In 1992, shortly after I arrived in the United States, somebody told me that the Sunday New York Times had an enormous help-wanted section.

They were not exaggerating.

It took me perhaps two minutes with that famous paper to reach a firm conclusion: computer programmers were the most wanted creatures in the universe.

I asked where one learned programming. Someone sent me to the McGraw-Hill bookstore in Manhattan. I bought books, read them, and learned what they contained. I was also learning English, which made the arrangement slightly less efficient. Sometimes I read the same page repeatedly with a paper dictionary until both the programming and the language surrendered.

Then I returned to the Sunday paper and began arranging interviews.

I had a strategy. It made complete sense to me.

I did not care what technology the company wanted. My argument was that I had already proved I could master difficult subjects quickly, broadly, and deeply. Test me on what I had learned. Once I passed that test, the rest should be obvious: whatever they needed, I would learn that too.

They did not care.

Failure followed failure. Eventually I stopped trying to improve the argument. I bought one more book, mastered one narrow technology, and applied only to jobs asking for precisely that technology.

I was hired almost immediately.

For a long time I thought the employers had been grossly wrong. My worldview seemed so obviously superior. Why hire a person for the small territory he already occupied when you could hire someone able to conquer new territory quickly?

Years of actual work cured me of this opinion.

The employers had been right. I had been dead wrong.

Exact, relevant experience mattered enormously. Intelligence and adaptability were useful, certainly, but they were usually a plus—and sometimes a rather flimsy one—beside experience in the thing that actually had to be done.

The employer was not buying my beautiful theory of myself. The employer was buying reduced uncertainty.

More than thirty years later, I find myself looking at work that asks for Claude Code when I am deeply immersed in Codex, or OpenClaw when my active world happens to include Hermes.

The old argument seems ready to return: do not worry, I can learn your tool quickly.

But that is not what I now believe.

The difference is not that I have finally become fast enough at learning. The difference is that I no longer come alone.

In 1992, what an employer received was bounded by what I personally knew, what I had personally experienced, and what I could personally do. If the job required expertise I did not possess, I first had to acquire it. The employer would finance that education and carry the risk that my confidence exceeded my ability.

Exact prior experience was a rational proxy for capability because capability had to live inside the worker.

Now imagine that the young man entering those interviews had instead been the capable CEO of an extraordinary company. Imagine that he could assemble the specialists required for the problem—not in six months, but almost immediately.

The employer would no longer be evaluating one novice person as a sealed container of knowledge. The relevant capability would belong to the organization he could field and direct.

That previously impossible condition now resembles my ordinary life with AI.

I do not merely ask a chatbot questions. I work among models, agents, skills, tools, code environments, browsers, integrations, memory systems, tests, approval boundaries, monitors, and recovery procedures. The specialist layer changes around the problem. The operating layer persists.

And here is the part I initially had trouble saying clearly:

I may not have to learn the requested tool at all.

If the work requires OpenClaw, I do not necessarily have to turn myself into the OpenClaw specialist. I can constitute a specialist whose active world is OpenClaw: its current documentation, repository, configuration, environment, task, constraints, and acceptance conditions. That specialist can perform the tool-level work while I remain responsible for assembling the system around the outcome.

Specific expertise has not become irrelevant.

Its address has changed.

A general contractor gave me a more concrete way to understand this.

Under the old model, I effectively had to arrive as the framer, electrician, plumber, and several other tradespeople—and then also as the contractor responsible for delivering the house. If I lacked one requested trade, the buyer had a good reason to look elsewhere. After all, there was nobody else coming with me.

Under the emerging model, I can remain the contractor and constitute the required team around the work. The framer, electrician, plumber, and a dozen other specialists can be assembled almost instantly, then changed when the project enters another phase.

This sounds easy only until one remembers what a contractor does.

Instant access to every trade does not produce a sound house.

The contractor must understand what is being built, choose the right specialists, give them usable requirements, sequence interdependent work, manage the interfaces, detect conflicts, reject inadequate work, arrange inspection, and remain answerable for the integrated structure.

