Imagine an engineer who asks AI to draft an implementation, write tests and documentation, summarize papers and prepare a presentation. Work that once took most of a week can sometimes fit into an afternoon.

The next morning, salary and equity are unchanged. The calendar fills with tasks. The engineer is now expected to produce more, and the saved time does not become free time.

That tension stayed with me while rereading Eric Jorgenson’s The Almanack of Naval Ravikant. Its wealth advice predates everyday generative AI, but one distinction now feels sharper:

“Seek wealth, not money or status.”

When I read it beyond startup advice, the sentence concerns control. Money is a claim on future goods and labor. Status is a position relative to others. Wealth includes assets, capabilities and relationships that can keep producing value without selling every future hour again.

AI changes production costs. It leaves open who owns the result, reaches a market, bears risk or decides how to use productivity gains.

Agency is the ability to make and act on one’s own choices. I want to know how the tools change work, ownership and agency, as well as how fast they make us work.

Wealth changes which choices are possible

If wealth becomes a final score, income, savings, valuation, visibility and consumption blur into one number.

A high income can come with very little control. Someone may earn well but depend on one employer, one visa, or one narrow skill the market currently rewards. A successful company on paper can leave a founder unable to step away. Modest savings, meanwhile, can provide time to refuse a bad project, leave a damaging workplace, or wait for a better collaborator.

The common element is room to choose without immediately needing someone else’s permission.

This is also why wealth includes more than money. A trusted reputation lowers the cost of the next collaboration. Public work lets strangers evaluate what a person can do. A reusable tool keeps solving a problem after its author closes the laptop. Good relationships create possibilities no job listing contains.

Reputation, tools and relationships cannot pay rent by themselves. Material security comes first. Above that floor, I would ask how much of my future I can direct before asking how impressive the balance looks.

This view fits with an argument I made in Work Is Important Enough Not to Become Everything. Work can form skill, responsibility, and craft without becoming the sole judge of a life. Here I ask the next question: can the way we work gradually give us enough agency for a life that rests on more than employment?

AI makes leverage abundant

Naval’s name for what lets one decision reach beyond one person’s immediate effort is leverage. Its forms include capital, labor, code and media. Code can serve another customer while its author sleeps. An essay can reach another reader without being rewritten. Their low reproduction cost mattered to his framework.

Generative AI pushes this logic further. Code can now help write the instructions that it can run at scale. Media systems can generate drafts, variants, translations, images, and clips. An individual can research, prototype software, analyze data, and operate a small service with far less support than before.

This is real leverage. It is also widely available leverage.

Wide access changes the value of output. If everyone can make a plausible landing page in an hour, those pages become less scarce. Thousands of competent market summaries can become commodities. A team shipping ten experiments instead of two may face competitors doing the same. Attention, trusted distribution and worthwhile problems remain limited.

The 2026 Stanford AI Index reviews productivity improvements in structured customer support, software development and marketing work. It also reports smaller, sometimes negative gains on deeper reasoning or poorly matched tasks. Work that is easy to describe, repeat and check is often a good fit.

That is exactly the work whose price is most likely to fall as supply expands.

I prefer Arvind Narayanan and Sayash Kapoor’s AI as a normal technology framing to a race against an independent machine. “Normal” does not mean minor: they include electricity and the internet. Invention, application and adoption take different times, shaped by firms, professions, schools and laws. Model weights alone do not contain the future of work.

David Autor offers one version of the more hopeful path. In Applying AI to Rebuild Middle Class Jobs, he argues that AI could help more workers make decisions now reserved for a narrow group of experts. He calls this a possibility, not a forecast. The same technology can extend expertise or centralize it, make workers more capable or make their work easier to monitor and unbundle.

Job design decides much of that difference. Can workers question the system and learn? Do authority, credit and pay follow responsibility? A claim that AI helps workers needs to say who benefits, in whose organization and on what terms.

