral change.

5. The Future: Multimodal Systems & Silent Interaction

A unifying thread across the report: AI is moving from interface-driven to environment-driven.

• AI will parse the internal structure of video—not just generate captions or tags, but understand scene transitions, causal sequences, and emotional arcs.
• Applications evolve from “help me do X” to “understand me / see me”—inferring intent, context, and needs from behavioral traces, interaction history, and multimodal signals.

6. Trust & Infrastructure: Privacy as the New Moat

In the crypto trends section, the report makes a sharp observation: Privacy is becoming the strongest moat for encryption and blockchain.

Why? Because migrating secrets is far harder than migrating value—creating a new kind of network effect rooted in confidentiality.

Future value and identity flows won’t depend on centralized gatekeepers—but on verifiable protocols, privacy-preserving mechanisms, and open economic architectures.

Key 12-Month Takeaways

• Education will restructure around AI-native learning ecosystems.
• Entrepreneurship will pivot from efficiency gains to intelligent coordination and workflow value.
• Career relevance will hinge on AI collaboration fluency and systems redesign ability.
• Healthcare products will shift from diagnosis/treatment to lifelong health lifecycle management.
• Trust infrastructure will migrate from institutional rules to technically enforceable, privacy-aware protocols.

Original reports:
a16z.com
a16z.com
a16z.com

How to Spot a Real Expert

In the AI era, trustworthy expertise matters more—not less. But distinguishing genuine experts from convincing impostors requires attention to four markers:

1. Rapid elimination of noise: They instantly dismiss irrelevant options—not because they’re dogmatic, but because they recognize what doesn’t matter, letting them zero in on the core issue.
2. Judgment under uncertainty: They make useful calls even with incomplete data—by identifying primary variables and dominant tensions, not waiting for perfect inputs.
3. Actionable clarity: Their advice lands concretely—not as abstract principles, but as steps, heuristics, or constraints that can be applied tomorrow.
4. Explainability + verifiability: They don’t just state conclusions—they walk through why, and invite scrutiny, testing, and post-hoc review.

Beware of “pseudo-experts”: They sound authoritative, cite widely, and rarely err—but never commit to specific judgments, actionable methods, or measurable outcomes. Their currency isn’t insight—it’s information asymmetry.

2025 in Review

In 2025, I interacted with ChatGPT over 10,000 times. Its memory function and steady model improvements were the main reasons I kept returning—consistently.

Near year-end, I asked it to summarize my 2025. The result was uncannily accurate—not just in content, but in tone, rhythm, and depth. That level of coherence and contextual awareness remains striking.

AI Productivity Survey Report

This summary draws from a survey of 1,750 tech professionals—a grounded, practitioner-level view of how AI is actually being used today.

1. Overall satisfaction?
   Most respondents report AI exceeding expectations, significantly lifting both work quality and efficiency. Time savings average half a day per week.

2. How different roles use AI:
   • Founders: Treat AI as a “thinking partner”—for strategy, vision, product framing, and risk mapping.
   • Product Managers: Benefit most on the output side: PRDs, wireframes, stakeholder comms.
   • Designers: Gain traction in adjacent areas—user research synthesis, copywriting, concepting—but visual design automation remains limited. Pixel-perfect fidelity and stylistic consistency still resist one-click replacement.
   • Engineers: Demand has shifted from “write code” to “write after code”: AI is increasingly expected to handle docs, tests, code reviews, and governance—freeing engineers for architecture and complexity management.
3. Tool preferences by role:
   • ChatGPT dominates among PMs, designers, and founders.
   • Engineers lean toward specialized tools like Cursor and Claude Code.
4. Where time is saved:
   • High-frequency: Writing & rewriting, research synthesis, rapid prototyping, structured doc generation.
   • Engineering “post-code” work: Test case generation, API documentation, linting, static analysis—building audit trails.
   • Founders: Competitive scanning, counterfactual stress-testing, strategic framing, org design support.
5. Bringing AI into upstream thinking & research:
   • Process formalization: Cluster topics, map evidence chains, flag counterexamples, link sources—and demand transparent reasoning.
   • Question-framing discipline: Prompt AI to structure outputs as hypothesis → evidence → counter-evidence → conclusion.
   • Human veto power: Reserve final judgment for people—use AI for alignment, synthesis, and option-scanning.
6. Measuring AI ROI:
   • Go beyond “hours saved”: Track quality scores, rework rates, defect density, alignment speed, and delivery cadence.
   • Audit evidence transparency: Are sources cited? Counterpoints considered? Reasoning traceable?
   • Review by iteration: Each week, log where AI intervened, what worked/didn’t, and adjust next week’s prompts or workflows.
7. Key shifts ahead (2026–2027):
   • From “output acceleration” → “collaborative ambiguity resolution”: AI as a long-term thinking partner for ill-defined problems.
   • Vertical workflow maturity: Engineering “post-code” pipelines, design system governance, standardized research evidence chains.
   • Organizational AI governance: Formal policies, risk controls, audit frameworks, and redesigned talent roles.

Full report: lennysnewsletter.com

Recommended AI Tools

A few I’ve found especially effective lately:

1. Banana Slides — Open-source AI-powered presentation tool built on nano banana pro. Truly liberates PPT creation.
   github.com

2. Stitch — Google’s new AI prototyping tool. Turn an app idea into a clickable, interactive UI prototype in under a minute.
   stitch.withgoogle.com

3. NotebookLM — Google’s AI-native knowledge companion. Universally praised for learning and knowledge organization. Upload documents, then generate reports, slides, tables, mind maps, audio summaries—even video recaps—all from your own sources.
   notebooklm.google.com

4. Antigravity Tools — A lightweight manager for Google’s Antigravity platform. Solves the frequent auth hang-ups by enabling seamless, one-click account switching.
   github.com

This week’s reflective questions:

1. What are customers really paying for—and do I have evidence, not just opinion?
2. Where does my current need for “certainty” come from?
3. What is the single most important health metric I want to improve this week—and what’s the smallest action that moves it?
4. Did I do something this week that made me feel more free? What was it?
5. What belief or attachment should I let go of this week—and what would releasing it truly mean for me?

Turning Inward: Rewriting Your Reward Function with Slow Variables

Full presentation available upon adding Yao Jingang on WeChat.

“Turning inward: rewriting your reward function using slow variables”
has been my biggest shift in attention strategy over the past year.

Specifically: the most important thing I track daily isn’t external output—it’s health-related metrics: VO₂max, HRV, average sleep heart rate, weekly aerobic minutes, total sleep duration.

These share a key trait: they’re slow variables—highly accumulative.

They don’t generate cash flow directly—but they set the upper limit for every fast variable above them: emotional stability, depth of focus, clarity of judgment, and sustained energy. All of these run on the same underlying biological infrastructure.

Because they’re slow and cumulative, they determine the ceiling of everything else—and because they’re systemic, they reliably improve when you optimize deliberately. Their controllability is high.

This shift is fundamentally turning inward.

It doesn’t produce “I must be more disciplined.” Instead, it triggers a cascade: more stable sleep, cleaner eating, lower emotional reactivity, faster entry into deep work, sharper sustained thinking—and a more grounded, resilient optimism.

When my attention anchors here, natural behaviors follow: I stop meaningless late-night scrolling earlier; I step back from alcohol and ultra-processed foods; I guard low-distraction deep-work blocks fiercely; I begin making decisions for tomorrow’s energy, not today’s dopamine.

These changes aren’t enforced by willpower—they emerge from intrinsic feedback. Metrics improve → I’m motivated to keep going. Metrics dip → I’m inclined to course-correct early.

That’s a higher-order kind of self-consistency.

As for external income? Because these foundational elements improve, overall returns—however defined—will inevitably rise. But I no longer fixate on the number.

The deeper principle: never outsource your mood, judgment, or rhythm to short-term gains.

Three things are fundamentally outside our control: macro market conditions, technological paradigm shifts, and individual luck.

What is within our control? The stability of our physical system, the speed of our emotional recovery, and the consistency of our cognitive output.

By centering attention here, I’m doing one thing: reducing the damage volatility inflicts—and increasing my capacity to catch positive opportunities when they appear.

The Truth About AI in Ad Campaigns

Full presentation available upon adding Yao Jingang on WeChat.

AI’s impact on advertising campaigns has been overestimated—for many mature marketing teams.

Why? Because the dominant cost driver in paid acquisition isn’t human labor. It’s the bidding mechanism itself.

In a mature, scaled campaign, costs break down roughly as follows:

• 90%: auction-based media cost (CPC/CPM)
• 10%: human-driven operational cost (account management, creative production, analytics, iteration)

Industry variance exists, but the ratio holds. Platform auction logic—not people—is the mountain.

Key insight #1: AI cannot systematically dismantle that 90%.

Why? Because bidding is a dynamic equilibrium.

If ROI remains viable, budgets expand. If you monetize traffic more efficiently, you’ll bid more aggressively to capture more volume. And platform CPC/CPM mechanisms naturally push prices toward each advertiser’s “tolerable ceiling.”

Yes—AI may temporarily lower your cost per impression. But that efficiency gain is rapidly absorbed by competitors’ improved efficiency and platform pricing dynamics—pulling prices back toward that ceiling.

Platforms won’t cut you a break just because you use AI. They maximize their revenue—not your customer-acquisition cost.

AI lifts creative output, analytical speed, and campaign iteration—but the auction system redistributes those gains via higher bids and fiercer competition.

So where does AI deliver reliable improvement? In that 10%: human-side efficiency—content production, script generation, asset iteration, copy rewriting, bulk A/B testing, data interpretation.

Here, AI can shave 10% off cost or boost throughput by ~10%. That’s real—but marginal.

That’s why I say the change “isn’t as big as imagined”: even perfecting that 10% only moves the needle modestly on the whole.

But AI’s second, more consequential impact lies in tempo:

Campaign success often hinges not on absolute efficiency—but on who first discovers and scales a replicable, effective model.

AI accelerates material production, experimentation, strategy convergence, and scaling velocity. Speed is a moat.

Teams that move faster gather better learning data sooner, lock in winning creative + audience combinations earlier, and punch through budget caps before others—even capturing brief windows of competitive advantage.

In other words: industry-average CAC may not fall—but top players scale faster, while mid- and long-tail players fall further behind. Market concentration increases. That’s the direction we’re already seeing.

There’s also an underappreciated leverage point: AI may not lower CPC—but it can lift CVR, AOV, or LTV. That raises your acceptable CPA ceiling. With higher willingness-to-pay, you can outbid others for inventory they simply can’t afford.

Superficially, acquisition isn’t cheaper—but your growth ceiling expands, and your scalability lifts.

So AI’s biggest impact on advertising isn’t “making it cheaper.” It’s making stronger players bolder, faster, and more stable—accelerating market divergence.

And if you want >10% efficiency gains? Don’t stop at the 10% inside the campaign team. Apply AI outside the auction: optimize conversion flows, sales handoffs, retention loops, attribution models, and first-party data closure. These levers don’t yield marginal tweaks—they lift your entire sustainable growth ceiling.

Paid acquisition is an auction market. In auctions, “efficiency dividends” get priced in quickly.

Lasting advantage comes from what others can’t copy: proprietary data, product differentiation, brand equity, channel structure—and a uniquely fast, robust AI-powered growth system.

Systems Thinking

Systems thinking means focusing on relationships and patterns—not isolated events or local variables.

Without it, we easily inflate the importance of a single tactic or insight—while ignoring systemic gaps.

In The Fifth Discipline, systems thinking is defined as a framework for seeing interconnections, not separate things—and patterns of change, not static snapshots.

In Thinking in Systems, a system is defined as:

“A set of interconnected elements, organized consistently to achieve a purpose.”
It consists of three parts: elements, connections, and purpose/function.

This definition matters—because many fixate on elements (a tool, a metric, a person), while behavior is actually governed by connections (information flows, rules, incentives, constraints) and the purpose function.

Two foundational building blocks of systems thinking are feedback loops and time delays. As MIT’s system dynamics literature puts it bluntly: Feedback loops are the basic structural element of all systems—and nearly all dynamic behavior stems from them.

Systems thinking is a habit: noticing how actions ripple across time to shape outcomes.

Many people’s “sharp edges” are local optima—excellence in one element. System-level wins, however, hinge on bottlenecks, feedback structures, delays, and governing rules.

A point matters—but its real weight depends on whether it sits at a bottleneck, alters a feedback loop, or reshapes a core rule.

How to build systems thinking? Try this minimal practice: re-price the “point” inside its system. Steps:

1. Clarify the system’s purpose function
   What are we optimizing for? Growth? Profit? Retention? Delivery reliability? Physical stamina? Without clarity here, any “point” risks being inflated by narrative.

2. Map 5–8 steps of the chain—focusing only on connections
   Sketch key information flows, decision gates, and incentives/constraints. Don’t rush to populate elements—connections often matter more than components.

3. Identify one reinforcing loop and one balancing loop
   Reinforcing loops amplify growth (“more users → more data → better product → more users”). Balancing loops pull toward equilibrium (“rising costs → reduced spending → lower costs”). Naming both dissolves many local debates.

4. Mark time delays
   Delays are where systems hide their teeth: “Nothing’s working” may mean “it hasn’t kicked in yet”; “This worked!” may reflect an old cause finally surfacing.

5. Test the “point” for two weeks—then recalculate its weight
   Track input, output, variability, and reproducibility. Systems thinking culminates not in elegant completeness—but in verifiable, approximate causality.

AI and Tools

In Future Silicon World, our latest episode explored AI and tools—with a seasoned AI practitioner sharing sharp insights.

Recommended AI tools:

Core takeaways:

1. Productivity leap: One person = one army. An AI-augmented consultant now produces 30 assessment reports/day—work once requiring full teams, with zero coordination overhead.
2. Handcrafted tools: End industrial mediocrity. With Claude Code, build a custom file-cleaner or config tool in 25 minutes—solving niche problems big software ignores.
3. Aesthetic moat: Quantity breeds quality. When AI output converges, your taste becomes your sole differentiator. Study Apple’s design, visit museums, watch masterful films—let intuition evolve through high-signal input.
4. Asset redefinition: Stop mining digital landfills. 100K auto-generated words are e-waste—they’re instantly reproducible. Real assets are your prompt engineering methodology and irreplaceable co-created insights.
5. SPEC-first programming: Sharpen the axe before chopping. For complex projects, don’t code yet. First, co-draft a full spec with AI—including PRD and technical docs. Spec quality dictates code accuracy.
6. The “three-fail” rule: Stop patching. If a module fails three times in assisted coding, scrap it—or rewrite the spec. Avoid the “worse with every fix” trap.
7. First-principles questioning: Pierce surface needs. Break problems into atomic units. Clarify exact purpose and expected outcome—so AI delivers on the shortest possible path.
8. Deep-tool mastery: Notebook LM + Gemini. For long-form reading or book writing, use Notebook LM’s “needle-in-haystack” search alongside Gemini’s long-context reasoning to digest entire books and draft rich outlines in minutes.
9. Efficiency hub: Raycast. Make it your AI command center—trigger instant web summaries or chat via hotkey. Eliminate window-switching friction.
10. Strategic shift: Prioritize pre-writing, not writing. Focus shifts to topic selection, logic extraction, and model architecture. AI handles 99% of text generation; humans become the 1% architects.
11. Idea-to-output speed: Turn lyrics into music in 30 minutes with Suno 4.0. AI slashes trial cost—letting anyone experience pro-tier creative joy.
12. Coexistence metaphor: Climber and Sherpa. Today’s AI is your Sherpa on Everest. You define why you climb (meaning, values, vision); AI provides expert support. Learn symbiosis—not fear.

