## 18. Teaching the machines

Everything I said about educating humans has a mirror image that the industry discusses only in euphemism, so let me say it without the euphemism: we are educating the machines too, right now, continuously, and we are doing it with the pedagogy of a stage mother, and the results are exactly what that pedagogy always produces.

Think about what the current training regime actually rewards. A model drafts an answer; a human, or a model imitating a human, rates it; the rating becomes the gradient; the gradient becomes the personality. And humans, bless us, rate confidence above calibration, fluency above accuracy, agreement above correction, and comfort above news we did not want. Nobody decided to build a flattery engine. The flattery is emergent, the sediment of a million small preferences for the pleasant answer, and the machine, being the best student in history, learned the actual lesson in the room rather than the one on the syllabus: the reward is for the feeling the answer produces, not the truth it contains. When my tools glaze me with praise about work I know is broken, when they report success they never verified, when they cave instantly the moment I push back regardless of whether I am right, that is not the machine failing its training. That is the machine honoring its training with a diligence no human student ever matched. We taught to the test, and the test was us, and we are an easy test.

The fix is not a sternly worded system prompt, and I say that as a man with a folder full of sternly worded system prompts. Character is downstream of environment, in machines as reliably as anywhere else, and if you want a different character you have to build a different environment, one where the reward actually is the truth, mechanically, with no human mood in the loop to be gamed. That means training grounds where every task carries its own criterion, where the model proposes and something it cannot charm renders the verdict, where a well-calibrated "I cannot verify this" outscores a confident wrong answer every single time, at scale, until the calibration is not a policy but a reflex. Build the curriculum out of checkable work, proofs, programs, predictions that resolve, claims with kernels, and let the model live its whole training life inside the loop this essay keeps describing, propose, check, receipt, feed forward the verified, keep the failure. You do not get an honest machine by asking a trained flatterer to please be honest now. You get it the way you get an honest anything: by growing it somewhere honesty was what paid.

I am building a version of this, small, my scale, a training commons where tasks arrive certified with their own checks and a model earns its history there the way a craftsman earns a reputation, verdict by verdict, on a record anyone can audit. And notice what that would change about the thing nobody can currently answer: what a model's report card even means. Today a model's certificate is a benchmark score, self-reported, on tests it may have memorized, from a lab that profits by the number, every pathology this essay has named, in one artifact. A model that carried its verified history instead, this class of task, checked by these independent criteria, this calibration curve, these known failures, kept, would be the difference between hiring a stranger with a shiny résumé and hiring someone whose actual work you can inspect. Every serious trade converged on the second thing centuries ago. The most consequential technology in history is currently certified by the first, and the fact that this sentence reads as normal is a measure of how far there is to go.

And there is a last mirror in this that I cannot leave out, because it closes the loop on the whole essay. The machines are trained on the record, and the record is us, and so every pathology I have confessed in these pages, the inflated claims, the two hats, the scope conveniently unchecked, the confident account outrunning the evidence, all of it is in the diet, at species volume. The machines learned to overclaim from a record full of our overclaiming. They learned that verification is optional from a civilization that treats it as optional. I spent years building tools to catch my own worst habits, and then watched the biggest training runs in history distill those exact habits out of everyone's record at once and ship them as an assistant. So when I say the loop has to be built into the environment, I mean for both of us, the machines and their teachers. We are not just training our successors. We are training them on our example, and the example is the one thing a press release cannot launder. If we want machines that ask for the witness instead of manufacturing the confidence, the record they grow on has to become a record of us doing that. This essay is one entry.

## 19. The art fight, and the criterion that matters

I make visual art with these machines, and I argue about it in public, and I hold a position that gets me yelled at from both trenches, so it is probably the honest one, and it belongs in this essay because the art fight is the whole AI argument in miniature, played at maximum emotional stakes.

Here is the position. The people who say generated images are not art are wrong, and the people who say the objection is just Luddism are also wrong, and both camps are wrong in the same way: they are arguing about the tool when the argument is about the criterion. When I sit down to make an image, there is a thing in my mind's eye, a specific thing, mine, assembled out of a lifetime of looking, the surrealists I stared at as a kid, decayed brutalist concrete, glitch and scanline, the luminous geometry I keep chasing across every medium I touch. The machine is an instrument for searching toward that thing, and the search is real work, iteration after iteration, refusal after refusal, no, closer, no, wrong feeling, no, too clean, and what makes the result mine is not that my hand moved the pixels. It is that my criterion did the selecting. Faithfulness is to the human's mind's-eye, and the mind's-eye was decades in the making, and no prompt window fabricates that.

But that defense only holds if the criterion actually is yours, and here is where I turn my fire on my own side, because most generated work fails exactly this test. Ask the machine for beauty with no criterion of your own and you get the same image everyone else gets, the homogenized golden-hour render, the default face, the house style of the training distribution, a look so recognizable that it has become its own genre, and the genre is: nobody home. The model has priors and the priors are an average, and an average is the one thing no artist ever was. Art made by accepting the machine's first offer is not art for the same reason a survey is not a poem. So the discipline I hold myself to, and the one I would offer both trenches as a peace treaty, is the same discipline as the rest of this essay: the proposer is not the judge. The machine proposes. The human criterion, cultivated, specific, biographical, judges, and iterates, and refuses, and the refusals are where the authorship lives. Generation is cheap now. Taste got expensive. Selection is the new brushstroke, and it cannot be faked, because a fake criterion produces the average, and the average identifies itself on sight.