One update to Naval is that access to the tools does less to set someone apart as machine intelligence becomes widely available. Like electricity or cloud access, it can matter enormously without deciding who captures the return.

Distinguish tool access from ownership

There is a seductive sentence in the current AI conversation: everyone now has a team of agents working for them.

The image of an agent team encourages delegation and makes new projects approachable. Most people, though, rent model access, rely on platform distribution, build on interfaces that can change, and sell through discovery systems they do not control.

The distinction matters because a rented lever can increase output while leaving bargaining power unchanged.

Two researchers might use the same models. One produces more internal reports. The other publishes a benchmark and method, earns trust and maintains a tool. Both become more productive, but the second leaves work that others can use beyond the next review. For software teams, speed can become more durable through a direct user relationship and a product people would miss.

Founding is one option, not an obligation. Separate tool access from ownership by asking:

• Does the work leave behind a reusable artifact?
• Does it build a relationship with users, collaborators, or a field?
• Is there any claim on the upside if the work succeeds?
• Can the capability travel when the current job, platform, or model changes?

AI can make the answers better. It can also make them worse by creating an endless stream of disposable output.

Fig. 1. A productivity gain does not automatically become free time. Without a change in the bargain, it can return as a higher baseline of expected output.

Specific knowledge after general-purpose models

The book calls specific knowledge an unusual mix of ability, interest and experience beyond a standard training pipeline. Its claim that this cannot be automated has aged faster than the underlying idea. Models now do parts of programming, design, analysis, translation and writing once thought safely creative.

Specific knowledge was never just stored facts, one tool or a title. I understand it as knowledge of a particular situation that is hard to copy.

A researcher suspects a benchmark gain is an artifact because they remember how the dataset was assembled. An infrastructure engineer recognizes a scheduling problem inside a latency regression because they understand the workload and runtime together. A designer sees a polished interface solving the wrong problem after months of watching users hesitate.

AI may even supply the decisive clue. What stays hard to copy is the context that makes the clue meaningful, the judgment to act on it and the responsibility for being wrong.

This suggests a more durable definition for the AI era:

Specific knowledge is the combination of context, taste, relationships, and demonstrated judgment that lets someone use available intelligence unusually well.

The work that sets someone apart moves toward judgment. Writing code now matters less than knowing which system should exist and how to verify it. Producing prose matters less than having something worth saying and a reader who trusts the voice. In research automation, recognizing which result should change the question matters more than generating more experiments.

Skill also changes the learner, beyond helping them stay scarce. Practice develops patience, standards, courage and a more accurate view of reality even when markets price them poorly.

Apprenticeship develops these qualities through effort, resistance, correction and another attempt. The Notre Dame essay AI, Work, and Human Dignity describes a risk: AI can produce a skilled-looking artifact while bypassing the practice that develops its maker. Output can improve while the worker weakens.

A junior engineer who delegates every difficult implementation can ship more while losing the mental models needed to diagnose tomorrow's failure. A researcher who uses AI to summarize every paper can cover more literature while losing contact with how arguments are assembled. A student can submit polished prose without discovering that the underlying idea was vague.

Difficulty has value when it develops judgment, rather than for its own sake. Boilerplate automation can free attention for architecture. Translation can open a field. An explanation at midnight can keep learning going. Remove wasted effort while keeping direct contact with the problem.

Attempt a design before asking for one. Read some papers in full. Trace a failure beneath the generated fix. Use AI to widen the field of practice, not to eliminate practice.

Getting an answer and becoming someone who can answer responsibly are different achievements. The second still takes time.

What still compounds

AI accelerates production cycles. Compounding still rewards things that survive them.

Judgment compounds through feedback. Record the choice, expectation and result to build a history of consequences. The aim is to bring expectations closer to results and make that improvement visible enough for others to trust larger decisions.

Reputation compounds through attributable work. A consistent public record helps future collaborators reduce uncertainty. A name exposes someone to failure while keeping their good decisions attached to them.