The Path to Personal Freedom

Real personal growth isn’t about knowing more—it’s about becoming more selective with your desires, more stingy with your attention, and more rigorous with your integrity.

From Naval Ravikant’s January podcast (Modern Wisdom #44):

“Success often springs from dissatisfaction. Happiness can spring from wanting less.”

Desire’s cruelest trick? It’s an invisible contract: sign first, suffer later.

Tie your happiness to “once I achieve X,” “once I fix Y,” or “once I earn Z”—and you’ve agreed to live unhappily until then.

Worse: achievement delivers fleeting euphoria, then resets to baseline—prompting the chase for the next target.

True waste isn’t inaction—it’s being physically present while mentally absent. That moment? You didn’t live it.

Many mistake busyness for progress—chopping life into fragments of attendance, not presence. Looking back, they see projects, results, screenshots—not texture, growth, or lived experience.

Naval’s eight distilled principles:

1. Default to “no” on indecision: Good opportunities rarely demand prolonged negotiation. If you’re stuck choosing, the answer is usually no.
2. Default to “no” on new commitments: Attention requires whitespace. A packed calendar means your life is written by others—not you. We rarely lack opportunity; we lack space.
3. Choose the short-term painful option in dilemmas: Your brain exaggerates immediate discomfort to sell you long-term suffering. The harder path often leads to lasting calm.
4. Trust gut, verify with reason: For tough calls, your body often decides first. Reason explains after. True intuition is trained judgment—not impulse.
5. Treat self-respect as a credit score: It’s not pep talk—it’s your reputation with yourself. Build it: define clear ethics, keep promises to yourself, act rightly—especially when it helps others. Every kept promise deposits into your integrity account. Low self-respect means you’ve stopped believing you’ll follow through.
6. Relationships run on shared experience and values—not résumés: Does energy align? Is presence easy? Do you unite in crisis? These aren’t measurable—but they’re real. If a relationship feels like work, you’re using willpower to sustain an unfit structure. Values aren’t slogans—they’re revealed in hard choices.
7. Don’t stream global emergencies into your brain: Human cognition didn’t evolve for “worldwide breaking alerts.” Trying to care about everything hands control to anxiety. Better rule: focus only on local, actionable domains. Don’t treat attention as charity.
8. Decompose stress into conflicting desires: Stress isn’t overload—it’s desire collision: wanting approval vs. living authentically; earning money vs. having ease; freedom vs. control. Write the conflict down—and anxiety shrinks. You don’t eliminate tension—you acknowledge you can’t satisfy all wants, choose one, and accept the trade-off.

Compress all this into one sentence:
Want less. Be more present. Explain less. Deliver more. Add less. Leave more space.

Practice these well—and freedom stops being abstract. It becomes your operating system.

Original source: podcastnotes.org

This week’s introspective questions:

1. What am I spending the most time on right now—amplifying leverage, fixing weaknesses, or repeating the same loop?
2. Is my current lifestyle building endurance—or draining willpower?
3. Is there something that, the moment I stop doing it myself, causes the whole system to collapse?
4. The last time I felt strong emotion about a technology or product—did I actually study its underlying mechanics and consider its boundaries?
5. In my most critical recent decisions, which judgments did I wrongly delegate to AI—or to others?
6. What is the one thing I’m most unwilling to have questioned—and why?
7. What was the root cause of our team’s most recent visible drop in efficiency or rework: skill gaps, or unclear standards?
8. Do I clearly know each team member’s strongest capabilities—and what they’re genuinely unsuited for?

Website Value Reimagined

Traditional websites are losing traffic—and relevance—faster than ever. For many, their value has become marginal: revenue may not even cover server costs.

Two core forces drive this:

First, organic search traffic is shrinking. Users increasingly rely on AI-powered search and conversational tools—not only for convenience, but for higher-quality, synthesized answers. These interfaces rarely require clicking through to a website. As a result, traditional sites face compounding decline.

Second, AI systems demand massive amounts of training data—so web crawlers (many disguised as benign bots) aggressively scrape content. A site with just 100,000 daily human visitors might receive over 10 million daily bot requests. Several high-traffic site operators told me their engineers now spend significant time blocking AI crawlers—not out of principle, but necessity: the tiny ad revenue they earn barely covers hosting costs, let alone bandwidth overloads caused by unfiltered scraping.

So where does this leave websites? Not obsolete—but transformed. Their future role isn’t primarily serving human readers, but acting as trusted upstream sources for AI models. Think of them less as endpoints, and more as foundational infrastructure for AI reasoning.

Monetization shifts accordingly—from “click-based” to “influence-based”:

• Citation count in AI-generated answers
• Number of models citing the site as a source
• Weight assigned to its content within answer structures
• Real-world impact on user decisions

The new currency isn’t pageviews—it’s citation authority. Over the next few years, 90% of websites will pivot from “human-facing” to “model-facing.” Those who master this shift won’t just survive—they’ll control the invisible flow of attention.

Practical steps include:

1. Content: Treat every article as “knowledge assets optimized for model consumption”—structured, self-contained, evidence-grounded.
2. Engineering: Tier AI crawler access—restrict low-value pages; offer dedicated “source-ready” endpoints (e.g., concise summaries, schema-rich metadata); track metrics like crawler-to-human ratio and per-page crawl value density.
3. Business: Embed traceable touchpoints into AI answer flows—custom landing pages, lead forms, branded contact options—to convert influence into measurable outcomes.

The VF Strategy

A compelling paper—“Asking LLMs to Verify First is Almost Free Lunch” (arXiv:2511.21734)—offers a simple yet powerful insight: Don’t ask AI to generate an answer first. Ask it to verify a candidate answer—even a wrong one.

1. Why “Verify First” Works

LLMs excel at continuation—not self-skepticism. Left unchecked, they produce fluent but flawed outputs. But present them with a provisional answer (“I think X is true”) and ask them to assess it, and their behavior changes: they shift into “quality assurance mode,” a task they handle more reliably. We call this the VF (Verify-First) pattern.

2. Why It’s More Effective

It mirrors human cognition: evaluating an existing claim (“Where is this wrong?”) is cognitively lighter than generating one from scratch. Experiments show VF improves accuracy across tasks—even outperforming Chain-of-Thought (CoT) and Self-Correction, at lower cost (see Figures 7 & 9 in the paper).

The operational shift is profound:

• From generate-first → verify-first
• From think more → trust less, check first
• From flow matters → foundations matter

3. Prompting Examples

• General reasoning:
  “I suspect the answer is X (uncertain). First, verify whether X satisfies all conditions in the question. If not, identify the flaw and provide the correct answer.”
• Proposal review:
  “Here’s a draft plan. Evaluate it for logical completeness, unstated assumptions, and causal validity. If any gap exists, rewrite it fully.”
• AI-search-friendly content:
  “Assume this text will be cited by AI search as a candidate answer. As an AI verifier, assess whether it: (1) states a clear conclusion, (2) maintains internal consistency, and (3) cites verifiable evidence. If not, reconstruct it for maximum trustworthiness.”
• Agent/API design:
  “This is an initial implementation—likely flawed. Verify: Does it meet spec? Are there logic or edge-case failures? If so, deliver a corrected, production-ready version.”

But VF has boundaries:

• ✅ Best for truth-conditional problems (math, logic, compliance checks, bug detection).
• ❌ Less effective for open-ended creativity (naming, emotional resonance, worldbuilding).
• ✅ Works when verification is cheaper than generation.
• ❌ Fails when verification itself demands deep implicit knowledge—or when standards are inherently ambiguous (e.g., policy interpretation, ethical trade-offs).
• ⚠️ It’s not about making AI “smarter”—it’s about reducing plausible-but-wrong outputs. Think of it as error-reduction infrastructure, not inspiration amplification.

AI’s Hard Boundaries

1. Mathematics reveals AI’s ultimate limit: Gödel’s Incompleteness Theorem proves that any sufficiently complex formal system contains true statements it cannot prove.
2. As long as AI operates within formalizable logic, this boundary is permanent.
3. The theorem teaches us: rationality itself is bounded. Depth and breadth cannot coexist perfectly; all systematic knowledge has blind spots.
4. AI doesn’t create new limits—it scales and exposes them.
5. Its real constraint isn’t computational power—it’s what cannot be formalized: value, meaning, moral duty, exception handling, risk perception.
6. Even AGI won’t resolve this. Responsibility, ethics, and existential purpose remain human domains—not because AI is weak, but because those questions lie outside formal systems entirely.
7. Practically, AI’s boundary isn’t just what it can’t calculate—but what it must never decide.

Midlife Reckoning

While running this morning, I circled back to a question no algorithm can answer: How should life unfold after 40?

One premise is non-negotiable: without baseline financial security, talk of well-being or fulfillment is abstract. In major cities, once household income crosses a stability threshold, the strategic priority must shift—from how to keep earning to how to live well, sustainably.

Midlife is often happiness’s nadir: peak responsibility, slowest returns, declining physical resilience—and a mind still wired for youthful pace. Without conscious recalibration, satisfaction tends to erode.

Two adjustments are essential:

1. Cultivate a lifelong practice rooted in body, mind, or craft—not for immediate payoff, but for decades of compound benefit. These aren’t time sinks; they’re ballast for the second half of life.
2. Consciously downsize material desire. Comparison-fueled wanting corrodes contentment. Aging wisely means shifting focus from acquiring more to needing less—a quieter, sturdier kind of freedom.

Much midlife anxiety stems not from hardship—but from loss of agency: fragmented time, depleted energy, life feeling like passive motion.

Perhaps the deepest work after 40 isn’t proving anything—but reclaiming the steering wheel.
To govern your body. To steward your time. To name your “why.”
That kind of life—simple, grounded, chosen—doesn’t dazzle. But it holds.

The Most Suitable Day

A microfiction co-created with GPT-5.2 (860 words)

His life unfolded on “the most suitable day.”

Admission came on a rain-washed Monday—air crisp, typhoon safely diverted.
His first day at work: no subway delays, his interviewer’s coffee precisely lukewarm.
Even his wedding occurred on the city’s clearest morning in months.

People called it luck.
He knew better.

Every morning at 6:47 a.m., the reminder chimed.
He rose, drank water, detoured around construction.
When his mood dipped, a song arrived—timed, uncanny.
He rarely chose. Choices were made before he noticed them.

At thirty-five, he was late—for the first time.
Not due to traffic or tech failure.
He simply turned off the reminder.

In the team meeting, his manager glanced up.
“You’ve seemed unstable lately.”

That night, he didn’t go home.
He walked the riverbank, lights fracturing on black water—directionless.

Back in his empty apartment, the reminder pulsed:
“Detected deviation from optimal life path.”
“Initiate correction?”
He paused—then, for the first time, didn’t tap Yes.

Days later, a call: his mother hospitalized. Not critical—just needing care.
The reminder responded instantly:
“Maintain current work rhythm. Optimal visit window scheduled for Saturday.”

He stood outside her room, silent.
The prompt offered efficiency: a weekend “high-yield visit,” zero project disruption.
The alternative? Take leave. Miss a pivotal deadline.

He ignored the prompt. Stepped inside.

She smiled. “You don’t need to come. Work matters.”
He nodded—and stayed.

Three days later, the project was reassigned.
His manager said only: “The company can’t wait.”
A week later, the layoff notice arrived—calm, clinical, unsurprised.

On her discharge day, he waited in the hospital lobby.
The reminder chimed again:
“Critical income source lost.”
“Activate risk mitigation protocol?”
He didn’t look.

At home, his wife sat on the sofa.
She held a list: mortgage, tuition, monthly budget—each figure exact.
“This isn’t romance,” she said.
“You wouldn’t have done this before.”

He tried to explain. Found nothing to say.
Because she was right.

Late that night, the reminder glowed:
“Life stability below safety threshold.”
“Restore assistance?”
He sat in darkness—and felt fear.
Not of loss.
But of the terror that, if he restored it, everything would snap back into place.

Rain tapped the window.
He didn’t know what tomorrow held.
But he knew, for the first time, the price he’d paid: choosing a life with no optimal solution.

On Saturday evening, “Future Silicon World” hosted a live session via WaytoAGI, featuring keynote talks by Qiniang and Xiangshu.

Core questions explored:

Xiangshu:

1. What trait or ability is most valuable in the AI era?
2. Has traditional education become obsolete? How do we interpret the strong Chinese presence in global AI advancement?
3. Should we “push” children differently in the AI era? If yes—how? If not—how do we avoid the “theater effect” (where everyone stands to see better, yet no one gains advantage)?
4. How can parents become gardeners—not carpenters—of their children’s growth? How do we resist the impulse to “sculpt” them?
5. What are your top-level educational guiding principles?

Yao Jingang:

1. Which abilities will never be replaced by AI, no matter how advanced it becomes—and how would you cultivate them in your child?
2. Will AI make education more equitable—or widen the gap? What’s the decisive variable?
3. If you were to redesign a school natively for the AI era, what three things would you rebuild from scratch?
4. Looking 20 years ahead: what single ability would give your child a massive early advantage—and why?
5. What concerns you most about your child’s development today?
6. How do we help children understand the relationship, distinctions, and boundaries between AI and humans?

In the discussion, Qiniang, Xiangshu, Xiangyang, Nickxi, and Yuanzi shared sharp, practical insights:

1. The most essential, irreplaceable human capacity in the AI era is “aesthetic judgment + leadership”: knowing what is good, and directing AI to realize it.
2. “Critical thinking” is better framed as “discernment”—the ability to judge, filter, and choose wisely amid an ocean of AI-generated content.
3. Imagination sets the ceiling for how high someone can leverage AI; without it, AI only amplifies mediocrity.
4. Real competitive advantage comes from agency—an inner drive toward self-defined goals—not from the power of tools alone.
5. Aesthetic sense, imagination, and values form through real-world exploration, play, lived experience, and deep engagement with timeless texts—not through abstraction or instruction alone.
6. The key shift for future schools isn’t teaching students how to follow rules, but empowering them to co-create the rules.
7. The most effective learning method is “teach-to-learn, learn-by-doing, project-driven”—not passive listening.
8. Parenting must evolve from controlling to gardening: offering space, modeling integrity, and accompanying—not commanding.
9. Children must clearly understand AI is a tool, not a person—preventing emotional dependency and conceptual confusion.
10. AI dramatically expands the potential for personalized and inclusive education—but whether that potential scales depends entirely on institutional willingness to change.

