The Stronger AI Gets, the Busier People Become

Yuanzi recommended a tool called Cohub, where she archives live-stream transcripts from the “Silicon World.” I fed all those transcripts into an AI to distill counterintuitive insights—and the output was genuinely illuminating:

The stronger AI gets, the busier people become: Summarization tools don’t reduce podcast listening—they make people consume more long-form audio, not less.
AI devalues “supply-side fluency,” not all experience: What depreciates is rote execution—the ability to do things quickly and repeatedly. Demand-side experience—judgment, context, taste, and domain intuition—gains relative value.
As general intelligence becomes cheaper, human differences may widen—not narrow: Access to reasoning power doesn’t flatten inequality; it amplifies the advantage of those who know what to ask, how to frame, and when to stop.
In the AI coding era, code increasingly resembles debt: Real assets are judgment criteria, failure pathways, and collaborative context—not lines of code.
AI won’t refuse absurd requests like a human engineer would: It might generate 10,000 lines for a tiny feature. Success logs only say “it worked”; failure logs preserve intent, preference, correction attempts, and domain reasoning—context far harder to replicate than code.
The more rigorous the finance function, the more suitable it is for AI: Not because models “guess” numbers—but because financial data is highly structured, rule-bound, and verifiable. The Happycapy team mandates Python scripts for every numeric calculation, then uses a second agent to cross-check. In two months of internal use, error rates fell below human levels.
Long-term thinking isn’t always about working harder—it can be strategic laziness: Reducing decision fatigue, automating low-stakes choices, and building systems that compound quietly over time.
The smarter the agent, the less it should have full permissions: Trust must be scoped, auditable, and layered—not granted wholesale.
This AI wave may not favor startups: Model advances could entrench large companies’ advantages—especially in data moats, infrastructure, and vertical integration.
The best AI-era education may look like “wasted time”: Less AI tutoring, more literature, history, philosophy, music, travel, sport, and games—activities dismissed as “low ROI” in exam culture, yet essential for shaping taste, character, curiosity, and moral judgment.
AI’s average artistic output may already surpass the average human’s: Aesthetic judgment may no longer be an unassailable human stronghold.

Several colleagues from legacy media visited our office.

They argued that traditional outlets hold an edge in GEO—particularly around content quality, editorial rigor, and compliance standards.

I respectfully disagree. Their strengths don’t automatically translate to GEO effectiveness—because content optimized for humans and content optimized for AI operate under fundamentally different logics.

Key differences include:

Unit of consumption: Humans read whole articles, following narrative flow, emotional arc, and contextual buildup. AI retrieves and synthesizes at the paragraph, chunk, or evidentiary unit level.
What defines “good content”: Humans respond to headlines, storytelling rhythm, voice, argument, and authorial style. AI prioritizes precision: Does this passage directly define, quantify, compare, list steps, or specify boundaries?
Tolerance for implicit meaning: Humans infer who said it, about whom, and under what conditions from context. AI needs explicit entities, timestamps, subjects, metrics, and relational logic.
Path to persuasion: Human-facing writing builds trust via authority, rhetoric, and cohesive narrative. AI evaluates claims by triangulating across multiple sources—no single voice carries inherent weight.
Reading path: Humans follow the author’s sequence. AI rewrites the query, fetches from dozens of sites, ranks, deduplicates, extracts, and assembles—then cites.

At its core, AI-optimized content participates in a finite-budget evidence competition, passing through six layers: parameter memory → search planning → candidate source pool → ranking & assembly → evidence absorption → citation display.

Legacy media’s advantage lies mostly in candidate eligibility: institutional authority, timeliness, editorial oversight, and third-party neutrality.

But “quality” used to be judged by editors and readers. In GEO, it’s also filtered through search engine chunking, ranking algorithms, and model compression.

Truly dual-purpose content preserves human judgment, lived experience, and expressive voice—while structuring facts as stable, machine-actionable evidence.

A Real-World GEO Case: The Corporate Website as Primary Source

A friend’s company launched a new channel on its official website. Within one month, they published several thousand pieces; within two to three months, over 100,000—covering highly specific, long-tail terms in their domain.

Roughly one month in, these pages began appearing—and being cited—in AI search results.

The most immediate impact? High-intent, high-quality sales leads—both in volume and conversion quality—exceeded expectations.

That’s a critical signal.