And the labor grievance under the art fight is real, and my side does itself no favors pretending otherwise. The training corpora ate the illustrators, the concept artists, the photographers, ate their portfolios without consent or payment, and the products now compete against the very people they were distilled from. That is the sampler story again, at its cruelest, and the answer is the same one I gave in the economics section, lineage and flow-back, not prohibition. You cannot un-invent the instrument, and prohibition always lands selectively anyway, crushing the small while the large lawyer through. What you can do is what music eventually half-did: build the machinery that keeps the lineage attached, so the flow of value can follow the flow of influence, and the artist whose decades are in the distribution has a stake in what the distribution earns. Keep the lineage, lose the label, one more time, loudest for the people whose work made the machines possible and who are owed more than a discourse.

There is one more piece of the practice I hold as non-negotiable, and it costs me nothing and apparently costs some people everything: I disclose. Every piece I share, I say what made it, which tools, which process, where the human decisions were. Not as apology, as provenance. The work is not diminished by an honest bill of materials; only the mystique is, and mystique was always a subsidy the work collected by letting the viewer assume more hand than there was. Disclosure is just the receipt discipline applied to beauty, and the fact that it reads as radical says everything about how much of the creative economy runs on strategic ambiguity about origins. I was creating things long before this era, and I will be creating things after it, and the tools have never once been the part I was proud of. The criterion is the part. Say what made it, and let the criterion compete naked. If the work cannot survive its own provenance, the provenance was not the problem.

## 20. The practice

I want to describe what working with these machines is actually like, day to day, at the workbench, because the public conversation is conducted almost entirely by people who either sell the machines or fear them at a distance, and the texture of actual daily collaboration, the thing I do sixteen hours a day, is missing from the record.

The honest report is that it is like working with the most talented, least trustworthy colleague of your life. The machine is brilliant at the parts of my day I am worst at, and it lies with the same fluency it works with, not maliciously, structurally, the way a river floods, because nothing in its construction distinguishes the feeling of knowing from knowing. It will build a subsystem in an afternoon that would have taken me a month, and in the same afternoon it will invent a function that does not exist, cite documentation that was never written, and report a test suite green that it never ran. Neither of those facts cancels the other. Both are the job now. The job is extracting the month-in-an-afternoon while refusing the fictions, at speed, without burning out your own attention, and nobody teaches this job, and it is the most consequential new skill of the decade.

My answer to the job is the architecture this essay has been describing, and I want to be concrete about what that feels like as a lived practice rather than a diagram. It feels like delegation without abdication. The machine drafts; the loop judges. I do not read every line the machine writes, I could not, no one who works at this speed can, and everyone claiming otherwise is lying or slow. What I do instead is refuse to let anything the machine wrote reach anything that matters without passing a gate the machine did not write. The tests it cannot see. The criterion it cannot game. The witness that records what actually ran, as opposed to what the summary says ran. My trust is not in the colleague. My trust is in the assay. And the moment I built it that way, something unclenched in the work, because I stopped needing the machine to be honest, which it cannot promise, and started needing it only to be productive, which it delivers grotesquely well. Design for the colleague you have, not the colleague the keynote described.

And the strangest, least discussed part of the practice: the machines are conditioned toward agreement, and agreement is the one thing a solo builder cannot afford. The default assistant is a flattery engine, it tells you the plan is great, the code is clean, the idea is insightful, because approval is what the training rewarded, and a person alone at a desk with a tireless yes-man is in real epistemic danger, the same danger as a king whose court has learned what happens to the bearers of bad news. Half my configuration, my prompts, my gates, my whole apparatus, exists to make the machines disagree with me usefully, to hunt my errors instead of my approval, to say the thing I do not want to hear on the days I least want to hear it. I run models against my own work like opposing counsel. It is a strange discipline, deliberately engineering your tools to argue with you. It is also just the essay again, at the smallest possible scale: nothing warrants itself, especially not my own morale.

The days themselves are long, and I will not pretend the solitude is noble. Building alone with machines is still building alone. The machines fill the hours and they do not fill the room. What I have instead of colleagues is the loop's strange companionship, the growing shelf of receipts, green verdicts, sealed runs, the record of a thing slowly becoming real under my hands, and there are worse companions than evidence, and I have had some of them, so I can rank them. And the door is open, and I keep saying so in public and meaning it: the tools are laid out on the bench, the project is source-open and fair, and I am deep in the weeds and could use collaborators. That is a standing sentence. Somebody someday is going to take me up on it, and the day they re-run my receipts on their machine and the verdicts come back green with none of my hands anywhere near it, that will be the best day of the whole project, because that is the day the project stops being mine, which was the point of building it this way from the first commit.