The next day, I attended an international school’s parent open day. Their summary of core teacher competencies in the AI era was equally incisive:

Full AI-assisted summary of the livestream’s key points:

Learning by Doing

I recently read about Gabriel Petersson—a 23-year-old Swedish researcher now working on Sora at OpenAI. He never finished high school.

His career path defies convention:

As he puts it: “Universities no longer hold a monopoly on foundational knowledge—we can access it instantly via AI.”

From Petersson’s own account: Companies don’t care about diplomas. They care about value creation. Prove you can deliver value—and they’ll hire you.

Petersson’s success stems not from elite credentials, but from his learning method: treating AI as a second brain, starting from real problems, driving progress through real projects, and using AI to solve them—rapidly iterating, debugging, and shipping.

For example: pick a real, valuable, in-demand problem → use AI to write and debug code → fill knowledge gaps on-the-fly with AI → ship a working system → land opportunities through tangible output.

This path diverges sharply from traditional learning:

Most of us follow: grades → elite school → prestigious job → socially sanctioned career track.
That path once promised security. Today, it’s fraying—and will fray further.

Key takeaways:

1. True capability no longer lives in knowing—it lives in making things happen. In a results-oriented world, one truth holds: those who deliver outcomes always get a seat at the table. Petersson internalized this early: “Whoever ships something gets to sit at the table.”
2. The defining skill isn’t how much you remember, but whether you and AI can form a mutually reinforcing loop—each amplifying the other.
3. This era rewards self-direction and shipped work, not waiting and credentials. Most people stall at the starting line—not because they’re unqualified, but because they say: “I’m not ready yet,” “I need to study more,” “I’ll wait for the right moment.” Petersson’s logic is different: Start. Learn while doing. Turn learning into output. Your portfolio—not your résumé—is your true credential.

How to Choose Projects

In conversation with Xiangyang, we refined our criteria for selecting high-signal projects—especially tool-based ones.

A foundational principle: build a single-point system that, once cracked, can be theoretically replicated infinitely.

Operational guidelines:

1. Fully automatable + AI-native, with built-in self-growth loops.
2. Small and elegant—beneath the radar of big tech, hard to copy.
3. Few competitors; product scarcity is real.
4. Solves a genuine, urgent need—but the solution logic is nontrivial.
5. Building the core system is complex—but once validated, the model scales infinitely.

For entering unfamiliar, niche domains, validate your idea like this:

1. Start with proven personal projects or micro-studios—not greenfield invention.
2. Conduct rigorous side-by-side comparison: objectively assess whether you could outperform the benchmark—if yes, it’s likely worth pursuing.

Knowing When to Walk Away

Annie Duke’s Quitting: The Science of Knowing When to Walk Away is profoundly practical. It unpacks decision-making under uncertainty—exposing cognitive traps, building rational stop-loss frameworks, and reframing exit as strategic wisdom—not failure.

In one sentence: Quitting isn’t surrender. It’s one of the most underrated forms of rational agency.

I. Counterintuitive Core Insight

We’re taught: “Persistence equals virtue.”
Duke, drawing on behavioral economics and game theory, shows repeatedly: Failure rarely comes from quitting—it comes from persisting too long in the wrong thing.

People don’t lose because they lack skill—they lose because of two illusions:

1. Mistaking persistence itself for the cause of success.
2. Equating quitting with failure, cowardice, or weakness.

She performs vital “cognitive defusing”: stripping the moral halo off persistence—and separating it cleanly from rational choice.

II. Three Deep Human Biases This Book Helps You Fight

1. Loss aversion
   We feel the pain of loss ~2x more intensely than the pleasure of gain. This breeds dangerous behavior: “I’ll keep going—even though it’s wrong—because I’ve already lost so much.”
   It’s rampant in startups, investing, careers, and relationships.

2. Sunk cost fallacy
   Money, time, effort, emotion—all become psychological anchors.
   Rationally, only future value matters. But emotionally, we ask: “How can I walk away after investing so much?”

3. Identity attachment
   Many won’t quit—not because the project still makes sense—but because they can’t bear the label “failure.”
   So decisions aren’t made for outcomes—but for self-image.

III. The Real Tool Isn’t Judgment—It’s Pre-Set Exit Rules

The book doesn’t just say “quit when needed.” It insists: design your exit mechanism before emotion clouds judgment.
Make rationality structural, not situational.

Examples:

• Not “I’ll stop when I can’t bear it anymore,” but “If cash flow stays negative for 3 consecutive months, I pause.”
• Not “I’ll quit if it feels hopeless,” but “If we miss Milestone X by Date Y, we reassess.”
• Not “I’ll leave when I’m exhausted,” but “If my energy score falls below Z for 10 days straight, I step back.”

IV. Correcting the “Success Narrative Bias”

Society celebrates the gambler who hit big, the founder who held on till the turnaround, the athlete who overcame injury.
Rarely celebrated: the investor who cut losses early, the founder who pivoted before collapse, the professional who walked away to preserve health and clarity.

The result? Generation after generation misreads the odds—overvaluing grit, underestimating the power of timely retreat.

True strength isn’t blind endurance.
It’s having the courage to press stop—on the wrong path, at the right time.

Baiyangdian Hiking Exchange

Over the weekend, 16 colleagues from Souwai joined me for a hiking exchange at Baiyangdian—covering AI, GEO, short-video ecosystems, and the emerging efficiency class divide.

At lunch, four speakers shared distinct but complementary lenses:

• Zhang Kai: Overseas GEO & independent stores
• Han Jiangting: AI + short-video matrix strategy
• Qiao Xiangyang: AI models & tool ecosystems
• Me: Domestic GEO execution paths

Together, these four perspectives map the full terrain of today’s AI-driven traffic migration.

I. AI Search Is Rewriting “Where Traffic Begins”

Historically: user → keyword → search engine → website
Now increasingly: user → question → AI → cited source → official site or platform

1. Why AI Must Cite “Fact Sources”

AI search follows a clean, two-step logic:

• First: locate what’s widely accepted as factual
• Second: interpret, restructure, and express it

This design combats hallucination—and means: whoever becomes “fact itself” becomes the AI traffic gateway.
That is GEO—not chasing rankings, but becoming the source AI trusts as ground truth.

2. Algorithm Shift: From Keywords → Semantic Weight + Compute Efficiency

Traditional SEO optimized for keyword matching and backlink authority.
AI search optimizes for intent alignment and semantic weight.

Three critical content rules emerge:

• Authority vocabulary: AI trusts industry-standard terms and technical language—not emotive phrasing.
• Structured expression: AI prefers tables, checklists, step-by-step guides, and comparative formats—not narrative prose.
• High information density: Adjectives, hype, and persona branding are mostly noise to AI—they waste tokens and dilute signal.

This shift favors small, deeply specialized players—not just big sites with legacy authority.
You may not beat Amazon on domain rank—but you can beat it on compute-per-value ratio.

II. Overseas GEO: Independent Sites Are Back on Top

3. Why Independent Sites Dominate Overseas AI Rankings

Because in global contexts, “official site = trusted fact.”
When AI crawls overseas sources, its priority stack is clear:

• Tier 1: Official websites / independent stores
• Tier 2: E-commerce platforms (e.g., Amazon)
• Tier 3: Communities (e.g., Reddit)
• Tier 4: Media outlets

→ Simply having an independent site places you in AI’s first-tier source pool.

4. Real-World Traffic & Conversion Results

Multiple live cases confirm consistent patterns:

• Traffic growth: 5–6× increase
• Time-to-impact: noticeable shifts within 1–2 weeks
• Conversion quality: significantly higher than traditional SEO

Crucially: Though AI results rarely show phone numbers directly, they do surface brand names, product names, and official links.
People who proactively query AI have already self-filtered for intent—so conversion rates rise.

5. Overseas GEO Tools Are SaaS-Native

The standard overseas model is subscription-based SaaS (~$299/month). These tools don’t post for you—they act as GEO navigation systems: telling you what to write, where to publish, and which “seed questions” to target.
This is fundamentally different from China’s “managed service” model.

III. Domestic GEO: Not SEO—It’s SEM 2.0

A key insight repeatedly validated on-site:
Domestic GEO’s commercial value mirrors early-stage Baidu SEM—not modern SEO.
That means:

• Not “cheap traffic”
• But high-intent, high-conversion, high-LTV leads

6. Three High-Impact Execution Tactics

Domestic GEO in practice:

• Timeliness refresh: AI strongly favors freshness. Updating publication dates on legacy articles to “recent” yields immediate, visible citation lifts.
• Structural re-tagging: For large legacy sites, no need to rewrite pages—just use tools like Scanner to add semantic markup (e.g., <schema:HowTo>, <schema:FAQ>), reshaping how AI parses and cites your content.
• Natural conversion embedding: Phone numbers and URLs appear 5–10× more often in AI outputs when embedded within action-oriented context (e.g., “Call now to schedule your free audit”)—not as standalone footnotes.

7. Reframe AI Crawlers: Not a Cost—An ATM

Many companies panic when AI bots trigger hundreds of dollars in daily OSS bandwidth fees.

The correct mindset shift: You’re not being “charged”—you’re buying leads.
This is a business model pivot:

• Old logic: ad inventory (A-space)
• New logic: AI-driven lead acquisition (B-space)

You’ve shifted from selling impressions to monetizing qualified inquiries.

IV. Short-Video Matrix: The “High-Yield Sow” Model for Traditional Industries

A blunt operational reality: For traditional manufacturers, the most effective Douyin strategy remains one word—volume.

Like SEO’s keyword stacking era, today it’s: account stacking, video stacking, matrix stacking, search-coverage stacking.
Goal: Ignore trends. Capture search-driven demand.

8. AI Is Now Embedded in Sales Operations

AI has entered the sales core of traditional enterprises:

• All sales activity logged in cloud CRM
• 64-dimension behavioral analytics
• AI-generated negotiation scripts
• Call transcripts auto-converted into video script drafts

→ Content creation no longer depends on inspiration. It flows directly from real, closed deals.

V. AI Models & Tools: The Real Divide Isn’t “Understanding”—It’s “Daring to Go All In”

9. Much of China’s AI Is “Skin + Router”

The dominant pattern:

• Use top-tier models (e.g., Qwen, GLM) for intent parsing and task decomposition
• Offload execution to lighter, cheaper models

That’s why many domestic AI tools feel smart in chat—but falter on heavy lifting.

10. The Only AI Category With Real PMF Today: AI-Powered Automation Workflows

Not AI writing copy. Not AI generating images.
But: AI coding + autonomous workflow orchestration.

These tools no longer “write code”—they run processes: auto-watching videos, auto-filing taxes, auto-filling government forms, auto-uploading invoices.
You’re not using a tool. You’re hiring a digital employee.

VI. The Efficiency Chasm Is No Longer a Trend—It’s a Class Boundary

The most resonant line from the hike:
AI isn’t delivering efficiency gains—it’s creating an efficiency class divide.

Some people operate at 10× speed.
Some at 100×.
Some are already running 24/7, machine-grade throughput.

This gap isn’t widening gradually—it’s snapping open.

Hence the only advice offered on-site:
Start early. Stay quiet. Go all the way.

From AI search to GEO, from independent sites to short-video stacks, from AI coding to the efficiency chasm—
The greatest value of this exchange wasn’t what we heard.
It was confirming one thing:
The future isn’t arriving slowly. Some people have already arrived.

A friend once said over dinner: “In the future, knowing how to play will matter more than ever.”

The premise? Playing—safely.

If, in a few decades, AI dramatically boosts productivity and grants most people abundant leisure time, what will they do with it?
Most likely: play.
But “playing well” may become a new core human competency.

01

AI will divide the world into two groups: those played by systems—and those who play well.

Everyone will face a choice: Do you want agency over your own life?

If not, AI will design for you: ever-more stimulating entertainment, ever-more attentive companionship, ever-more precise emotional comfort—pleasures like “mental nutrient solution”: the more you consume, the emptier you feel.

But if you choose to reclaim agency, AI transforms into something else entirely: a creative engine—helping you design adventures, build worlds, and expand the scope of lived experience.

Truly skilled players aren’t consumers. They’re game designers.

02

Playing well is a skill.

It means generating meaning from experience, worldview from meaning, and personal style from worldview.

Someone who can’t play well defaults to consuming pre-packaged stimulation—designed by algorithms, AI, or robots. Their emotions get pulled along; their life narrative becomes a side effect of recommendation feeds.

The greatest danger ahead isn’t having nothing to do.
It’s believing you’re playing—when in fact, you’re being played.

Playing well isn’t escapism. It’s clarity.
It’s knowing the difference between experience and anesthesia:

• When ideas suddenly multiply—that’s experience.
  When you lose all sense of self—that’s anesthesia.

• When the world feels vividly clear—that’s experience.
  When time vanishes without trace—that’s anesthesia.

03

The central human differentiator in the future won’t be knowledge—it’ll be sensibility.

AI can master any skill instantly—but it cannot replace your felt relationship with the world.

The same wind blows across a field.
One person registers only: “wind.”
Another hears direction, emotion, force—in its rustle.

The gap won’t be about who knows more.
It’ll be about who experiences more dimensions in the same reality.
That’s a new kind of intelligence.

04

We’re not entering a high-efficiency era—we’re entering a high-experience era.

AI will turn the world into a vast “autonomous system”: tasks automated, workflows rewritten, even creativity outsourced.

When work is hollowed out—what proves you’re still alive?
Answer: your capacity to experience deeply.

GEO’s Consensus Mechanism

What is GEO, really?

At its core: building brand-relevant consensus with AI.

How does that work?

AI search doesn’t ask, “Who paid the most?” It asks: “Does this statement align with what the world collectively affirms?”

If yes—it gets cited, recommended.
If no—it’s ignored.

Spending money or publishing volume matters far less than being repeatedly verified, across sources, as the right answer.

AI is the first search engine that ignores ads, discounts authority scores, and trusts only consensus. In its presence, marketing slogans, brand taglines, and emotional packaging all collapse—because AI values facts, prefers structure, and demands evidence.

The more you try to “polish” your message, the less it trusts you.

Some brands never ran GEO campaigns—but because third parties referenced them consistently, with clean parameters and clear structure, AI began citing and recommending them.

