The Age of the Deep Generalist: Who Gets to Teach the Machines

Somewhere upstream of every frontier model sits a payroll that would have made no sense a decade ago. Thousands of contracted humans, paid by the hour, producing nothing you could ship. Their entire output is judgment: grade this answer against that one, rank these five drafts, catch the subtle error in this proof, explain precisely why this argument fails. The models learn from the verdicts, and the verdicts are the product.

The prices in this market say something the industry’s marketing does not. The firms supplying human data to the labs publish rate cards that climb with the rarity of the judgment being rented: modest rates for routine annotation, comfortable professional rates for skilled reviewing and tutoring work, several hundred dollars an hour for practicing physicians and senior attorneys, and rates brokered toward four figures for the judgment of experienced founders and operators, the people whose taste cannot be looked up. Read the gradient for what it is. The machine-learning economy, the same one flooding storefronts with free competent text, has quietly become one of the highest-paying buyers of human discernment on earth, and what it pays for most is whatever discernment it cannot yet manufacture.

The interesting question is who ends up holding that scarce discernment, and the answer is changing shape for the third time.

Three Waves of Teachers

The first wave was the crowd. Early data work was mass annotation, thousands of workers labeling images and flagging toxicity at commodity rates, and it fit the era: the models were far below ordinary human competence, so ordinary human competence was the teacher. That work built the foundations, and then the models absorbed everything the crowd could show them, and the crowd’s lessons ran out.

The second wave is the specialist, and it is the current regime. Once models write cleaner prose than a casual worker and code better than a junior developer, the only humans left with something to teach are the ones above that line: the doctors, the lawyers, the mathematicians, the published writers hired to grade creative output. This is exactly what the rate cards reward, and the labs’ appetite for it has turned credentialed judgment into an export industry. The pattern underneath is worth stating plainly, because it predicts the third wave.

A model is taught by whoever still knows something it does not. The frontier of what the model lacks decides, year by year, who gets to be the teacher.

The third wave follows from where that frontier is moving. Specialist teaching has a built-in horizon: a specialist can lift a model toward the specialist’s own lane, and once the model performs at that level inside the lane, the lane has little left to give. Narrow benchmarks saturate one after another; each is a lane closing. What does not saturate, what the current systems still reliably fumble, sits between the lanes, and between the lanes is territory specialists were never hired to patrol.

Agents Broke the Lanes

The shift from chatbots to agents forced the issue. A model that answers questions can be graded a question at a time, by a matching expert a question at a time. A model that executes a workflow, researching a subject, drafting the text, generating the art direction, wiring the landing page, planning the launch, produces one integrated artifact that crosses half a dozen crafts before lunch. Somebody has to judge that artifact, and judge it whole.

Hand the finished campaign to a marketing specialist and they will assess the funnel while missing that the prose contradicts the brand’s voice. Hand it to the writer and they will fix the voice while missing that the visual identity drifted. Every specialist grades their slice honestly and the seams go unexamined, and the seams are where agent work actually fails: each step locally plausible, the whole quietly wrong. Evaluation of end-to-end work is only as good as the breadth of the person evaluating, which yields a rule the training economy is just beginning to price.

The judge has to be at least as broad as the work being judged. Agents that span crafts create demand for humans who span them too.

This is visible in how the expert-data firms describe their hardest searches. Filling a queue with board-certified physicians is a solved problem; the market found tens of thousands of them. Finding evaluators who can follow an agent across an entire project, holding a professional standard in each craft it touches and noticing what slipped between them, is the search that keeps coming up empty, because the credentialing system that produced the second wave was built to certify depth in one lane and has no certificate for the space between lanes.

What the Generalist Catches

Be precise about what this person brings, because generalist has spent decades as a polite word for shallow, and the teacher the third wave needs is anything but. Depth in one domain, passing familiarity everywhere else, describes a dabbler, and dabblers have nothing to teach a frontier model. The profile that matters is the deep generalist: real, practiced depth in several crafts, plus the connective judgment that only forms when one mind has worked all of them.

The cross-domain contradiction. The plot point that is fine for the story and fatal for the game balance, the cover that is beautiful and wrong for the genre shelf, the launch plan that maximizes reach while quietly breaking the world’s tone. No single-lane expert flags these, because each half of the contradiction lives in a different lane.

The plausible whole that is wrong. Systems that no longer make sentence-level errors fail at the level of the entire artifact, coherent in every paragraph and mistaken in sum. Catching that requires having shipped whole things, because only someone who has carried a project end to end knows what a finished-and-right one feels like.

The missing question. The sharpest expert feedback is usually not a correction but a question the work never asked: who is this for, what does it cost, what breaks downstream. Those questions live at the joints between disciplines, and the joints are precisely what a career inside one discipline never exercises.

None of this is mystical, just ordinary taste compounded across crafts, and it is learnable the way taste is always learned, by making things in more than one medium and being answerable for how they land. What has changed is that a market now exists that pays for it directly.

The Writer Was Already the Prototype

Here is the turn that matters for anyone reading this from inside a writing career. The profile the third wave is searching for, depth in several crafts plus connective judgment, is rare not because it is superhuman but because the twentieth-century economy punished it, sorting everyone into lanes and paying for depth alone. The people who escaped that sorting are disproportionately the ones the lanes never had room for, and working genre writers, especially independents, are prominent among them.

Look at what running a small fiction operation actually requires now. Story craft, obviously. But also art direction for covers, because the shelf is visual. Audio judgment, as audiobooks became half the market. Funnel economics, read-through math, pricing. Community, presentation, the discipline of a coherent world maintained across years. Nobody called this generalist training while it was happening; it was called making a living. It was also, sentence for sentence, the same exercise the evaluation economy now runs at scale: hold a standard in one craft, then another, then judge the whole.

The strategic point is larger than a side income grading model outputs, though that income is real and the rates are public. The same faculty that teaches a model in a lab is the one that directs models in a studio. Every writer who has learned to brief a system precisely, judge what comes back across prose and image and sound, and catch the plausible-but-wrong before it ships, is already doing third-wave work, just with the pay attached to the books instead of the verdicts. The deep generalist existed long before the training economy noticed. What is new is a market that finally bids for one.

Breadth Compounds Now

For most of industrial history, the advice was to specialize, and the advice was correct, because organizations existed to assemble specialists and the assembling was somebody else’s job. The machines have now taken over most of the assembling. What they cannot yet supply is the judgment that spans the assembly, and everything about the current evidence, the rate gradients, the saturating lanes, the agent workflows nobody narrow can grade, says that judgment is where the scarcity has moved.

So the practical counsel inverts. Add a craft, deliberately, and take it to real depth rather than tourism. Finish whole things, because wholes are where the rare judgment forms. Treat the connective questions, who is this for, what does it cost, what breaks downstream, as skills to drill rather than instincts to hope for. The specialists taught the second wave, and the models absorbed the lanes one by one. The third wave belongs to the people who can stand between the lanes and see the whole board, and for once, the market is paying them what that view was always worth.