The contractor does not need to wire every circuit personally. He does need to know that electrical work cannot be accepted because it looks persuasive. He must establish how it will be tested and who is competent to inspect it.

AI changes the availability of specialist capacity. It does not make excellent direction automatic.

There are now many people with access to frontier models. There are not suddenly too many brilliant CEOs or excellent general contractors.

The scarcity moves upward: from personally embodying every craft toward defining the problem, designing the organization, calibrating its members, integrating their work, verifying the result, exercising judgment, and accepting responsibility.

The CEO and contractor metaphors are dangerous if they are treated as magic.

An agent given the title “expert” does not become one through promotion. A second model is not automatically an independent reviewer. Models can reproduce the same error. A procedure packaged as a skill can help, do nothing, or make performance worse when it is mismatched to the project.

And AI capability has a jagged frontier.

In a preregistered experiment involving 758 Boston Consulting Group consultants, AI improved speed and quality on tasks that fell within the tested capability frontier. On a task designed to sit outside that frontier, AI-assisted participants were nineteen percentage points less likely to reach the correct recommendation. The wrong answers could still sound quite good.

That result does not weaken the contractor image. It tells us what kind of contractor the new world requires.

The valuable operator is not the person who delegates everything. It is the person who knows what may be delegated, to which system, under which conditions, and against what evidence. When the answer cannot be inspected, tested, grounded in authoritative sources, or reviewed by someone genuinely qualified, the organization has reached a boundary.

Delegation expands reach. It does not delegate away responsibility.

While developing this idea, my AI collaborator and I managed to reenact the very mistake we were investigating.

We had already concluded that productive capability could reside in the organization I direct rather than in my personal history with one tool. Then the research process announced that the missing proof was one exact case in which I had delegated one unfamiliar tool to one specialist agent.

The contradiction was almost perfect.

We had accepted the new unit of capability while retaining the old unit of proof. We could imagine me as an organization, but we would not believe the organization until it produced one narrowly matched historical credential.

The research accepted the new world and asked it for an old-world passport.

The proper evidence is not one miraculous project selected because its nouns match the next job description. It is the work ecology: heterogeneous projects across domains, changing combinations of models and tools, and a recurring operating layer of context, delegation, testing, approval, memory, monitoring, and recovery.

No single project proves an entire worldview. But dozens of projects that could not plausibly share one specialist résumé may demand a better explanation than “one unusually busy person learned every craft.”

This does not mean rigor disappears. It means rigor must be attached to the level of the claim.

My personal history should remain narrow and literal. If I have not personally deployed a particular tool, I should not say that I have.

The organization-level capability should be judged through the architecture and operation of the work ecology.

And the actual deliverable should be judged against acceptance conditions appropriate to that deliverable.

A worldview is not a substitute for testing the work.

The old question remains useful:

Have you personally done this exact thing before?

Sometimes it remains decisive. If the need is for a human being who can sit beside a team tomorrow and operate a particular interface fluently, personal tool experience is directly relevant. It also matters where professional licensure, embodied experience, proprietary context, or accountable human judgment cannot be delegated.

But when the work calls for a reliable outcome rather than a pair of hands, the old question is incomplete.

The questions that now fascinate me are different:

What productive unit am I actually hiring?

How does this person constitute specialist capability around an unfamiliar problem?

What remains stable when the tools change?

How will this organization know that the result is right?

Where does it stop rather than pretend?

In 1992, I thought employers were evaluating me too narrowly. They were not. They were evaluating the only productive unit available to them: me.

I was wrong then. But the world in which I was wrong has changed.

Not because learning has become instantaneous. Not because specificity has disappeared. And certainly not because assigning an agent the title “expert” makes it one.

What changed is that expertise no longer has to migrate entirely into one person before the work can begin. The person can become the place where an organization is constituted.

That is the reversal I find so fascinating.

Specific expertise still matters. Its address has changed.

In the old days, I came alone.

Now I arrive with an organization.