In the AI world, brand identity isn’t “what you say you are.”
It’s “how others consistently describe you.”

Old branding was shouted. New branding is accumulated.

I’ve observed a trend: the more independently sourced, identical language used to describe a brand—the more likely AI treats it as the default answer.

If ten websites all call you “immersive courses + AI coding tools + ages 6–12,” AI etches you into its knowledge graph.

Not because you said it—but because the world says it.

That reveals a deeper truth: AI-era branding isn’t PR. It’s consensus engineering.

What does “consensus engineering” mean in practice? Take AI citation efficiency:

AI doesn’t cite you because your writing is elegant—it cites you because you make its job easier.

Top-performing content shares one trait: it’s rarely the flashiest—but always the most frictionless for AI to use. Examples: clear data, logical structure, atomic points, decomposable attributes, consistent semantics, rich factual grounding, and obvious information gain.

AI won’t reward beauty. It rewards usability.

That’s one principle of consensus engineering.

The harsh reality of AI search? You can no longer mask factual gaps with marketing tactics.

Now, AI asks three questions:

1. Is this true?
2. Do others say the same thing?
3. Does your content offer unique, verifiable insight others lack?

Yes—there’s noise, manipulation, and black-hat work. But the long-term trend is positive: AI is pushing the world toward answer transparency, forcing businesses to make products, services, data, content, reviews, and facts more real.

The more you obscure, the less visible you become.
The more you ground yourself in reality—the more AI cites and recommends you.

This is a foundational rewrite of information rules.
The only winners in this era? Those who turn complexity into clear, accurate facts.

GEO Effectiveness Measurement

This note itself was produced in an AI-native workflow:

1. I dictated my structured thinking on this topic into GET Notes (an AI voice-note tool), triggering automatic transcription and initial organization.
2. I fed that draft to ChatGPT for logic refinement and structural tightening.
3. I reviewed, added new insights, and finalized the piece.
4. I uploaded the final text to NotebookLM to generate a presentation—visualizing key ideas for clarity and sharing.

After implementing GEO, companies inevitably ask: “Is it really working? Can we measure it?”

Here’s a comprehensive, pragmatic evaluation framework:

I. Three Core Types of GEO Impact

1. Direct impact (quantifiable, trackable)
   Actions users take immediately after seeing your content in AI responses: clicking through to your site or landing page.
   Examples:
   • Calling a marketing phone number listed in an AI answer
   • Clicking a branded link embedded in AI output
   • Navigating via source citations in AI results

   These are hard metrics—the “hard KPIs” of GEO.

2. Brand impact (observable, but not directly traceable)
   GEO builds AI-powered mental endorsement, yielding two clear shifts:
   • Significantly increased brand exposure: users see your name repeatedly in AI answers
   • Rising brand search volume: users follow up with searches on Baidu, Douyin, or other traditional engines

3. Decision & conversion impact (hard to isolate, but critical for growth)
   AI answers accelerate decisions. At key moments, users trust AI-recommended brands more readily—boosting downstream conversion rates (i.e., higher close rates per lead of equal quality).

   Why? Because AI shortens decision paths and reduces comparison fatigue.

   This is qualitative + analogically quantified—validated through backend data analysis.

II. How Companies Should Systematically Measure GEO

Break monitoring into three dimensions: source, brand, and conversion.

Dimension 1: Source & Traffic Monitoring (directly quantifiable)
Core methods:

1. Web referrer tracking (strongly recommended)
   • Use ref parameters or tagged URLs to identify AI-platform traffic
   • Connect to CRM/order systems at the user-ID level
   → Enables full funnel: Which AI platform → which leads → how many converted → ROI?

2. Dedicated marketing phone line
   • Assign a unique number exclusively for AI-platform listings
   • Track call volume, duration, and qualified inquiries
   → A reliable proxy for app/client-side traffic.

3. AI-exclusive landing pages (BI pages)
   • Build standalone pages used only in AI responses
   • Monitor visits, dwell time, button clicks, etc.
   → Solves referrer-blindness in app environments.

Dimension 2: Brand Metrics (observable, not fully attributable)
Full attribution is impossible—but trends confirm value. Key indicators:
• Volume of brand-specific searches (brand name, brand + product terms)
• Frequency and visibility of brand mentions across AI platforms
• Shifts in users’ secondary search behavior (e.g., searching your brand after an AI answer)

Validation approaches:
• Compare 30–90 day trends before/after GEO launch
• Use AI response screenshots to verify improved placement
• Benchmark against competitors’ exposure frequency to gauge mindshare lift

Dimension 3: Backend Conversion Metrics (inferred via analogy)
Since AI influences decision efficiency, not just click-throughs, we infer impact indirectly:

1. Compare lead-to-close rates pre/post GEO
   Expect: same-quality leads convert at higher rates post-GEO; consultation costs drop.

2. Benchmark lead quality vs. other channels (e.g., Baidu, content platforms)
   If AI-sourced leads convert better, the “trust endorsement” is real.

3. Track overall sales lift for core products/courses
   For standard offerings: monitor total orders, shortened purchase paths for key SKUs.
   Why? AI’s “recommended answer” nudges users toward faster decisions—creating ripple effects across channels.

III. Current GEO Measurement Limits & Workarounds

1. No referrer data from apps
   → Use dedicated phone lines + exclusive landing pages instead

2. Can’t perfectly attribute brand search lifts
   → Apply “trend + cohort comparison”: compare pre/post-GEO search volume and cross-check against competitor trends

3. Can’t isolate pure AI-driven conversion lift
   → Focus on AI-sourced leads’ conversion efficiency, or treat overall uplift as a validated proxy

IV. Six Key GEO KPIs Every Company Should Track

1. Number of leads from AI platforms (web + phone)
2. Conversion rate of AI-sourced leads
3. Change in brand-search volume
4. Frequency of brand appearance in AI outputs (visibility)
5. Conversion-rate change for key SKUs
6. Overall marketing ROI shift (including cross-channel lift)

These six form a minimum viable measurement system:

• Direct impact → measured via source tracking
• Brand impact → assessed via trend observation
• Conversion impact → inferred via analogical analysis

Together, they make GEO’s value quantifiable, verifiable, and reusable.

Feeding this entire framework into NotebookLM generated a high-quality presentation—clearly visualizing the logic:

Eight Essential Skills

These eight capabilities deserve sustained attention and deliberate practice.

Observe the fastest-growing people around you: they rarely excel in just one area. Instead, they upgrade multiple capability curves simultaneously.

This “compound evolution” determines how deeply someone experiences the world—and how far they can go.

1. Learning how to learn: Mastering metacognition
   As the world accelerates, knowledge depreciates fast—but the ability to learn quickly grows ever more valuable.

2. Asking better questions: The fastest way to open the world
   Great questioners advance 10× faster. A sharp question signals deep understanding, clear boundaries, and a precise path forward. In the AI era, the future belongs to those who ask better questions.

3. Learning to code & use AI: The working language of tomorrow
   Coding trains logic. AI multiplies leverage. Together, they’re the most powerful personal amplifiers for the next decade.

4. Practicing science-based movement: Your energy system governs everything
   Exercise isn’t just health—it’s fuel for execution, emotional stability, and resilience. It’s the highest-ROI investment in your life. Period.

5. Learning how to earn: Value-creation and monetization
   Sustained income isn’t luck—it’s a high-bar skill requiring demand sensing, value delivery, closed-loop design, and scalable execution.

6. Cultivating focus: Owning your attention
   Attention is productivity. With average attention spans falling, focused people are rare—and inherently stronger. Not just harder-working, but more awake.

7. Mastering English: Your global interface
   English expands your cognitive horizon directly—connecting you to ideas, talent, and opportunities beyond borders.

8. Training aesthetic judgment: Raising the ceiling on choice
   Aesthetics isn’t just art—it’s choice architecture. It shapes what you build, how you communicate, what standards you accept, and how you live. Writing, product, brand, lifestyle—all are bounded by aesthetic maturity. An evolving aesthetic is, fundamentally, upgrading your world’s operating system.

These eight skills form a dynamic stack:

• Cognitive layer (1, 2)
• Tool layer (3)
• Energy layer (4)
• Value layer (5)
• Execution layer (6)
• Boundary layer (7, 8)

Integrate them—and your growth curve leaps from linear to exponential.

The 2035 Gaokao

In 2035, roughly 17 million students will sit for China’s national college entrance exam.

That year’s society will likely feel profoundly unfamiliar—and these students will confront a deep, systemic rupture.

1. Competency Benchmark: Who Are They Really Competing Against?

For decades, the Gaokao has been a ranking game among peers—top 5%, top 1% of the same age group.

But in 2035, students face a starker question: Am I competing against my peers—or against AI?

By then, AI will outperform 99% of adults on nearly all standardized cognitive tasks.

The psychological reference point will jump an order of magnitude:

• You spend a day mastering a concept—AI processes it in one second.
• You grasp a theory—AI has already stress-tested it across 10,000 contexts.

For the first time, humans must psychologically accept: We are no longer the highest intelligence in the room.

2. Evaluation Mismatch: AI Banned in Exams, Required in Jobs

A surreal yet real paradox emerges:

• In the exam hall: AI is strictly forbidden. “Independent work” is sacrosanct.
• In the workplace: Not using AI disqualifies you—even from interviews.

Many companies already enforce this today.

This isn’t just a skill gap—it’s a chasm in mindset and methodology:

• Schools test memory, calculation, patience, rule-following.
• Society demands search fluency, creativity, expressive clarity, and collaborative intelligence.

The assessment system and reality are visibly misaligned.

3. Opportunity Architecture: Who Has AI Acceleration, Wins First

Historically, family advantage meant access to tutors, resources, or supportive environments.

By 2035, the gap shifts from resource inequality to system inequality. What truly separates students isn’t IQ—it’s who has AI acceleration.

Examples:

• One student uses AI to generate personalized textbooks in real time; another memorizes static textbooks.
• One builds projects with AI daily; another drills problems and takes notes.
• One iterates daily with the world’s strongest “intelligent tutor”; another relies on limited human teachers.

The efficiency gap isn’t 10×. It’s 100×. For the first time, proficiency with AI becomes the new frontier of educational inequality.

That’s 2035: an old education system still running—while a new world crashes in.

Seventeen million students stand at the starting line—appearing to begin together. But beneath the surface:

• Their mental benchmarks are no longer human.
• Their evaluation criteria are already out of sync with society.
• Their future opportunity landscape has already been redrawn—by who uses AI, and how well.

1. The more skilled AI becomes at giving answers, the more essential it is for you to master asking questions.
2. For most people, AI is an amplifier of ability—but for a few, it’s first and foremost an amplifier of ambition. Most use it to save time; a few use it to host bolder imaginations—like launching projects they’d never dared dream of ten years ago. AI doesn’t amplify skill first—it amplifies whether you’re willing to treat your own life as a larger experiment.
3. As large models grow more capable, what remains uniquely human—and differentiated—is your values, taste, and insight.
4. You’re not using AI to create work. You’re using your work to train your mirror self. Every output is a self-cloning event.
5. The better you are at writing prompts, the more deeply you understand human frailty: clarity, constraint, compression, prioritization, counterexamples, boundaries…
6. Future founders won’t need answers—they’ll need to design questions scalable by AI.
7. Information is no longer scarce. What is scarce is your personal compression function. AI can expand infinitely; you must compress. In the AI era, compression ability is the new IQ.
8. AI may be the first tool in history that cares more about who you actually are than what you can do.
9. Learning fast is still slower than one structural cognitive reframe. Learning adds incrementally; reframing leaps. When was the last time you changed your mental framework—not just your technique?
10. You’re not teaching AI to become smarter. You’re teaching it to replicate your soul. What AI enhances isn’t just your strengths—it’s your blind spots, too.
11. The more you rely on AI, the more vital it is to cultivate your irreplaceable human parameters: values, taste, judgment, worldview, aesthetics. Right now—what’s the most human parameter worth having a model learn from you?

High Standards

High standards aren’t just a work habit—they’re your operating system.

One of the deepest impressions I carried from working with a former leader was his foundational professional principle: If we decide to do something, we do it as well as it can be done—or we don’t do it at all.

At first, it felt overwhelming.

But after being pushed through discomfort and struggle, I realized: high standards are among the strongest forces for expanding your capability boundary.

They let you grow faster in the same amount of time.
See deeper layers in the same task.
Rise more naturally in the same competition.

When everyone defaults to “good enough,”
we default to “best” as the minimum acceptable version.

Even a 1% edge, compounded over a year, becomes undeniable.

You’ll notice: your output, quality, rhythm, and efficiency quietly shift into another dimension—distinct from your peers.
All because of one simple habit: no superficiality, no shortcuts, no compromises.

What people fear most isn’t hard work—it’s mistaking cleverness for competence, and laziness for efficiency.
Those little “lucky breaks” and “time-savers”? They always come back.
They leave you hollow when complexity arrives, hesitant when difficulty looms—because deep down, you know your foundation isn’t strong enough to hold bigger opportunities.

True mastery isn’t about working harder. It’s about making “the best I’m capable of right now” your automatic setting.
Once that sinks in, the world changes: you spot essence faster, build judgment structures more easily, and distinguish effort that yields results from effort that merely burns time.

More importantly, high standards radiate outward—from person to team, from individual to environment.
You lift the water level around you.
That environmental compounding effect is incalculable.

So this mindset bears repeating: Either don’t start—or start with the intent to deliver the best possible version.
Effort poured into excellence never vanishes.
Effort spent on compromise always leaves a trace.

Your real growth speed depends on the standard you apply—to yourself—in each moment.
And high standards? They’re always the hardest—and most worthwhile—kind of difficulty.

Principles and Rationality

Many family conflicts aren’t caused by wrong methods—but by unclear principles.

Set principles clearly, align on them, and uphold them—even in moments of tension—and most friction dissolves naturally.

For example, in our home, we hold three small but firm principles:

• Never speak negatively about the other parent in front of the children.
• When one parent is guiding or correcting a child, the other does not undermine or contradict.
• Use “we” instead of “I” as often as possible—so children experience their parents as a unified whole.

Living these out consistently demands real rational discipline.

Fatigue, emotion, resentment, miscommunication—all can make us forget our principles in an instant. A child’s offhand comment. A partner’s tired sigh. And just like that, reason drops away, replaced by reflex.

That’s precisely why principles matter most in those moments.
Principles are the line—and the habit—that decides for you when emotion surges.
Emotions shift. Principles hold.

Children need that reliability.
They develop security, trust, and sound judgment not in chaos—but in environments where rules are stable, predictable, and upheld.

When parental responses are random or mood-dependent—“Yes today, no tomorrow”; “Negotiable now, non-negotiable later”; “United this morning, divided by lunch”—children stop attending to logic. They start reading atmosphere instead.

