Training Yourself Like a Large Language Model

There’s a way to grow: treat yourself like a large language model—and train accordingly.

For example, release yourself as a new version every so often: v1.0, v2.0…

The core concepts of LLM training—pretraining, dataset curation, post-training, feedback loops, reward functions, and iterative correction—apply just as meaningfully to human learning and growth.

1. Pretraining

Pretraining is the first phase of LLM development.

For humans, the pattern is similar.

Every day, your attention is your training data pipeline.

What books you read, who you talk with, which projects you take on, what content you consume—all shape the features and quality of your internal dataset.

Attention determines what you learn, how deeply you integrate it, and how coherently it becomes part of your judgment.

So managing input—and especially attention—is foundational.

To improve output quality, start by upgrading your training data.

Input starts the process. You must also process, compress, and restructure information—through writing, sharing, designing solutions, or making decisions.

Only when you form independent judgments—and consistently produce high-quality output—do ideas truly become part of your “parameters”: internalized, operational, and reliable.

2. Post-training

For people, the most critical post-training happens in the real world.

Ship a product. Deliver a proposal. Close a sale. Then wait for reality to respond.

Each meaningful feedback loop recalibrates your understanding.

Execution generates feedback; reflection turns feedback into proprietary knowledge—your own curated fine-tuning data.

Growth thus becomes a cycle: Input → Think → Output → Execute → Feedback → Correct → Execute again

Much of what’s truly valuable—insight that can’t be Googled or taught—comes only from this kind of firsthand, execution-anchored learning.

3. Reward Function

Another vital layer: your personal reward function.

Put simply: What do you consistently use as your metric for “good”?

What are you optimizing for—long-term value, solving hard problems, or short-term validation?

If you habitually reward yourself for external approval, you’ll get better at winning approval while meaningful work may receive less training signal.

If you reinforce curiosity, rigor, impact, and integrity, your internal model will gradually converge toward those traits.

You strengthen whatever you reward.

Your long-term trajectory depends less on raw talent than on the reward function you consciously design—and periodically revise.

So pause regularly: audit your inputs, update your reward logic, refine your model—and ship your next version.

Becoming the Question-Setter

There are two ways to live: as an answerer, or as a question-setter.

Answerers work within problems defined by others: exams, assigned tasks, client briefs.

Question-setters define the terrain itself: launching research directions, spotting market gaps, redesigning product logic.

The difference lies in problem ownership.

Answerers solve other people’s problems. Question-setters create their own—and invite others to help solve them.

Research consistently shows that breakthroughs often start with superior question-framing, supported by the ability to answer once the right problem is chosen.

Real growth often begins as you move from finding better answers to asking better questions.

Management, at its core, is about setting one strong question after another.

Helping teams own meaningful problems is the highest-leverage act of leadership.

A good question satisfies three criteria:

Defines boundaries Clear goal, scope, and success state—leaving space for execution. A great question clarifies “Why?” and “What does ‘done’ look like?” while leaving the implementation path open.
Generates ownership It must feel like their question, with ownership replacing passive compliance. When a task becomes “my question,” responsibility, creativity, and commitment rise sharply.
Leaves room for discovery Hold back the urge to supply the “right answer.” Real questions have no single solution. What matters is keeping the space open—so teams explore, test, and learn, expanding beyond validation of pre-existing assumptions.

Answers are becoming cheaper by the day. What’s scarce—and increasingly decisive—is the ability to frame the right question, and to cultivate organizations that keep doing so.

Managing Attention

In conversation with Xiang Yang, he recommended The Attention Merchants—a book whose central thesis is simple yet profound: Attention defines who you are.

We need active attention management: selection, protection, and periodic review.

From a commercial perspective, individual attention is the raw material of the advertising industry.

The First Principles of the Attention Economy

Attention is finite—and non-storable. Every focus choice excludes other possible experiences.
Giving attention equals giving time—and the experience that time could have generated. Your lived reality is shaped by what you attend to.
Attention allocation literally constructs your subjective reality.
Commercially, attention is harvested, priced, and resold—like any industrial commodity.
“Free” content is a trade: users pay with time, cognition, and behavioral data; advertisers pay cash; platforms broker the exchange (e.g., TikTok, Baidu).
The attention industry thrives on new frontiers: each medium—newspapers, radio, TV, WeChat, short video, live streaming—extends its reach deeper into private time, space, and relationships.
Your life experience is built from sustained attention. Whoever shapes what you see, think about, and wait for—shapes how you live.

