The generalist's moment

Every second post promising you the generalist's decade has the mechanism backwards. The pitch is that AI rewards breadth, that your scattered curiosity is finally the asset, that the machine loves a generalist. It does not. AI does not reward breadth. It commoditises depth, which is a different event with a different winner, and the winner is a narrower person than the post wants you to be.

Watch what actually got cheap. The thing you used to pay a decade of someone's training for, the contract drafted, the query written, the page styled, the analysis structured, now comes back competent from a prompt at a price that rounds to nothing. That is depth, priced like a coupon, on tap. The specialist's whole moat, the ten years nobody else put in, is now a field anyone can rent by the call. Not degraded. Rented. The expertise is real and it is still good and it is available to the person next to you for the cost of the electricity.

So the specialist did not lose their skill. They lost the scarcity of it, which was the part that paid.

What did not get cheap is the judgment. The model drafts the contract. It does not tell you the contract was the wrong instrument, that what you needed was a different deal in a different domain. It writes a clean query. It does not know the number the query returns changes the growth plan, which changes the burn rate, which means the answer it just handed you is the answer to a question you should stop asking. It gives you ten competent outputs. It will not tell you which one is wrong in a way that only bites three domains downstream, at the seam where legal becomes pricing becomes a trust problem and nobody who speaks only one of those three languages can see it coming. That work has a name, and the name is not new. It is integration. Choosing which skills to combine, judging what the combination is for, directing it across fields that do not share a vocabulary. It has never been buyable, and the tools did not make it buyable. They made the thing it sits on top of free.

The economics is plainer than the vibe around it. A task splits into the depth, executing the narrow skill, and the integration, deciding which skills and aimed at what and judged how. The tools drove the price of depth toward its marginal cost, which is roughly the cost of nothing. When one input to a product goes abundant and free, the return does not evaporate, it moves. It concentrates in whatever input stayed scarce. Depth is the abundant input now. Integration is the scarce one. The margin goes where the scarcity is, and the scarcity moved. That is not a prophecy about generalists winning. It is where money goes when a complement gets commoditised, which is a thing money has always done and will keep doing whether or not anyone writes a post about it.

And this is the part that kills the flattering reading, so sit in it. Cheap depth does not favour breadth. It favours one narrow capability that happens to require breadth: running the system across domains, which you can only do in fields you took far enough to judge the output and steer it, not just request it. You cannot tell a good legal draft from a plausible one if you never went past the wall in law or something next to it. You cannot combine skills you only sampled. The generalist who captures this is the finished one, deep in at least two fields, conducting the cheap rest. The dabbler does not get the moment. The dabbler gets an upgrade to the machinery of dabbling: more fields sampled, none judged, a longer list of things they can now request confidently and still cannot direct. The same cheap tools that hand the integrator an orchestra hand the dabbler a faster way to be wrong in ten new fields before lunch.

Why integration resists the commoditising in the first place, mechanically. Value here is not the sum of the skills, it is the fusion, the pair test's whole point. A model trained to do X well and Y well does not thereby know that X's answer changes what Y should do on this exact problem. The economists have a dry name for where new value comes from, recombination: novel combinations of things that already exist, and the space of possible combinations grows faster than the stock of things being combined. The printing press was a wine press, movable type, and metallurgy, none of them Gutenberg's invention, fused into one thing by someone who could see across all three. Recombining across domains is the scarce act. It scales with the number of domains you can actually judge, which is a human quantity, and a finished one.

It needs aim and doubt, and the tool supplies neither. The wand executes tirelessly and hands you no aim and no doubt, which is cluster 1's whole point and it holds here. Integration is mostly aim, which problem and framed how, and doubt, which of these ten cheap outputs is wrong in a way that matters three fields downstream. That is discernment stretched across domains, and it is exactly the input the tool multiplies instead of replacing.