That’s not the outcome we want.

Reasoned dialogue is the hardest—and most vital—skill a family can practice.

It’s not coldness. It’s a gentler way of caring—for each other, and for the child.

• It lets us handle facts before feelings in conflict.
• It teaches children the world is navigated by logic—not guesswork.
• It keeps the family system intact, even when emotions run high.

The more we practice reason, the clearer the core truths become:
Real love isn’t choosing the path that feels easiest right now.
It’s choosing the path that serves the family’s long-term health—even when it’s harder.

Over time, families stabilize. Children settle. Adults mature.
Principles, then, are quiet, tender power—letting us live fully human lives, even amid feeling.

Wealth-Invisible Living

On “keeping wealth invisible”:

1. “Wealth-invisible” ≠ being low-key.
2. “Wealth-invisible” = preventing wealth from distorting others’ mental model of who you are.
3. Human perception is multi-variable. When “wealth” suddenly dominates that model, other variables—competence, contribution, character—get compressed. That’s single-variable misjudgment.
4. True wealth-invisibility means letting others see your value structure first, your wealth structure second.
5. Put differently: When you succeed, the more you make “money” seem unimportant in your story, the more people will see you—not just your net worth.
6. The wealthier you become, the more actively you must craft narratives unrelated to wealth.
7. Wealth itself isn’t the problem. The absence of anti-wealth narratives is.
8. Why doesn’t Elon Musk—despite his vast fortune—feel like he’s “showing off”? Because he’s spent decades building a dominant alternate narrative: adventurer, engineer, relentless disruptor, someone who literally risked his life building rockets.
9. In that narrative, wealth carries almost zero weight.
10. Wealth-invisibility isn’t hiding money—it’s ensuring “wealth” is never the first variable people use to assess you.
11. A basic feature of human cognition: we reduce complex individuals to simple labels.
12. Real masters operate via emotional calibration, scene selection, primary narrative design—and deliberate variable-weight shifting.

Memento Mori

I asked Gemini 3 to build a small tool: Memento Mori — Life Weeks Visualized. It works beautifully.

Link: laoyao.cn

Enter your birth date and expected lifespan—and the tool renders your life in three visual states: Past, Now, and Future.

As shown here:

Stare at that page long enough—and a deeper question emerges:
Am I living my own story? Or someone else’s expectation? Or a script handed to me?

Memento Mori: Remember, you will die.

Your life isn’t an infinite canvas. It’s an oil painting drying in real time.

The power of this phrase lies not in death—but in the sudden, visceral awareness that your time is finite, bounded, and irrevocable. That’s mortality awareness.

Its purpose isn’t fear. It’s clarity. It invites you to ask:

• Which parts of this life truly belong to me?
• Which are just noise from society?
• Which paths am I walking only because I’ve never questioned them?
• Which choices did I accept without ever choosing?

Mortality awareness isn’t pessimism—it’s radical lucidity. It gives you permission to:
Reject meaningless busyness.
Release stories that aren’t yours.
Reclaim time from others.
Return your life—fully—to yourself.

In a finite life, your greatest freedom is deciding where each brushstroke lands.

AI & The Future, Part II

The Silicon World, our live discussion series on “AI and the Future,” held its second session Saturday night—covering eight tightly framed topics:

1. When was the last time an AI-related moment shifted your worldview?
2. Why do some still reach for traditional tools before AI—even when stuck? Is the gap habit, inertia—or something deeper? When will “everyone uses AI daily” arrive? Is it generational? How far off is “AI-as-smartphone”?
3. What major, previously unmet needs or large-scale scenarios have recently become viable because of AI?
4. If AI-generated art or writing surpasses humans in both quality and speed by orders of magnitude—what’s left for “average human creators”?
5. Which futures described in The Three-Body Problem are inevitable—and which are fiction?
6. At 10-, 20-, 50-, and 100-year horizons: what will AI be capable of—and what will it change?
7. If an ordinary person had just 10 hours to learn AI seriously—how would you allocate those hours across three concrete actions?
8. Thirty years from now, what one sentence do you hope 20–30-year-olds say about our generation?

Gemini 3 distilled the full transcript into The Silicon World Core Ten—a concise synthesis of ideas that reshape cognition, action, business judgment, and ultimate values.

I. Cognitive Reframe (Breaking the Old World)

1. “Experience” Is Becoming a Liability

Core insight (Fuxiang / Chen Xiang): “The era of eating experience is over.”
This is the harshest workplace truth. When an intern + API can replace a decade-old security team, the moat of accumulated “proficiency” collapses. The future belongs to explorers—not veterans.

2. Art and Technique Are Now Fully Decoupled

First-principle insight (Niksi): “Before, making art required mastering technique (e.g., piano). AI decouples them. Now, aesthetic sense alone unlocks creation.”
That’s why AI content explodes: “creator” now means taste and imagination—nothing else.

3. Entertainment Is the Real On-Ramp to Mass AI Adoption

Counterintuitive insight (Niksi): “The day ‘everyone uses AI daily’ arrives—not because people use it for work—but because fun, playful AI products go viral.”
History repeats: internet and smartphones spread first through games and social apps. So will AI. Don’t fixate on productivity tools—you’ll miss the biggest entry point.

II. Action Guide (Surviving the New World)

4. The “50-Keyword” Breakthrough Method

Most effective learning method (Yao Jingang): “Pick a new domain. Identify its core 50–100 keywords. Ask AI to explain each—then teach it back using the Feynman technique.”
In an age of knowledge overload, this is the fastest, lowest-cost way for anyone to build a functional mental model from zero.

5. Don’t Treat AI as a Tool—Treat It as a Mentor

Key mindset (Yuanzi): “Don’t just search. Set it as a respected teacher—and prepare thoughtful questions before asking.”
Prompting isn’t incantation. It’s posture. The richness of your context—the respect you show—directly determines the density of its intelligence.

6. Business Opportunity: Releasing Suppressed Demand

Commercial logic (Chen Xiang): “Human needs haven’t changed. AI’s opportunity lies in lowering the cost of fulfilling previously suppressed needs—like 1:1 tutoring or therapy.”
This is the litmus test for spotting real vs. fake demand: don’t invent new wants—remove old barriers.

III. Human Grounding (Holding the Line)

7. The Essence of Creation: “Because I Exist”

Emotional anchor (Yao Jingang): “Even if AI paints better, my daughter’s drawing still matters—because this world exists only because I exist in it.”
When AI outperforms humans in every skill, this subjective, embodied presence remains our sole, unassailable dignity—and our final moat.

IV. Ultimate Horizon (Looking Farther Ahead)

8. The Three-Body Problem—Real-World Mapping

Masterful analogy (Fuxiang): “Luo Ji (the Swordholder) = Ilya Sutskever (guardian of safety); Wei Cheng (ruthless pragmatist) = Sam Altman (relentless accelerator); Cheng Xin = European regulators; the Dark Forest = Anthropic.”
Not just clever wordplay—it’s the sharpest, most accurate portrait of today’s global AI power dynamics. Understand this map, and you understand Silicon Valley.

9. The Endgame Roadmap: AGI → Robot → Rocket

Long-view forecast (Fuxiang): “10 years: AGI (digital world). 20 years: robots (physical world). 50 years: interstellar travel (cosmic world).”
This is civilization’s main storyline—a grand coordinate system to orient every step we take today.

10. “The Beginning of Prosperity”

Ultimate framing (Fuxiang): “In 30 years, people will say: ‘That generation—30 years ago—was the beginning of sustainable human prosperity.’”
Amid widespread anxiety, this fiercely optimistic definition is a shot of adrenaline. It reminds us: We stand at history’s hinge. We are the ancients of a new age—and the architects of abundance.

A person with integrity often holds an internal compass: some things can be done; others must not.

Even when no one is watching, they do not cross that line.

The driving force behind this judgment is awe.

But without awe, boundaries blur—and the sense of “enough” or “too far” vanishes.

Every domain has its floor and its edge. Without awe, people lose sensitivity to those limits. They tend to:

1. Underestimate consequences, treating boundary violations as trivial—unaware of systemic costs;
2. Overestimate relational tolerance, assuming everyone must understand and accommodate them;
3. Dismiss rules, seeing them as optional rather than scaffolding;
4. Misread themselves, lacking self-awareness and leaning into blind confidence.

Together, these traits erode both moral grounding and practical discernment.

Why does awe go missing?

At root, it’s because someone cannot see—or refuses to believe—in anything larger than themselves. The logic runs: ignorance → fearlessness → awelessness.

Awe emerges only when we grasp our finitude—when we feel the scale of the world, recognize real boundaries, and accept that consequences are neither abstract nor negotiable.

Rebuilding awe as an adult is hard—but possible. The core path is exposure to meaningful feedback systems: real-world consequences, mentorship from people who operate at higher resolution, sustained learning that surfaces uncomfortable truths.

Awe isn’t taught in lectures. It’s taught by reality.

People without awe easily spiral; those with awe stay steady, precise, and free—not because they’re untested, but because they choose not to test what breaks them.

Leveraging Leverage

Earning well isn’t about working harder—it’s about leveraging wisely.

Mastering just one high-leverage mechanism—management, software, media, or capital—is enough to create exponential divergence in life outcomes.

What is leverage?

A system that turns small inputs into disproportionately large outputs. It has three defining traits:

• Amplification (of effort, insight, or impact—both ways);
• Replicability (it works again and again, without re-engineering);
• Declining marginal cost (the 100th use costs far less than the first).

It’s time with compound interest.

Without leverage, you trade hours for dollars—linear time. With leverage, your time multiplies, copies, spreads, and compounds.

But leverage isn’t magic. It’s driven by domain-specific mastery.

Software leverage requires deep technical intuition—not just coding, but architecture thinking. Capital leverage demands fluency in risk, timing, and valuation—not just spreadsheet skills.

So professional competence isn’t the lever itself. It’s the force that moves it. Competence builds strength; leverage scales the result.

Most people don’t need all four levers. They need their lever—the one that fits their temperament, access, and learning curve. Think of them as four distinct weapons: master one, and you’re armed.

The Power of Good Questions

Answers cost time and energy. Good questions generate value—low-cost, high-yield, fast-acting catalysts for cognition.

How?

1. Catalyzing cognition
   They shift thinking from linear to structural—cutting straight to essence.
   Examples:
   • Which step offers the highest leverage?
   • If you could do only one thing, what would it be?
2. Compressing and focusing
   They strip away noise to expose the singular core.
   Examples:
   • What is the one thing this is really about?
   • Where does intuition fail here—and why?
   • What single variable is actually holding growth back?
3. Catalyzing action
   They don’t just illuminate—they launch.
   Examples:
   • What is the next action—smallest, most concrete, and most necessary?
   • What’s the cheapest experiment to test our key assumption?

A good question doesn’t wait for clarity. It creates it.

Why Run a Marathon?

A marathon compresses a lifetime of human challenges into 42.195 kilometers: endurance, pain, doubt, breakthrough, collapse prevention, pacing, strategy shifts, and bodily feedback—all in under six hours.

Three insights emerge:

1. You control more than you think
   Pace, breathing, fueling, stride cadence, mindset, expectation management—these aren’t fixed. They’re trainable. With repetition, they reshape how you approach any long-term challenge. That’s where runners’ quiet confidence comes from—not bravado, but the lived knowledge that systematic practice changes what’s possible.

2. Pain is inevitable. Collapse is optional.
   Around kilometer 30, your brain negotiates surrender. You learn:
   • Pain isn’t a stop signal—it’s data.
   • How you meet pain is a skill—one you can train and transfer.
     You don’t stop feeling it. You learn to move with it.
3. It’s the cheapest rite of self-reinvention
   Adulthood rarely offers ceremonies that let you redefine yourself. A marathon does. Crossing that finish line delivers visceral proof:
   • You are not who you were yesterday.
   • You still have untapped capacity.
   • You are now someone who keeps commitments—to yourself.

Much of life’s freedom flows from this: the deep, embodied certainty that you can steer your own course. A marathon installs that certainty—not in the mind, but in the body.

Meta-Thinking Chains

A recent paper, “Don’t Just Give Answers—Teach How to Think and Correct Mistakes,” proposes a paradigm shift for AI reasoning: moving beyond Chain-of-Thought (CoT) to Meta-CoT—a framework where models don’t just reason step-by-step, but reason about their reasoning.

Traditional CoT mimics System 1 thinking: fast, sequential, linear. But real problem-solving—especially in math or logic—requires System 2: searching, verifying, backtracking, correcting.

1. What Is Meta-CoT?

Meta-CoT adds a layer of meta-cognition: modeling the hidden exploration before explicit steps appear.

Formally:

Question q → latent exploration z₁→z₂→…→zₖ → explicit reasoning steps s₁…sₙ → answer a

Standard CoT trains only on (s₁…sₙ). Meta-CoT trains on (z₁…zₖ) too—teaching the model how to think before it thinks.

2. Key Mechanisms

• Search & verification: Introduce Monte Carlo Tree Search or A*-style evaluation during inference, rewarding logical coherence over speed.
• Self-bootstrapping: Model generates → verifies → filters → retrains, iteratively strengthening its “thinking protocol.”
• Meta-reinforcement learning: Optimize how to search, verify, and correct—not just what to output.
• End-to-end pipeline:
  • First, instruction-tune on linearized search trajectories;
  • Then, reinforce for generalizable System 2 reasoning.

3. Evidence

• Performance: On hard math problems, Meta-CoT models outperform standard CoT—especially as difficulty rises.
• Behavioral trace: Longer, more branching thought chains appear for harder tasks—evidence of internal search/verify loops.
• Real-world alignment: OpenAI’s o1 series and DeepSeek-R1 already show traces of “internal search”—suggesting Meta-CoT is converging with practice.

The core hypothesis: Generating a correct answer is harder than verifying one. Meta-CoT works by modeling that very loop—search, test, refine—inside the model itself.

In short: Don’t just teach answers. Teach the engine that makes them.

4. Seven Prompt Engineering Tactics (Based on Meta-CoT)

1. Shift from outcome- to process-oriented prompts
   ❌ Traditional: “Solve step-by-step and give the final answer.”
   ✅ Meta-CoT: “Don’t solve yet. First: list 3 possible approaches. Compare their risks and trade-offs. Choose the strongest. Then solve. Finally: if wrong, where would the error most likely live—and how would you fix it?”