Relationships compound because repeated cooperation reduces the work needed to negotiate. People who have solved hard problems together know how the other reacts to ambiguity, credit, and mistakes. Models can make introductions. They cannot compress five years of earned trust into a prompt.

Reusable systems compound. A script, dataset, course, library, publication archive, or well-designed process lowers the cost of the next useful action. Some earn money directly. Others create opportunities, bargaining power, or time.

Financial slack creates room for better decisions. Before investment returns become substantial, savings can let someone wait, learn, negotiate or leave. Scarcity shortens horizons. A modest buffer can protect longer plans from one bad month.

These advantages reinforce one another. Judgment improves work. Attributable work builds reputation and relationships. Collaborators expose better problems, while reusable assets free time for another considered decision.

This loop is slower than generating another hundred artifacts. That is why it is defensible.

Build ownership without risking the floor

Naval distinguishes income from selling hours and the possible nonlinear upside of ownership. I find the direction important and easy to misuse.

For someone without savings, stable residency, health, or support, a salary helps manage risk. It does not show a failure of imagination. Founding can create autonomy. It can also concentrate risk in one asset that is hard to sell and turn every waking hour into work.

Where circumstances allow, I would build ownership gradually rather than imagine a dramatic exit from employment.

An employee can keep a portable record of learning without exposing confidential work. A researcher can maintain open tools alongside papers. An engineer can turn a recurring task into a documented system. A creator can build a direct archive. A team can negotiate profit sharing, authorship, or equity when its contribution creates durable value.

Let some effort survive its immediate transaction. AI can reduce the cost of maintaining small assets, though fashionable output decays quickly. An asset lasts through its connection to a real need.

The right progression may be modest:

1. Use work to build capability and material stability.
2. Use AI to reduce the cost of creating something reusable.
3. Attach that artifact to a problem and a group of people you understand.
4. Keep enough ownership that success changes your future choices.
5. Do not risk the floor that makes long-term thinking possible.

This is slower than the fantasy of instant independence. It is also more compatible with an actual life.

Productivity does not decide distribution

The Anthropic Economic Index survey reports that many users believe AI improved the speed, scope, and quality of their work. These are self-assessments from people already using one provider's products, not a labor-market forecast. Those self-assessments help explain why adoption is attractive: people can feel the gain in capability immediately.

The International Labour Organization’s 2026 review finds real but uneven productivity gains. Reported time savings have not yet consistently become higher output, earnings or employment. It also flags risks to entry-level opportunity, worker autonomy and inequality. An IMF working paper models how existing capital owners may capture more adoption returns even as some wage gaps narrow.

Individual skill still matters. Productivity and bargaining power are different variables.

If a company owns the models, customer relationship, data, and resulting intellectual property, greater employee output may strengthen the company without creating a proportional employee claim. If entry-level tasks disappear, telling young workers to develop judgment skips the stage through which judgment was once learned. If access to the best tools depends on organizational budgets, “permissionless leverage” is less permissionless than it first appears.

Organizations and institutions therefore matter. Who receives equity or profit sharing? Whose work trains the system? Are productivity gains returned as pay, shorter hours, or only higher targets? Can junior people still perform enough real work to learn? Do workers have portable benefits and enough mobility to leave a bad bargain?

Jaron Lanier’s data dignity proposal extends the ownership question to human language, images, recordings, judgments and classifications behind models. Tracing outputs to contributors is technically difficult, and his proposed mechanisms remain contested. Describing contributors' work as coming from nowhere still makes them easier to ignore.

Attribution, licensing, professional associations, unions, data trusts, and collective bargaining may all become part of the answer. “Build a personal moat” is incomplete when value was produced collectively and captured through infrastructure no individual controls.

An individual strategy cannot decide how the gains are shared. I want both personal freedom where available and collective rules that prevent only a few owners from capturing technological gains.

A person is more than a portfolio

Compounding can also consume its purpose. Once judgment, reputation, relationships and health all become capital, every quiet hour looks like an underperforming asset.