Now the strongest version of the case against, because it is good and it is half right. The objection: this has it backwards, AI is a depth-amplifier for specialists, not a leveller for generalists. The hard problems live in wicked domains, Epstein's word for the ones with unclear rules, delayed and misleading feedback, patterns that do not repeat. To judge a model's output in a wicked domain you need deep domain expertise, which is precisely the specialist's asset. And the field data agrees, at least in software. The Copilot experiments across thousands of developers found the least-experienced gained most, closing the gap on routine work. The METR trial went further and stranger: experienced open-source developers using an AI assistant ran about nineteen percent slower on their own repositories while believing they were faster. Translate that: the generalist is the novice in every field. Cheap depth lets them generate confident output in domains they cannot evaluate, which is not useless, it is worse, it is fast plausible error. The person who actually captures AI is the specialist who now has a tireless junior, and your "integrator" is a project manager the model will commoditise next. So this is the specialist's moment, and "generalist" is the kind word for someone about to ship confident nonsense at scale.

That objection is right about the failure mode and wrong about who it names. It describes the dabbler, exactly, the person who samples a field, cannot judge its output, and now produces plausible error fast. That is the person this essay excludes on purpose, and I concede the danger at full volume, because it is the reason the claim has to be narrow. Three things. One, the Copilot evidence is the thesis, not the rebuttal to it: least-skilled gain most on routine tasks is depth getting cheap and the floor rising, arriving on schedule, and the objection quotes only the half that stings. Two, "you need depth to judge output" is not an argument against integration, it is the definition of the integrating generalist. The integrator is not deep in zero fields, that is the dabbler again. They are deep in two or more and conducting the cheap rest, and the whole bet is that judgment in your finished fields plus cheap output everywhere else beats deep-in-one plus expensive-or-absent everywhere else, on problems that cross domains. Three, the specialist with a tireless junior is real, and inside a kind, single-domain problem the specialist should win, granted, no argument. But the value is migrating to the seams, the pricing-and-legal-and-trust problem that no one field frames, and at the seam the specialist's single language stops helping and the integrator's cross-domain judgment is the only thing that does.

One honest qualification, because a sharp reader is already holding it. Integration does get sold. Consultancies sell it, systems integrators sell it, there is tooling that sells orchestration. True, and it does not rescue the objection, because what those sell is integration inside a domain, or integration as a generic service bolted on from outside. What stays unrentable is the judgment across the specific domains you personally finished, aimed at a specific seam you can see because you hold both sides of it. You can buy a systems integrator. You cannot buy the person who knows that this legal edge case is a pricing decision is a trust problem, because knowing that requires having gone deep in more than one of those and nobody sells that by the hour. And a scope flag on the receipts, since I lean on them: METR and the Copilot experiments are about software developers specifically. That is real evidence that depth is getting cheap and that novices gain most on routine work, and it is software. Do not stretch it into a law of all knowledge work. It is a strong instance, pointing the direction, not a proof that covers every field.

So the move, and it is one decision you can run this quarter, not a mindset.

Name your two finished fields. Not your interests, your finishes: two domains where you went past the wall and could survive the depth check, an output a stranger could point to. If you cannot name two, this essay is not yours yet, and the honest thing is to go finish one, because that is the prerequisite and there is no clever route around it.

Find one live problem that sits in the seam between them and would have needed a whole team a year ago. Specific, not abstract. You know you have the right one when each single field would frame it wrong, because the problem is really about what one domain's answer does to the other. That seam is your position.

Then conduct, do not execute. Rent the depth of every field you are not deep in, let the model draft the specialist's work, and spend your scarce hours on the two things it cannot do: aim, deciding which problem and how it is framed, and doubt, catching the cheap output that is wrong in a way that bites three domains down. And the diagnostic that closes it, which you can run on this week without asking anyone's permission. Count the hours you spent executing depth a model could have rented, against the hours you spent judging and directing across fields. If the first number is the bigger one, you are still competing as a specialist in a market that just made your depth cheap, and you are leaving your own moment on the table for someone who figured out they were supposed to be holding the baton.