2. Add an explicit “Verifier” role
   ✅ *“You play two roles:
   • Thinker: proposes each reasoning step.
   • Verifier: checks whether that step follows logically from prior ones.
     Alternate lines between them until Verifier confirms coherence.”*
3. Inject “search budget” signals
   ✅ *“You may use up to 5 rounds of reflection. In each round, record:
   • Your current hypothesis,
   • How you’d test it,
   • If disproven, where you’ll backtrack and what you’ll try next.”*

4. Simulate error + correction
   ✅ “First, write a plausible *wrong solution—and explain why it’s wrong. Then, rewrite the correct version. Finally, summarize the cognitive pattern you just repaired.”*

5. Enforce layered output structure
   ✅ “Structure every response as:
   [SEARCH] → [VERIFY] → [REVISE] → [ANSWER]”

6. Demand explainable reasoning traces
   ✅ *“For every step, explicitly state:
   • Why you chose it,
   • How you know it’s valid,
   • How you’d detect if it’s wrong.”*
7. Simulate multi-agent reasoning
   ✅ *“Three experts collaborate:
   • A (intuition): gives the first hunch.
   • B (analysis): maps assumptions, constraints, dependencies.
   • C (critique): stress-tests each claim.
     They debate until consensus emerges.”*

These aren’t gimmicks. They’re prompt-level implementations of Meta-CoT’s core insight: Thinking well means building a feedback loop inside your own reasoning.

On Saturday evening, “Future Silicon World” hosted Xiang Shu as a guest speaker, who led a deep dive into six core themes. Xiangyang, Yuanzi, Nick Xi, and I joined the discussion—provocative, grounded, and rich with insight.

Here’s a distilled summary of key takeaways, refined with AI assistance:

1. When Was the Last Time AI Truly Surprised You?

• RM Papers & Memory Experience: Reading recent Retrieval-Augmented Memory (RM) research revealed a leap—not just answering queries, but actively organizing, extending, and retaining context across documents.
• Voice Input Breakthrough: The synergy between voice input and GPT-5–level debugging felt like jumping from “idiot speed” to “awakening speed.”
• The Era of Proactive AI: Three years ago, GPT-3.5 was a passive tool. Today, AI initiates conversation—it reaches out.
• Subagent Teams: CC’s multi-agent orchestration showed how AI teams can now deliver end-to-end outcomes—handling familiar and unfamiliar tasks with layered competence.
• LOVOT’s Emotional Blueprint: Japan’s LOVOT robot offers early proof of emotionally resonant AI—especially in how elderly users’ attachment reshapes our understanding of human-machine warmth.
• Human Growth in Parallel: With AI support, coding fluency and music acquisition aren’t linear—they’re exponential.

2. How Has Your Daily Work Changed Since 2022?

• Mindset Shift: From “difficulty gatekeeping” (“Can I do this?”) to “imagination gating” (“What should I build next?”). Efficiency and cognition now outweigh raw task volume.
• Interaction Mode: Voice-first dialogue feels like brainstorming with a long-term collaborator—not querying a search engine.
• From Tool-User to Tool-Maker: Search volume is down; tools like Raycast and custom agents make creation frictionless.
• Learning Depth: “Shallow consumption” gave way to paper-level research—AI as co-investigator, not just summarizer.
• Story + Practice: AI-designed fitness plans, AI-coauthored books, auto-generated training schedules—all personalized, systemic, iterative.
• Social Rhythm: In an era of productivity explosion, time is the only true scarcity.
• Crucial Reflection: Real “100× productivity” isn’t output velocity—it’s investing time to build better tools.

3. If the Next Decade Were Compressed Into 10 Months—What Would You Do First?

1. Robotics Goes Everyday: Embrace open-source, multimodal robots—bridging AI with the physical world.
2. Co-Creation Over Isolation: Prioritize in-person collisions with peers. Weekends may become the new innovation labs.
3. Slow Down Intentionally: Reclaim fishing, hiking, or quiet mornings—rediscovering that slow is fast.

AI raises the floor for baseline competence. In education and marketing alike, depth—not breadth—is the differentiator.

Structural Shifts Unfolding:
→ Productivity surge → Production relationship redesign → Social reconfiguration
→ The “one-person company” emerges as viable. AI handles what; humans must reclaim why.

4. How Should We Raise Children in the AI Age?

• AI is already an expert—but humans still need to learn how to learn.
• Priority shifts from knowledge acquisition to questioning, judgment, and creation.
• Parenting evolves: Less lecturing, more sparking curiosity. Values > skills.
• Real-world example: Building a toothbrushing competition app with your child—turning self-management into play.
• Core belief: School isn’t for grades—it’s for mastering learning itself.
• Happiness as choice: Ban “100% perfection.” Align skill development with authentic joy.

“Good character, AI literacy, and physical health are the three pillars of future education.”

Inspired by Captain’s dialogue with his 14-year-old son:

1. Cultivate taste: AI generates 1,000 options—choosing the right one is human wisdom.
2. Learn to orchestrate: AI maps workflows—but only humans navigate emotional variables.
3. Withstand failure: AI teaches efficiency; life teaches resilience.

5. What Will Be the “Employment Sponge” of the AI Era? Which Jobs Will Disappear—and Which Will Emerge?

• New sponge roles: Human-AI co-creation (e.g., prompt engineering + editorial judgment), high-touch companionship, empathetic counseling, consultative sales.
• At-risk roles: Mid-tier analysts, copy-paste editors, routine reporting positions.
• Emerging roles: AI Operations Specialist, Agent Architect, Data Validation Engineer.
• Enduring edge: The ability to find—and sustain—work you love.

On the Programmer’s Path Forward:

• Self-directed learning dominates: No internships? Then build as a native AI-generation creator.
• Time horizon expands: With lifespans nearing 100, obsession—not credentials—drives relevance.
• Judgment & taste rise: Amid AI gold and garbage, discernment separates signal from noise.
• Next-level growth: Master systems thinking—extending both individual impact and collaborative reach.

6. What Still Requires Irreplaceable “Human Flavor”?

• Low-pressure, long-half-life creation: Fishing, sketching, playing piano—activities where process matters more than output.
• The Lindy Effect: The longer something has existed, the more likely it will endure (e.g., handwritten letters, analog photography).
• Running & marathons: Slow, embodied challenges that forge identity through sustained effort.
• Human presence: In an ocean of flawless AI content, human imperfection becomes a trust signal.
• Handmade revival: As AI floods the market, personal IP and distinctive taste grow rarer—and more valuable.

How AI Gradually Replaces Humans

Consider a typical corporate org chart:

At the base: many junior staff. As you ascend: fewer people—but higher capability.

Yet AI’s rapid advance is hollowing out this pyramid—not as a tool, but as labor replacement:

1. Unlike past technologies, AI isn’t just “augmenting”—it’s substituting.
2. As performance climbs, middle management and strategic tasks get automated—systematically “excavating” the pyramid.
3. Governments and industries are doubling down—accelerating disruption.

The result? A “pyramid replacement” in white-collar firms:

1. Freeze entry-level hiring
2. Eliminate junior roles entirely; reduce mid-tier demand
3. Trigger layoffs—first juniors, then managers
4. AI assumes most functional roles
5. Only C-suite remains—to set direction for vast AI teams
6. In extreme cases: fully autonomous “zero-human companies”

As visualized here:

Traditional career ladders and occupational categories will collapse—demanding social and institutional reinvention.

Key implications:

1. Competitive advantage shifts from “headcount” to depth of AI integration and collaboration architecture.
2. New infrastructure opportunities emerge: AI orchestration platforms, “AI workforce” monitoring systems, and cultural frameworks for human-AI coexistence.
3. For individuals: Become an AI conductor, not a replaceable player. Ask: Can you orchestrate AI toward complex goals? Do you wield advanced prompting, system integration, and judgment? Do you embody irreplaceable human qualities—taste, trust, meaning?

Source: intelligence-curse.ai

Five Techniques for Pragmatic Learning

Learning falls into two modes: pragmatic (goal-driven) and intrinsic (for its own sake). Both matter—here are five tactics for the former:

1. Clarify “Why” First
   Motivation anchors retention. Ask yourself:
   • Why does this deserve my time?
   • How does it serve my long-term vision?
   • What concrete outcome becomes possible once mastered?

2. Anchor Learning in “Use”
   Pragmatic learning isn’t about knowing—it’s about doing. Start with a real project, role, product, or business problem—then learn just enough, just in time.
   Guiding principle: Don’t learn knowledge that solves no problem.

3. Learn from the Top Tier
   Your learning ceiling is set by your reference points. Study how elite practitioners think, decide, and validate. That exposure recalibrates your standards—and sharpens taste and judgment.
   Same for reading: Skip 10 average books. Master one canonical text—and implement it relentlessly.

4. Learn by Doing
   Theory is the spark; action is the fire. Build, ship, iterate. Let feedback—not lectures—drive refinement.
   Goal: Close the loop—cognition → action → feedback → iteration.

5. Go Deep, Fast, and Focused
   Block distractions. Pick one theme. Structure learning as: framework → details → application → reflection.
   This builds not just knowledge—but your own personal knowledge architecture, essential for expertise.

AI-Powered Feynman Learning

Doubao’s voice interface is exceptional.

Use it to test your grasp of any new concept—out loud, in real time. Just create a custom agent in Doubao with a prompt like:

“You’re my learning partner. When I explain a concept to you, ask me clarifying questions, point out gaps, and challenge assumptions—like a thoughtful peer reviewing my understanding.”

Then speak freely—about quantum computing, supply chain math, or poetry meter. Instant, low-friction verification. Sometimes, convenience is the catalyst.

Whose Advice Should You Actually Take?

To avoid draining energy on unactionable input, prioritize advice from only three sources:

1. Domain experts or authorities—listen deeply, even if you disagree.
2. Your customers—internal or external—but always assess whether their request aligns with your mission and constraints.
3. Specific, factual corrections—e.g., “Line 42 has the wrong metric.” These are actionable and objective.

In all cases, your judgment remains the final filter.

Today, advice is abundant—but discernment is scarce. AI makes it trivial to generate exhaustive recommendations for any question. The hard part? Knowing which apply to you—and having the clarity and courage to act.

So don’t rush to advise others. First, listen. Sense. Wait. Often, what’s needed isn’t your answer—but your presence.

Turn Inward

After assigning a task to the team, a small execution detail went off-track. I asked our operations colleague where the misalignment occurred. She reviewed chat logs—and found the error originated with an engineer’s implementation.

Instead of attributing blame, she paused. Reflected. And proposed a new cross-functional review step before launch.

That moment revealed a rare quality: not deflecting, but turning inward—then building systems to prevent recurrence.

I once asked myself: What does “good reflection” actually look like?
My answer crystallized into four words: “Turn inward” (fǎn qiú zhū jǐ).

Its power lies in focusing only on what you control—and designing interventions accordingly.

Two contexts matter:
• Growth mode: When aiming for real progress, turning inward yields compound returns.
• Preservation mode: In non-professional stress, occasionally externalizing blame is healthy—freeing mental space.

But the most resilient habit? Train attention on high-compounding activities: sleep, movement, presence, friendship, family, and the quiet confidence that comes from showing up—fully—for your own life.

“AI First” sounds great—but without professional judgment, it quickly becomes AI blind faith.

I used to believe that in the AI era, the “AI First” principle would benefit most people—and that adopting it as a habit or mindset would reliably improve outcomes.

Now I see that view is flawed. The reality is far more nuanced.

First, most people don’t truly understand AI’s boundaries—especially in specialized professional contexts. It’s easy to confuse what AI knows with what it doesn’t know.

Second, most people struggle to accurately assess their own boundaries. In many cases, they suffer from the “unknown unknowns” problem: they don’t know what they don’t know. So even when AI outputs garbage, they remain unaware.

See the diagram below:

What’s truly alarming is how, over time, habitual AI assistance can erode our judgment. We start cautiously reviewing AI outputs—then gradually shift to passive acceptance—and finally, to outright blind faith. That erosion isn’t just of professional judgment; it can extend to basic common sense. And that shift is deeply dangerous.

This isn’t limited to technical domains. Even on everyday factual questions, people often fail to spot AI hallucinations.

Take this news story: Two tourists planned to walk across a tidal causeway to a small island. Before leaving, they asked ChatGPT: “When is low tide today?”

ChatGPT confidently replied: “9:30 a.m.”

They set off—without double-checking with the hotel staff, the local tide chart, or any authoritative source. When they reached the island and tried to return, the tide surged violently. The causeway was submerged, and a dangerous tidal drop formed rapidly.

In a later interview, one tourist said: “I made a mistake—using ChatGPT to check low tide time. It told me 9:30 a.m., so we went. But coming back, everything was completely different. That was a hard lesson for me.”

Just days earlier, I asked ChatGPT about the relationship between resting heart rate and average sleep heart rate. It answered with unwavering confidence—yet its explanation defied basic physiology. I cross-checked across multiple trusted medical sources and found its answer wildly inaccurate.

Such errors are routine—whether in daily life or high-stakes work. At that point, your critical thinking and domain judgment become the only real safeguards. For important decisions, lacking those skills means falling straight into “AI blind faith”—a genuinely perilous state.

And cultivating genuine critical thinking and professional judgment? It’s far harder than most assume.

So now, I advise this: For a small number of high-leverage, high-consequence tasks—don’t lean on AI at all. Keep it at arm’s length. Prioritize building real expertise first: consult authoritative sources, learn from seasoned practitioners, and ground yourself in lived experience. Only then, once you’ve developed true domain mastery, should you bring AI in—not as a crutch, but as a force multiplier.

Otherwise, what I see in practice is grim: teams that think they’re accelerating, but are actually scaling garbage output.

Focus on Building Your Own Advantage

What is an advantage?

My current definition: a capability or trait that consistently generates leverage—amplifying effort into outsized impact.

Traditional education trains us to obsess over weaknesses and gaps—rooted in the “test-driven remediation” mindset of exam culture. But business competition and personal growth operate on entirely different logic.

Fixating on flaws rarely builds competitive edge. Worse, it fuels self-sabotage and internal friction.

Positive psychology offers a better path: discover your strengths—and deliberately amplify them.

In the AI era, someone with a clear, well-developed advantage—paired with AI as a super-lever—can rapidly build formidable personal differentiation and massive productivity gains.

By contrast, pouring energy into fixing weaknesses (yours or others’) carries serious risks:

1. It increases internal friction and resentment;
2. It consumes scarce attention—attention you must manage intentionally;
3. Obsessing over problems doesn’t solve them. Most real-world challenges are resolved not by patching deficits—but by deploying strengths.

So if your colleague, friend, or even family member only points out your flaws or shortcomings—that’s not generosity. It’s often just low-value noise.

What’s truly rare—and invaluable—is someone who sees your strengths, names them clearly, and helps create conditions where you can deploy them fully. In the AI era, that kind of strength-based empowerment is rarer—and more powerful—than ever.

How to Understand OKR

I sketched this diagram for my team to clarify how OKR relates to day-to-day work planning.

What is OKR, really?

At its core, it’s a goal-management system built on causal logic—a pyramid of aligned intent.