How Attention Turns Into Revenue

Demand engineering: First create desire, then slot products into that desire—this is advertising.
When products attach to identity, class, emotion, or values, purchase motives transcend utility.
Once measurable (via ratings, surveys, user profiles, click data), attention enters industrial trading—segmented, compared, priced, optimized.
Uncertain rewards drive compulsive checking. Feeds and notifications exploit variable reinforcement—you keep returning because you don’t know when the reward will arrive.
Platforms subtly steer choice: ranking, headlines, push alerts, and recommendation engines manipulate probability before you click—designing the environment where decisions happen.

How Attention Reshapes Power & Relationships

Platforms amplify preexisting human drives—narcissism, curiosity, belonging, comparison, voyeurism, validation—and make them easier, faster, and more frequent to satisfy.
Media creates the illusion of intimacy: one-way attention to influencers or celebrities feels like relationship—then converts into commercial leverage.
Fame becomes capital. Reality TV and social media industrialize celebrity-making—ordinary people now curate visibility as deliberate labor.
Attention converts into money, obedience, willingness to fight—or even sacrifice. Advertising and propaganda share the same root chain: occupy consciousness → organize emotion → direct action.
Freedom depends on available options and awareness of those options. When platforms decide which facts, products, or political ideas enter your field of view, they shape your range of possible choices.

Industry Cycles & Personal Agency

Competition for attention pushes tactics toward ever-greater intrusiveness—stronger stimuli, deeper engagement, longer retention.
Public backlash forces adaptation: regulation, resistance, and tech-based countermeasures—eliminating crude methods and replacing them with subtler, more embedded ones.
What shocks one generation becomes routine for the next. Habituation continuously shifts social boundaries—letting commerce penetrate further into private life.
Innovation, expansion, and resistance form a loop: new media opens attention territory → industry scales harvesting → public pushes back → rules evolve → next tech opens fresh ground.
Attention autonomy requires both individual discipline and collective guardrails. We need to protect certain times and spaces as off-limits—and institutional limits on how deeply attention merchants may enter our minds and hearts.

A New Paradigm for Technologists

A tech leader once told me: “You’re all coding now—I’m going to be obsolete.”

I replied: “The hardest parts—architecture, system design, security—still demand deep technical judgment.”

In the AI era, technical value shifts toward higher-leverage layers.

Focus moves toward infrastructure: API design, agent orchestration, standards, safety, and interoperability.

Meanwhile, UI layers and basic application logic increasingly fall within reach of domain experts—via vibe coding: intuitive, low-code, prompt-driven development.

Business teams now own the “last mile” of software—defining needs, iterating UI, deploying internally, and refining daily based on feedback.

Technologists steward the underlying infrastructure—the rails, gates, and guardrails—that make this safe and scalable.

This reflects three foundational shifts:

Business people gain software agency: workflow simplifies to define need → vibe code → call APIs → build UI → deploy → iterate daily. Whoever understands the problem best sits closest to the code.
Technical leverage rises upstream: A great CTO’s core skill becomes designing infrastructure that lets 100 non-engineers safely vibe-code while keeping the system stable.
The essence is clear: business owns the application layer; tech owns the infrastructure layer.

As this evolves, engineers may increasingly split into two archetypes: platform engineers, building shared foundations—or functional domain engineers (FDEs), embedded directly in business units.

A company’s future technical strength may hinge on a new metric: How many domain-savvy people can safely, efficiently build software—without compromising security, consistency, or scalability?

New Management in the AI Era

Managers’ most vital work increasingly centers on finding and framing questions.
Methodologies are mature and widely accessible. The scarcest judgment is what to solve. Good questions are rarer than good answers.
The new capability loop: spot problems → define them clearly → select the highest-leverage ones → allocate resources → establish evaluation criteria.
This is the shift from answer competence to question competence.
The core flywheel of AI products: product ↔ data ↔ evaluation ↔ model, in continuous feedback.
More teams will be small—3–5 people—deeply focused on one domain. The strongest AI-native organizations run on trusted, autonomous experiments: tolerate failure, prioritize learning.
Great AI teams optimize for information gain across experiments. Innovation speed = high-quality experiments per unit time.
So AI-native organizations resemble this loop: hypothesis → experiment → eval → feedback → adjustment → repeat.
Every engineer increasingly acts like a leader, orchestrating multiple coding agents simultaneously.
Management evolves into: Leader → person + agent.
A new foundational skill emerges: breaking goals into AI-executable tasks, then designing context, tools, skills, and evaluation criteria for each.
A key managerial muscle: continuously identifying, defining, and assigning problems worthy of joint human-AI effort.
Part of future management is leading digital workers who operate outside offices and physical bodies.
Cross-functional collaboration hinges on building mutual trust. The hardest part is empathy, perspective-taking, and shared mental models.
In the AI era, context—structured, trusted, up-to-date knowledge about your domain, customers, and systems—becomes one of your most strategic assets.