For enterprises and their industries, the corporate website is a primary source—not just a brochure. Its overall domain authority may lag behind major media or government platforms, but within its own vertical, it sits closest to real products, live use cases, frontline customer problems, and operational reality.

Given sufficient depth, specificity, and consistent updates, a company’s site can become how search engines—and AI—learn to understand that business.

The real challenge for enterprise websites isn’t whether to publish at scale—it’s doing so continuously, affordably, at scale, and without sacrificing quality.

AI seems to solve that. But after reviewing dozens of implementations, I’ve found most miss the foundational logic—resulting in mass-produced, low-signal noise.

To sustainably publish at volume, you must first solve for quality and informational gain: each piece must rest on credible sources, answer one precise question, and add industry-specific insight missing from public documentation.

The underlying principle: quantity expands reach; knowledge density determines whether content gets indexed, cited, and ultimately converted.

Three capabilities are non-negotiable:

A high-fidelity, living internal knowledge base, continuously updated and structured: case studies segmented by industry, scenario, customer pain point, solution, and outcome; methodologies, FAQs, frontline observations, and capability maps—all validated, deduplicated, and transformed into publishable content.
A data-driven keyword selection and iterative content generation system, calibrated to search demand and competitive gaps.
Agent-powered review at every stage—fact-checking, coherence validation, SEO alignment, and tone consistency.

When an organization cultivates a self-renewing knowledge asset, content production gains a reliable source. Over time, the website evolves a clear chain: internal experience → structured knowledge → public content → search & AI citations → qualified leads.

Long-term, a company’s content competitiveness hinges on its ability to convert daily operations into trustworthy, searchable, and reusable knowledge.

ChatGPT Begins Targeted Evidence Retrieval

A recent monitoring report from an overseas GEO tool revealed a notable shift: ChatGPT Search now explicitly selects specific websites to retrieve evidence.

Observed changes:

Queries containing site: jumped from ~0.37% to 16.8%—a 46× increase.
Average subqueries per response rose from ~1.08 to 1.83.

In effect, AI has upgraded from “one-shot broad search” to a two-stage process: broad discovery, then targeted evidence retrieval—sharply improving answer reliability.

GEO implications:

Design each page to answer one clear question—and ensure your core channels (especially your official site) cover breadth and precision.
Make titles, dates, authors, and data sources explicit and machine-readable.
Isolate prices, specs, case studies, and definitions onto dedicated, indexable pages.
Maintain crawlability: clean HTML, semantic markup, logical internal linking, and fast rendering.
Cite original data sources and display last-updated timestamps prominently.
Minimize “fluff” pages—those heavy on slogans but light on verifiable facts.

simonwillison.net

A 17-Year-Old’s ICML Paper

In Zhang Xiao-Jun’s latest episode, she interviewed Su Tinghao—a 17-year-old student at a Hong Kong international school.

He recently published a paper titled Attention Projection Mixing with Exogenous Anchors as a sole author in the main conference proceedings of ICML—the International Conference on Machine Learning, one of the world’s most prestigious AI research venues. Papers undergo rigorous double-blind peer review before acceptance.

What problem does it solve?

In Transformers, word representations pass through successive layers. Deeper layers capture richer semantics—but early token identities risk being “washed out” by repeated transformations—a phenomenon known as over-smoothing.

Existing methods try to “look back” to the first layer for identity cues. But that creates tension: the first layer must both preserve stable identity information and perform its own semantic computation. Su calls this the First-Layer Tension.

His solution: ExoFormer, which adds a separate exogenous anchor module alongside standard Transformer layers.

Extracts a stable representation directly from input token embeddings.
Generates attention Q/K/V and gating signals G from that anchor.
Allows every layer to mix its current state with the anchor.
Lets the model learn how much to mix—and adapt the ratio dynamically per input.
Applies RMSNorm to anchors before injection, preventing scale drift across layers.

Think of writing a long article:

The exogenous anchor preserves “Who is this about? What’s the raw material?”
Transformer layers handle analysis, relationship mapping, and conclusion-drawing.
Every layer can consult the original source—no need to carry foundational identity forward as baggage.

He frames this as the offloading hypothesis: anchors retain lexical identity; the main network focuses on higher-order features.

Why it stands out:

Identifies a concrete architectural tension few had named.
Extends cross-layer reuse beyond simple residual connections—into the full attention pathway.
Delivers measurable gains: average accuracy across six downstream tasks rises from 48.80% (gated attention baseline) to 50.27%; validation perplexity drops from 14.64 to 14.09—with minimal added compute.
Offers testable, mechanistic explanations—not just empirical wins.