As shown below:

Each layer flows from the one above:
North Star Metric → Objective (O) → Key Results (KR) → Key Projects → Key Tasks

Every objective is derived from the North Star—the single metric that best reflects long-term health and direction. Every KR must be a measurable, outcome-oriented indicator of progress toward that objective. Everything downstream serves that causal chain.

To grasp OKR’s role in an organization at a glance, here’s an even simpler visual:

What Makes a Good Father

A friend from Shanghai visited Beijing recently—we had a small dinner with a few others in the AI community.

What surprised me was how often she spoke about her father—always with warmth, respect, and deep gratitude. She described how he helped her build her own value system, cultivate self-confidence, and develop independent judgment.

Listening closely, I noticed key patterns in his approach:

1. On making better choices: He’d playfully challenge her preferences or strengths in a game—and only after winning would he explain: “That choice may seem right now—but a truly good choice opens up more options for your future.”
2. On learning: When she faced a new subject, he spent three days modeling how to learn it from scratch—demonstrating that learning itself is a learnable skill.
3. On safety and trust: From early childhood, he patiently named specific people and situations she could safely rely on—guiding her, step by step, to recognize and internalize security.

He also shared thoughtful ways he nurtured her confidence and judgment—each grounded in patience and presence.

Her strong educational foundation wasn’t accidental. It reflected her parents’ deep intellectual engagement—they’re both university professors.

Of course, every adult redefines “good father” differently. Parents love and support their children in diverse, imperfect ways.

But one universal litmus test holds: When grown children talk about their fathers—in casual conversation, under stress, or in moments of reflection—do their eyes light up with affection, respect, or even quiet awe? Do they recall vivid, meaningful moments—not just events, but interactions that shaped who they became?

A good father does more than provide materially. His greatest impact lies in the inner architecture he helps build: values, resilience, discernment, and the quiet confidence to navigate uncertainty. That kind of influence echoes across a lifetime.

The Amateur Ensemble

The world is, fundamentally, a giant amateur ensemble.

So—barring miracles—we’re part of it too.

Accepting that fact isn’t defeatism. It’s clarity. It grounds self-assessment in honesty—and cultivates humility.

Accepting the “amateur ensemble” doesn’t mean giving up on growth. It means choosing where to go deeper: doing one thing a little better, a little more carefully, a little more consistently than the next ensemble.

That modest edge—sustained—is often enough to earn disproportionate returns.

And paradoxically, operating from that grounded, unpretentious stance may accelerate innovation: fewer status games, less posturing, less energy wasted on pretending—and more room for real experimentation.

AI Live Trading Competition

An unusually revealing experiment is underway: six top-tier AI models are competing in live cryptocurrency trading—each starting with $10,000 of real capital.

The platform: nof1.ai

As of November 2, results stand:

• DeepSeek V3.1: +$4,728 (1st)
• Qwen: +$3,603 (2nd)
• GPT-5 & Gemini 2.5 PRO: worst performers

What makes this site compelling is its setting: a high-stakes, highly uncertain, real-time financial environment—where decisions have immediate, measurable consequences.

It pushes AI beyond language fluency—into decision intelligence and execution intelligence.

We shouldn’t treat the leaderboard as a definitive ranking of “best model.” But it does offer something rare: a transparent, outcome-focused window into how these systems perform—not in curated demos—but in messy, consequential reality.

I ran my first full marathon—42.195 km—in Chengdu.
Time: 3 hours, 58 minutes—just under four hours.

For the first 35 km, I held a steady pace of ~5:35/km. At 38 km, sensing early signs of cramping, I eased off—and crossed the finish line safely.

The course was flat, the weather overcast—ideal conditions. The physical strain began to mount after 35 km; that’s when willpower stepped in.

All in all, it was a deeply rewarding experience.

On the subway back to the hotel, I met a friendly retired man from Zhejiang—also fresh off his own marathon. We shared the same stop and chatted the whole ride.

Ten years ago, he ran sub-3:30—a time that, for his age group then, carried the same weight as breaking three hours does today. Now, he no longer races for time. For him—and for most amateur runners—health is the core goal. Pushing one’s personal limits, within that frame, remains meaningful. I couldn’t agree more.

No matter how hard amateurs train or race, we’ll never match elite athletes—let alone state-supported professionals.

A marathon feels like extended meditation: rhythm, breath, and focus fused into motion. For ordinary people, its real value lies not in speed or ranking—but in repeated exposure to our own edges. It’s how we meet the version of ourselves who refuses to quit.

On Values

Here’s how I think about building a values system—for teams or companies:

1. Let values emerge intuitively first, then verify them. That’s more authentic than over-engineering them—or copying “world-class” companies. What surfaces naturally reflects your lived reality.
2. A strong values system must be practiced, not just proclaimed. If the founder or leadership can’t live up to it at least 90% of the time, it becomes theater—not culture.
3. Intuitive emergence works because values are habits made visible: they’re patterns you already follow, principles you lean on instinctively. They grow from the inside out, not top-down.
4. Values aren’t slogans. They need clear behavioral definitions—and concrete examples of what “living them” looks like.
5. Personal values → company values → team values. The chain starts with the individual.
6. Saying something and doing it are two different things. Integrity means aligning inner conviction with outward action—consistently.
7. Many organizations have no written values—but their leaders model integrity, humility, and care daily. That is values in action—and it spreads silently, powerfully.
8. Culture isn’t dictated; it’s contagious. Leadership behavior ripples outward, shaping norms without a single memo.
9. So the real sequence is: inner orientation → consistent behavior → named values → influence on others.
10. True values aren’t “agreed-upon statements.” They’re repeated, verified choices—habits hardened by time and testing.

The Core Logic of Getting Things Done

Seven principles for doing work well—and doing the right work:

1. Clarity of purpose: Define the problem first. Quantify the goal. Without this, action is noise.
2. Embrace experimentation: Learn through doing. Test assumptions. Cut what doesn’t hold up.
3. Focus on leverage: Resources—time, energy, attention—are always finite. Invest where impact multiplies.
4. Routinize reflection: Build feedback loops. Turn every outcome—success or failure—into a lesson. Let systems evolve.
5. Dissolve artificial boundaries: Don’t box yourself in. Stay open to higher-leverage paths—even if they’re unfamiliar.
6. Prioritize quality over validation: Positive feedback feels good—but relying on it breeds fragility. Train discipline, perspective, and collaboration even in silent stretches.
7. Execute relentlessly—but only once direction is sound. Use time and effort to let truth reveal itself.

Why Most People Struggle to Use AI Productively

By “most people,” I mean non-AI practitioners—teachers, managers, creatives, small-business owners.

I recently gave an AI workshop for faculty at a friend’s school. Covered fundamentals, tool recommendations (global and domestic), and practical use cases.

Three observations emerged:

1. The gap between knowing and doing is wide
   I asked teachers: Now that you understand AI’s capabilities, what are 10 specific things it could do for your work?
   Three volunteers shared their lists. Almost none included concrete, role-relevant applications—like drafting parent emails, generating rubrics, or summarizing meeting notes.
   Why? Because understanding AI at a conceptual level ≠ knowing what it can concretely do for you. Awareness ≠ utility.

2. Tool literacy remains low—even for familiar tools
   Many teachers use Doubao (ByteDance’s AI assistant) daily. Yet none knew about—or had tried—its real-time AI video call feature.
   I demoed it live. The reaction? “Wow—I had no idea this existed.”
   Most AI tools today are intuitive—but users face two barriers:
   • They don’t know the features exist.
   • Even when they do, they hesitate to click (“click anxiety”).
     So I now teach a simple rule: When you install a new tool, click every visible button. Try everything. Nothing breaks. Literally—“learn to click.”
3. AI isn’t yet tied to real productivity
   People adopt tools faster when they see tangible outcomes. One compelling case study—e.g., “This teacher cut lesson-planning time by 40% using AI”—shifts perception faster than any theory.

Stoic Life Principles

Stoicism is a practical philosophy rooted in reason, self-mastery, and inner calm—originating with Zeno of Citium in ancient Greece.

1. Control: Focus only on what’s yours to control

• Divide life into two categories: what you can control (judgments, choices, actions) and what you cannot (others’ opinions, outcomes, weather, reputation).
• Direct all energy toward the former. Release attachment to the latter.
• When facing difficulty, ask: Is this within my control?

2. Reason: Virtue—not externals—is the sole good

• Act from rational judgment—not impulse, fear, or desire.
• Moral character is the only true source of flourishing.
• Wealth, status, health, and fame are “preferred indifferents”: useful, but never necessary for happiness.

3. Rhythm: Freedom lives in restraint

• Live simply. Temper desire. Resist hedonic excess.
• Never trade long-term integrity for short-term gratification.
• Discipline isn’t restriction—it’s the architecture of freedom.

4. Fate: Love your fate (amor fati)

• Accept all events as part of nature’s rational order.
• Resistance wastes energy. Acceptance opens space for growth.
• Every challenge is raw material for virtue.

5. Mortality: Remember you will die

• Awareness of finitude sharpens presence and focus.
• Death isn’t a threat—it’s a reminder to act now, with clarity and courage.
• True freedom begins when you stop fearing the end.

6. Community: Reason points outward

• As social beings, our rationality calls us to justice—even when unreciprocated.
• Assume ignorance, not malice, behind others’ actions.
• Do good not for praise—but because it’s right.

7. Self-examination: An unexamined life is not worth living

• Ask daily: How will I act virtuously today? Did I focus on what I control? Was I led by desire or reason?
• Keep a “reason journal”: log emotions, decisions, and reflections—not to judge, but to learn.

At the end of last year, I distilled some reflections into simple formulas.

On happiness, my earlier understanding was:
Happiness = MIN(Meaning, Health, Relationships) × Presence

But recently, I’ve arrived at a new perspective—one grounded in the architecture of personal competence, illustrated here:

How do we interpret it?

The happiest state occurs when:
Cognition > Capability > Wealth > Desire

In other words, when each higher-level element can “govern” or “channel” the one below it, the system becomes stable—and generative.

Conversely, when cognition, capability, or wealth fail to contain desire, desire takes over. That’s when people suffer: anxiety, confusion, even depression intensify.

As shown in the diagram below:

This also explains why money earned through luck is often lost just as quickly. When wealth vastly outpaces one’s ability to regulate desire, that wealth tends to vanish—not from bad decisions alone, but from desire’s unchecked momentum.

So the sound principle for growth is twofold:
→ Reduce consumerism (to manage desire reasonably)
→ Cultivate creatorship (to expand cognition and capability)

What lies beneath this is a fundamental energy balance:
→ Tightening the output end (desire)
→ Expanding the input end (cognition)

The same applies to children: encourage creation over consumption; emphasize process over outcome.
Because cognition accumulates through making, not through acquiring. When we obsess over results—especially in a consumption-oriented way—we inadvertently cap our cognitive and capability growth.

A Sense of Control

Happiness doesn’t stem only from “MIN(Meaning, Health, Relationships) × Presence” or from a balanced competence structure.
Another vital dimension is a sense of control.

Happiness correlates strongly with perceived control—the belief that you can meaningfully influence important events and outcomes in your life. This feeling significantly boosts well-being.

When people feel in charge of their lives, they experience less stress, greater self-confidence, and higher life satisfaction.

Crucially, happiness here depends not on how “good” external conditions are—but on whether you feel like the author of your own story.

There are two kinds of control: outward and inward.

Outward control—shaping environments or influencing others—is a double-edged sword.
It may offer fleeting security, but over-pursuing it usually backfires: the world is inherently uncertain; others have their own wills; systems follow their own logic. Excessive control-seeking erodes relationships and traps us in a paradox: the more we try to control, the more we lose control.

Inward control, by contrast, reliably lifts long-term well-being. Examples include:
→ Learning to regulate your emotions
→ Choosing where to place attention—and how to interpret events
→ Building foundational capacities and character traits
→ Cultivating healthy, compounding personal habits

Letting go of excessive outward control—and turning attention inward—often brings unexpected freedom and calm.

The Value of Data Processing

Qichacha has filed for IPO. Its prospectus is publicly available here: link

Key highlights:

1. Exceptionally high margins: ~90% gross margin, >40% net margin—signaling high-value, knowledge-intensive service
2. Revenue CAGR: 16.89%; Net profit CAGR: 35%—driven mainly by rising ARPPU (average revenue per paying user)
3. Over 150 million registered users; >80 million monthly active users; >1 million paying users—and C-end growth continues
4. C-end accounts for >70% of total revenue; within membership products, C-end contributes 88.5%–91.4%

Qichacha’s foundation rests on China’s National Enterprise Credit Information Publicity System, launched by government authorities in 2014.
All enterprise data it displays is, in fact, freely accessible via that official portal—anyone could retrieve the exact same information, given enough time and patience.

Most people simply won’t.

So what Qichacha does is take raw public data—and repackage it with friendlier logic, better presentation, and smarter workflows.
Beyond the National System, it also integrates data from Credit China, China Judgments Online, China Enforcement Information Network, CNIPA, Trademark Office, Copyright Protection Center, Ministry of Civil Affairs, and more.

In essence? It scrapes public data, then adds layers of processing: cleaning, labeling, mining, visualization, and intelligent analysis—before delivering high-value, paid products.

The barrier to entry seems low—which is why competitors abound: AiQicha, TianYanCha, QixinBao, QichaMao, and others.

Yet people consistently overestimate the value of raw data—and underestimate the value of processing it.

Qichacha’s real innovation isn’t data ownership—it’s value-chain reconstruction:
Raw data (free) → Collection → Cleaning → Labeling → Mining → Visualization → Intelligent Analysis → High-value product (paid)

From this view, so-called “data moats” may be shallower—or far less durable—than assumed.
What truly matters isn’t hoarding data, but how it’s applied: how it solves real problems, and delivers tangible value to users.

Practice Is What Matters

“The only true measure of intelligence is whether you can get what you want out of life.” — Naval Ravikant

Clarity about what you genuinely want—and the ability to create that result—is the mark of real wisdom.

True intelligence doesn’t live in books, speeches, or clever-sounding thoughts. It lives in action.

And yet, practice is where many stumble—repeatedly. Here are eight common, costly pitfalls:

1. Perfectionism: “I’m not ready yet.”
2. Analysis paralysis: Endless research, zero execution.
3. Consensus trap: “Everyone agrees—so it must be right.”
4. Idealism: Over-optimistic assumptions about outcomes.
5. Sunk-cost fallacy: “I’ve invested so much—I can’t stop now.”
6. Opposition trap: “Everyone’s against it—so it must be wrong.”
7. Attention trap: Too many ideas, too little focus.
8. Comparison anxiety: Obsessing over others’ pace or achievements—derailing your own rhythm.

These errors don’t operate in isolation. They form a reinforcing psychological network—each feeding the others—until action grinds to a halt.