That is not freedom. It is the logic of work following a person home.

Market value and human worth are different judgments. A model performing someone’s task does not reduce that person’s dignity. Nobody should need to prove they are uniquely creative, emotionally superior or permanently irreplaceable to deserve security and respect. Dignity is the floor of this argument, not a premium for scarcity.

Some important human activities look inefficient from the perspective of scale. Teaching requires noticing why this student is confused. Care responds to a body and history that do not fit a standard case. Friendship includes conversations with no deliverable. Play matters partly because nothing is at stake.

Allison Pugh calls part of this connective labor: seeing another person clearly enough for trust, learning, care, or cooperation to become possible. A teacher contributes more than information. A clinician contributes more than diagnosis. Attention can change what the other person is able to say and do.

AI can prepare lessons, translate, surface a medical pattern or remove administration. Employers can also turn those savings into larger classes, shorter appointments and empathy at production speed. Time promised for relationships can become another utilization target.

Fig. 2. Attention can create value precisely because it is given by one person to another, at a pace that cannot always be compressed.

Ask what responsibility the situation needs. Sometimes an accurate automated answer is enough. Sometimes it needs an accountable professional or another person who recognizes what happened and stays present. Treating these as identical services hides the social structure of care.

Amartya Sen’s capability approach connects wealth to what resources actually let someone do. The same salary permits different freedoms under insecure residency, inaccessible health care, caregiving or an uncontrollable schedule. An AI subscription gives little say if its user cannot question it, keep the benefit or refuse the pace.

Wealth, then, is effective freedom: real access to lives one has reason to value. It includes the ability to learn, care, recover, leave harm, and spend time without turning it into proof of ambition.

Assets can protect these freedoms: savings buy recovery time, ownership can strengthen bargaining and reputation can open doors. Their purpose matters. Accumulating some things should let other things escape measurement and sale.

I want to use powerful tools without making their easiest metric the measure of a good life.

How I would adapt without panic

Five principles seem durable to me.

Use AI aggressively, but keep a learning edge

Delegate repetition, first drafts, broad search, and low-cost experiments. Keep direct contact with a difficult layer where feedback can still surprise you. Read some sources in full. Debug some failures beneath the first generated explanation. Gain leverage without losing the ability to judge its output.

Turn a fraction of output into assets

Most work will remain transactional. Make sure some effort still accumulates: a tool, method, archive, audience, dataset, reputation, savings buffer, or ownership stake. If every productive hour disappears into someone else's queue, higher productivity mostly moves the queue faster.

Keep the cost of saying no within reach

Runway is a form of freedom. Lower fixed commitments, diversified relationships, portable skills, and adequate savings make it easier to reject work that consumes learning without building anything durable. It requires knowing which expenses buy a better life and which ones quietly make obedience mandatory.

Play long games with people, not only platforms

Platforms optimize discovery, then change their rules. Long-term collaborators remember context and make trust more valuable with use. Publish where work can be found, but build relationships that survive an algorithm change.

Define enough before the machine defines more

AI makes more drafts, experiments, features, and content possible. Without enough, every efficiency becomes another obligation. Enough is the boundary that lets capability return time to life instead of turning all life into throughput.

These principles offer a stance toward uncertainty: use the leverage, keep learning, protect the floor, and refuse to measure a life only by output velocity.

What the wealth is for

For me, the most useful claim in The Almanack of Naval Ravikant is that wealth is a means to something else. It offers a way beyond compulsive money problems and toward greater control of one's time.

That point matters more when the tools can work without rest. A person who defines wealth only as production will lose to the machine by definition. A person who defines wealth as status will remain trapped in comparison because AI expands the number of people who can look competent, prolific, and polished. A person who defines wealth as agency can use the same tools differently.