A few striking details:

He ran ~200 experiments (~¥20,000 RMB). There was no pre-planned roadmap—he discovered the First-Layer Tension and exogenous anchor idea only after extensive trial and error.
When scaling up experiments, he calculated an extra ¥4,000 cost. That night, he hesitated: rejection was likely, and the money might vanish. His parents and brother encouraged him to proceed—and he did. The paper bears only his name, but its completion relied deeply on their emotional and financial support.
~98% of the work was solo—no advisor, few peers to discuss modeling with. His most frequent collaborators? ChatGPT, DeepSeek, and online courses.
He openly credits AI: “Without ChatGPT, I couldn’t have navigated submission logistics, reviewer responses, or training scheduling.”
His personal goal? “To be someone who makes himself happy—and makes people he loves happy too.” To him, “happiness is fundamentally human”—AI or not.

Related resources:

Post-Search Marketing Strategy

This research paper tackles a persistent gap: e-commerce platforms lack effective cross-platform re-engagement strategies for users who search but don’t click or buy.

The team deployed a multi-agent AI system for product research and CRM reactivation. Here’s how it works:

Offline log mining: Identify users with strong purchase intent (e.g., searches containing “best,” “latest”), high spending capacity, and zero clicks—despite active exploration.
Multi-agent research pipeline: One agent interprets subjective intent and hidden preferences; another calls external search engines for competitive benchmarks; a third aligns findings to the platform’s inventory and generates personalized offers with real-time pricing/incentives; a fourth verifies factual accuracy (e.g., specs, availability).
WhatsApp delivery: Send concise, personalized recommendation reports + shortened links—prompting return visits.
Review Agent layer: Eliminates hallucination in product parameter generation.

They ran a 23-day live production test. Key results:

CTR increased 285% vs. conventional WhatsApp marketing.
Sent 15,061 messages → drove 37,258 page views.
Message open rate improved by 8%.

Three takeaways:

Trigger marketing at points of unfinished decisions: Zero-click searches, repeated comparisons, browsing without adding to cart—these are signals users need help concluding. AI should intervene here, not just after abandonment.
Make AI a trustworthy, personalized concierge: Fuse user intent, external reviews, internal pricing/stock, and verified facts—then answer clearly: “Why is this right for you?”
Measure AI marketing by incremental value: CTR reflects interest, but true ROI lives in incremental purchase rate, incremental GMV, gross margin, and cost per incremental order. Always run randomized control trials at launch.

Paper: arxiv.org

DeepSeek Harness: Curated Learning Resources

Here are 12 high-signal resources on DeepSeek Harness—including official docs, tutorials, architecture papers, and engineering guides:

Official GitHub repo: Source code, README, examples, architecture diagrams, version notes, and community links.
deepseek-harness
Official Web UI Guide: Installation, startup, model config, workspace setup, and quickstart walkthrough.
deepseek-harness.github.io
First Plugin Tutorial: Learn apply, inject, lifecycle hooks, patch loading, and minimal plugin structure.
deepseek-harness.github.io
Cordis Framework Paper: The theoretical foundation—effects, reactive dependencies, dynamic composition, hot swapping.
paper
Hierarchical Self-Improvement: Task-specific, evolvable agent frameworks—tested with DeepSeek-V4-Flash-Preview.
arxiv.org
Agent Harness Engineering Overview: Tools, context management, hooks, sandboxing, memory, and verification.
addyosmani.com
How Coding Agents Work: Models, system prompts, tool calling, state, and execution loops—explained step-by-step.
simonwillison.net
Harness Design for Long-Running Apps: Anthropic’s approach—Planner, Generator, Evaluator, task decomposition, cross-session handoff.
anthropic.com
Harness Engineering: Six-Layer Framework (Chinese): Layered architecture, context management, constraint enforcement, observability, recovery.
JavaGuide
What Does a DeepSeek Harness Researcher Actually Do?: Reverse-engineered from job posts—covers Agent Loop, Memory, Tools, Eval, Multi-Agent, and Self-Improvement.
reelos.ai
Anime Avatars ≠ Just Cosplay: How DeepSeek Harness reveals hidden patterns in AI culture.
mp.weixin.qq.com
DeepSeek Harness: From Setup to Plugins: System architecture, Profile/Bundles, plugin dev, tool design, lifecycle, debugging.
doc.laoyao.cn