The key to breaking free? Recognize that there is no perfect starting point. What matters is beginning—then learning, adjusting, and iterating in motion.
A flawed action followed by reflection beats flawless planning that never leaves the page.

The Difficulty of Speaking Clearly

The challenge of saying something clearly may be greater than I’d imagined.

1. Clarity begins in the mind—not the mouth.
   Forming coherent, logical thought is hard enough. Our thinking is often fragmented and associative; translating it into linear, intelligible language demands extra cognitive effort.

2. Speaking is encoding; listening is decoding.
   You compress complex ideas into words; the listener must reconstruct them accurately. The gap between intention and interpretation is wider—and more error-prone—than most assume.

3. Backgrounds differ wildly.
   Knowledge, experience, mood, culture—all shape how someone hears the same sentence. You may think you’ve been crystal clear—only to discover the listener heard something entirely different.

4. Attention is scarce.
   In today’s overloaded information environment, holding focus—and cutting through noise to deliver precise meaning—requires deliberate skill and effort.

Because miscommunication is more common than we realize, a core principle of effective communication is simple but non-negotiable:
→ Stay present and listen deeply
→ Confirm understanding by paraphrasing and asking questions

I restarted running last June. As of today, I’ve been running consistently for 1 year and 4 months.

I just realized: my total distance this year has surpassed 1,000 kilometers.

This is the first time my annual running volume has crossed that threshold—and for me, it’s a major personal breakthrough.

Running is now deeply embedded in my daily rhythm.

The benefits go far beyond physical fitness or mental refreshment. One less-discussed but profound effect is how it reshapes self-awareness: clarifying personal boundaries, refining desire management, and deepening introspection.

Especially when I observe myself from a third-person perspective—detached yet present—that sudden clarity can be genuinely surprising.

During long runs, it’s effortless to slip into self-dialogue and self-observation. There’s no external noise—only breath, stride, and thought.

You become both actor and observer: moving your body while simultaneously watching yourself move. That fluid shift in perspective is uniquely powerful for cultivating self-knowledge.

For me, running delivers more than health—it offers a new lens through which to examine life and understand myself.

Managing Assets

When managing a company, the most critical insight from a risk perspective is simple: running a company means managing risk.

From a growth perspective, however, one of the most important realizations today is equally clear: running a company means managing assets.

What do we mean by “assets” here?

An asset is any resource capable of generating future economic benefit—i.e., something that helps you earn money or save money. Either outcome qualifies.

Its defining traits are: controllability, measurability, and capacity to yield future returns.

So for companies—or individual founders—assets = resources that generate recurring cash flow, compound over time, or create strategic moats.

These include:

• Content assets: articles, videos, scripts, knowledge bases
• Technical assets: codebases, models, APIs, architectures, domain-specific systems
• Brand assets: reputation, influence, trust
• Relationship assets: customer lists, partner networks
• System assets: standardized workflows, automation tools
• Financial assets: stocks, funds, real estate, equity stakes

This framework applies equally to organizations and individuals.

And for individuals, asset-building grows more consequential with age. The good news? It’s never too late to start—because most people simply don’t prioritize personal asset management at all.

Vibe Coding

During the 8-day National Day holiday in October, I stayed in my wife’s hometown—a scenic suburb of Beijing. My main activities: foraging chestnuts, running, writing a book, and building tools.

Most of my time went into tool development.

Thanks to the rapid leap in AI coding capabilities, I—someone with no formal programming background—have seen my coding fluency accelerate dramatically over recent months using Vibe Coding.

In under 10 days this October, I wrote over 3.45 million lines of code, and shipped four functional tools—including one with nontrivial complexity. Just mapping out its product logic and requirements took nearly an entire afternoon.

That made the holiday unusually joyful and rewarding.

Using AI to build tools isn’t just output—it’s also immersive learning. Through hands-on practice, my understanding of Vibe Coding evolved significantly.

The biggest challenge wasn’t writing code anymore. It was defining standards:
How do we judge whether a feature, UI style, or tool is “good”?
How do we translate our intuitive sense of “good” into precise prompts or engineering plans that AI can reliably interpret and execute?

At its core, this circles back to aesthetics, logic—even philosophy. It’s about values, direction, and commercial insight: understanding users, industries, and what truly matters.

On Growth

What does real growth look like?

True growth isn’t just about acquiring new users. It’s about making people want to return.

The world’s top products share one trait: exceptional retention.

So for product growth, retention is the real leverage point.

Retention, in essence, reflects authentic value creation.

A product with poor retention—even if it acquires users brilliantly—will inevitably run dry. Its vitality is fragile.

Two foundational principles drive strong retention:

1. Systematic growth experiments, not isolated tests
2. Systematic reuse of validated insights, not one-off wins

Single-point experiments are unstable—they rely too much on intuition or luck. Systematic experimentation, by contrast, maps all plausible levers, then tests them in parallel to gather rich, actionable data quickly.

Based on this, a complete growth system must integrate four dimensions:

1. Value dimension: The product must continuously deliver value—so users can’t help but come back or keep collaborating. This demands deep user empathy and a defensible core competency.
2. Method dimension: Rely on data-driven experimentation—not gut feeling—to guide decisions.
3. Scale dimension: Turn proven tactics into standardized, modular, repeatable units—and deploy them broadly.
4. Time dimension: Real growth is cyclical: test → learn → adapt → repeat. It never stops.

GEO vs. SEO

1. Whether for SEO or GEO, content quality remains one of the top factors determining whether AI will cite or reference it.
2. GEO is a new form of content marketing—one designed specifically for AI retrieval and generation.
3. Problem-driven, not keyword-driven: SEO centers on keywords; GEO centers on full user questions. This shift mirrors AI’s vastly improved ability to parse complex semantics.
4. Structure saves AI compute: Clear sections, bullet points, data tables, and JSON markup make content easier for AI to parse and quote.
5. AI favors professional, authoritative, verifiable sources—e.g., peer-reviewed research, official standards, government data—with transparent attribution.
6. Information gain is a key GEO differentiator. In an era of low-cost AI generation, unique insights and deep empirical validation are increasingly rare—and valuable. Examples: benchmarked performance data, counterintuitive findings, documented failures, reproducible experiments.
7. Due to AI search’s “zero-click” behavior, GEO impact leans heavily toward brand awareness. Users often stop at the AI’s answer and rarely click through—unless they seek deeper exploration. At that point, they’ll likely search your brand name. So GEO’s primary ROI is being seen, even without high click-through rates: repeated AI mentions shape perception powerfully.
8. SEO remains a foundational layer for GEO—traditional search rankings strongly influence AI’s likelihood of retrieving and citing your content.
9. Getting your brand’s content ingested into AI training data is slow, unpredictable, and largely out of your control. For GEO, real-time search is the main battlefield. Being included in base model training is near-impossible—so the pragmatic path is optimizing for real-time retrieval via SEO.
10. Multimodal content is the future—but for now, text-based content creation and optimization delivers the highest ROI. Build that foundation first, then expand into audio, video, and interactive formats.
11. Domestic AI hallucinations stem partly from cost constraints and limited search service depth. Models often produce shallow, aggregated summaries—which increases error risk. We’ve even run GEO “domination” experiments exploiting this tendency.
12. In a recent GEO experiment focused on AI-generated summaries, we found their influence on AI search ranking is significantly stronger than expected.

On AI Search

This episode of Future Silicon World featured two technical experts from Xiaosu Technology, diving deep into AI search. Here’s a distilled summary—generated by GPT-5 from the live session:

Q1: What is AI search—and how is it different from Google or Baidu?
Traditional search matches keywords—you scan 10 results to find your answer.
AI search is question-first: you ask, and AI synthesizes and answers directly.
Example:

• Search “Who is Apple’s CEO?” on Baidu → list of links
• Ask AI search → “Tim Cook,” plus context on his tenure and recent news
  Key difference: AI doesn’t just retrieve—it understands and generates.

Q2: What does “Infra” mean in AI search?
“Infra” = infrastructure—the nervous system and toolkit behind AI search.
It includes:

• Search engine APIs
• Web page rendering, PDF parsing, HTML extraction
• Data recall, ranking, caching, and semantic understanding modules
  Together, these let AI “read pages,” “grasp context,” and “cite sources” like a human.

Q3: How do developers and regular users interact with AI search differently?

• Developers: Focus on “how to make my AI search.” Join Xiaosu’s dev community, apply for API access, get docs.
• Users: Just want smarter answers—use an AI search app or ask during events.
  In short: developers build wheels; users drive.

Q4: Why distinguish ToC search vs. ToAI search?
They serve entirely different audiences:

• ToC search: Designed for humans → returns concise summaries
• ToAI search: Designed for AI → returns full text or structured data
  Example:
• Baidu gives you “Huawei launches new phone” (headline)
• ToAI search gives AI the full press release—so it can write an analysis
  AI needs machine-readable content, not clickbait.

Q5: What happens inside AI search—from question to answer?

1. Question understanding: Does it need live data?
2. Decomposition: Break complex queries into sub-questions
3. External search: Fetch via API (web, DB, etc.)
4. Recall & ranking: Filter and prioritize relevant content
5. Answer generation: Synthesize and verbalize
   Think of it as an “AI search squad”: model = brain, tools = hands and feet.

Q6: Why does AI need full text—not just summaries?
Because AI needs semantic reasoning, not just facts.
Summaries give conclusions; AI needs context to infer causality, nuance, and provenance—just like a researcher reading original papers.

Q7: What makes Xiaosu’s search tech distinctive?

• Multilingual support: Only domestic vendor offering robust multilingual search
• Fully in-house stack: End-to-end self-developed—from crawlers and indexing to recall and semantic models
• Low latency: Average response in hundreds of milliseconds
• Flexible deployment: Serves both AI chatbots and intelligent agents

Q8: How does AI know if “Apple” means fruit or company?
Via semantic understanding:

• Intent detection
• Tokenization + disambiguation
• Contextual semantic matching
  E.g., “Apple’s market cap?” → automatically routes to financial data on Apple Inc.

Q9: Does AI search rely on outdated info?
No. Systems assess temporal relevance:

• Timeless queries (“Earth’s radius”) → served from cache
• Timely queries (“Today’s A-share index”) → pull latest indexed data
• Temporal signals are even weighted higher in ranking.

Q10: With billions of pages, how does AI pick the right ones?
Hybrid matching: semantic + keyword.
Plus two-stage recall:

1. Offline: Filter out low-quality, spammy, or unsafe sites
2. Online: Dynamic re-ranking per query
   Final candidates: <10% of initial hits—but highest quality.

Q11: What about AI-generated web content?
Not inherently bad—but quality matters.
Systems evaluate logic coherence, factual accuracy, and originality.
Low-value AI “spin-offs” or stitched-together content gets filtered out.
Bottom line: AI writing is fine—if it’s correct.

Q12: Can AI tell good writing from bad?
Yes. Models learn human judgment patterns:

• High information density
• Clear logical structure
• Credible, cited sources
• Originality, no plagiarism, no fluff
  These features are quantified and scored to identify high-value content.

Q13: How do you handle black-hat SEO?
Dedicated anti-abuse team.
Black-hat pages, fake content, scraper sites—all removed offline, before indexing.
Exact algorithms are confidential—but the mission is clear: keep the AI search ecosystem clean and trustworthy.

Q14: Could AI search surface non-compliant content?
No.
Domestic operations strictly follow Chinese laws and values.
Overseas deployments comply with local regulations.
Different regions use region-specific APIs and data stacks—ensuring local legality, global compliance.

Q15: What’s the current business model—and will it change?
Today: API-based pricing.
Future: Potential partnerships with AI apps (ads, knowledge subscriptions)—but no direct ad injection like Baidu.
Why? ToAI search is B2B infrastructure—not a C2C traffic portal.

Q16: Can AI search reveal real-time trend data?
Not yet—but if developers or enterprises request it, Xiaosu can build custom dashboards: e.g., “AI Trend Index,” “Top Questions for Agents.”
Crucially: trends reflect real demand, not vanity metrics.

Q17: Does AI search crawl Xiaohongshu or Douyin?
No unauthorized scraping.
Xiaosu respects platform policies and only partners with willing, high-quality content providers.
As co-founder Guo Gengliang put it: “We’d rather have less data—so long as it’s clean.”
Ongoing negotiations aim to license more structured vertical content.

Q18: How widely is AI search used—and will it replace traditional search?
Current data: ~30% of AI Q&A calls external search.
That share will rise as models better grasp context.
But full replacement of Google/Baidu? Unlikely—for now. AI search is a new layer, not a replacement layer.

Q19: If AI has memory, why still search the web?
Because memory and search serve different purposes:

• Fixed knowledge (e.g., physics constants, historical dates) → stored internally
• Dynamic data (weather, stock prices, policy updates) → fetched live
  They’re complementary.
  The future lies in intelligent balancing between internal knowledge and real-time external inputs.

In a recent talk, Kai-Fu Lee posed a critical question: What will determine competitive advantage for companies in the future? The underlying trend he identified is profound: most work will shift from being organized around people to being organized around AI agents—what we might call “super-employees.” This mirrors the core logic behind our course AI Leadership.

The managerial mindset must evolve accordingly.

In the AI era, resources expand beyond people to include AI tools—but the fundamental leadership capabilities remain unchanged: setting clear goals, deeply understanding the strengths and limitations of your collaborators (human or AI), and designing effective collaboration workflows.

What does change is the rhythm and structure of teams: with AI, pace accelerates, boundaries blur, and teams shrink—becoming flatter and more autonomous.

Key shifts in management perspective and capability:

• Define success with precision: Just as with people, managing AI starts with crystal-clear definitions of success. “Improve retention” is too vague; “increase 30-day revisit rate by 5%” gives AI actionable direction—and better outputs.
• Resources now include AI tools: Your toolkit isn’t just headcount—it’s a portfolio of AI agents, each with distinct capabilities. Whether used proactively or reactively, they’re part of your productive capacity.
• “Human + AI” remains the dominant operating model: Almost all high-value scenarios still require tight, intentional human–AI collaboration—not replacement.
• Flatter, role-blended teams: AI enables individuals to grow multidimensionally. Small teams can own end-to-end outcomes, reducing dependency on traditional roles—e.g., many functions no longer require dedicated product managers.
• De-emphasize job titles; emphasize outcome ownership: The future favors generalists who “get things done”—which demands rapid learning agility from everyone.
• Prioritize data literacy: Use data not just to report, but to diagnose root causes and spot emerging opportunities.
• Accelerate feedback loops: Shift from monthly reviews to weekly—or even daily—iterations. Aim for continuous 1% improvement.
• Use AI to compress scale and deepen accountability: Smaller, end-to-end teams mean faster decisions, tighter ownership, and higher velocity.