Agency means having a meaningful say over the problem, collaborators, pace, and terms. It is not complete independence. Almost no worthwhile life has that. We depend on institutions, infrastructure, care, and other people. The goal is not to eliminate dependence but to make it less coercive and more reciprocal.

That requires enough security to think beyond the next invoice, enough skill to enter a new domain, and enough reputation to be trusted. It also requires enough identity outside work that a failed project remains a failed project rather than a failed self, and collective arrangements that make choice possible for people whose job does not produce equity, public recognition, or a portable audience.

AI can widen both possibilities: help someone build an asset reaching millions, or help a company extract more output on unchanged terms. Good judgment may gain value, while bad direction may become more costly. The tools do not choose the bargain.

This is where defining enough becomes practical rather than philosophical. Some time should remain unoptimized because rest is not deferred maintenance. Some relationships should remain unscaled because attention is the relationship. Some interests should remain private because curiosity changes when it has to perform for an audience. A life in which every activity must become content, reputation, or optionality is not wealthy. It is a diversified work portfolio.

Enough will be different across lives and across seasons. It cannot be reduced to one savings multiple or career milestone. But it can still function as a boundary: this amount of security lets me refuse the worst bargain; this amount of work is compatible with being present; this ambition is mine even when nobody can see it.

The task is not to become busier than the machine.

It is to own enough of what compounds, and to build institutions that share enough of the gain, that a more capable machine gives people more life back.

The richest future is not one in which everything compounds. It is one in which enough things compound that not everything has to.

@article{ji2026wealthintheaiera,
  title = {When Execution Gets Cheap, What Still Compounds?},
  author = {Ji, Wenbo},
  journal = {fusheng-ji.github.io},
  year = {2026},
  month = {July},
  url = {https://fusheng-ji.github.io/blog/posts/wealth-in-the-ai-era/}
}

References

1. Eric Jorgenson, ed. The Almanack of Naval Ravikant: A Guide to Wealth and Happiness. The official free online edition used for this essay.

2. Understanding How Wealth Is Created. The book chapter collecting Naval's framework for wealth, leverage, ownership, specific knowledge, and judgment.

3. Stanford Institute for Human-Centered AI. The 2026 AI Index Report: Economy. A synthesis of current evidence on AI adoption, productivity, investment, and labor-market effects.

4. Anthropic. Economic Index report: Cadences. Usage and survey evidence on task exposure, automation, perceived productivity, learning, and job expectations among Claude users.

5. International Labour Organization. The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence. A review of experiments, firm data, platform studies, and worker surveys, with attention to uneven effects and job quality.

6. Emma J. Rockall, Marina Mendes Tavares, and Carlo Pizzinelli. AI Adoption and Inequality. An IMF working paper modeling how complementarity, displacement, adoption decisions, and capital returns can affect wage and wealth inequality.

7. David Autor. Applying AI to Rebuild Middle Class Jobs. An argument—not a forecast—for using AI to extend expert decision-making to a broader group of workers.

8. Arvind Narayanan and Sayash Kapoor. AI as Normal Technology. A framework that distinguishes AI capabilities from applications and adoption, and emphasizes the role of institutions in shaping outcomes.

9. Jaron Lanier. There Is No A.I. An essay proposing data dignity, attribution, and collective mechanisms for recognizing the human contributions inside large models.

10. Perla Khattar. AI, Work, and Human Dignity. An account of meaningful work, apprenticeship, craft, and the risk of producing skilled artifacts without forming skilled practitioners.

11. Allison J. Pugh. When AI Automates Relationships. An essay on connective labor and the human attention involved in teaching, care, counseling, and other relational work.

12. Internet Encyclopedia of Philosophy. Sen's Capability Approach. An introduction to evaluating well-being through people's effective freedom to pursue lives they have reason to value, rather than resources alone.

Note: This essay was drafted and polished with the assistance of ChatGPT (GPT-5.6), based on my reading of The Almanack of Naval Ravikant and the sources above. The illustrations were generated with GPT Image 2.

Updated 03 November 2026