The Taste.

The lab has spent twenty-one essays building the generators of intelligence: the Corpus the model is made of, the memory that carries it, the substrate it runs on, the Spark of connecting existing dots, the Frontier of creating new ones. Every one of those is a generator. Not one is the selector.

Who decides what is worth preserving, what is profound and what is cheap, what is powerful but ugly, what is correct but soulless, what is efficient but not worth living in, what is possible but should remain unborn? That is taste. Not taste as luxury — taste as the highest-level judgment there is, the selection pressure on everything intelligence generates. And the wager of this essay is that as generation collapses toward free, taste becomes the scarcest and most load-bearing thing in the world.

There is exactly one process known to physics that builds genuine novelty out of a world that contains none: variation plus selection. Darwin found it in 1859. Donald Campbell generalized it in 1960 to all knowledge growth — “blind variation and selective retention.” The crucial word is blind: the generator cannot know which trial is correct, because if it did you would not need selection at all. Which means the information — the value, the fit — does not enter at the generating step. It enters at the selecting step. Dennett put it exactly: natural selection is “design out of chaos without the aid of Mind.”

Now apply it to intelligence. Spark named reward #1 — connecting existing dots. Frontier named reward #2 — creating new ones. Both are generators; they produce possibilities. But there are astronomically more connectable and creatable dots than are worth having, and The Seat proved the brutal corollary: a connection can fully seat — click, lock, feel complete — and still be worthless. “The click certifies that a structure closed, not that it closed on the truth.” Most connections are cheap. Taste is the selection pressure over that space. Generation is the variation. Taste is the selection — the half we never gave its own essay.

The structure of the last three years of machine learning is unambiguous: generation is becoming free, and the scarce input is the objective. In 2022, InstructGPT delivered the cleanest proof in the field — a 1.3-billion-parameter model aligned to human preference is preferred over the 175-billion-parameter GPT-3. Scaling the generator a hundredfold is worth less than pointing it at the right objective. Selection beats scale, measured.

Then the verifier systems. FunSearch beat the best known human result on the cap-set problem with a deliberately shallow LLM and an external evaluator. AlphaProof reached silver-medal Olympiad performance because the Lean proof-checker is the objective. AlphaEvolve broke Strassen’s 56-year matrix-multiplication record and its authors said the quiet part out loud: “the new bottleneck is evaluator design.” The generation is free; the verifier is scarce. This is exactly what Frontier claimed about reward #2 — “the verifier is what does the inventing; humans are the source of new verifiers.” Taste is the verifier: the human-supplied objective the whole machine optimizes against. And the formal capstone — the No Free Lunch theorem (1997): averaged over all objectives, every optimizer is identical. There is no objective-free optimization. Generation is mathematically empty without taste.

Kant (1790) gives the mechanism. A determinate judgment applies a given rule to a case — exactly how data-driven inference works. A judgment of taste is reflective: it has the particular and must find the universal, with no rule to apply. His formula — “the beautiful is that which pleases universally without a concept” — names the strange thing taste does: it claims the force of a rule while having no rule behind it. And he doubles it: genius, the generator, is also rule-less — “the talent that gives the rule to art.”

Frank Sibley (1959) is the cleanest statement, and the essay’s epigraph: “there are no sufficient conditions, no non-aesthetic features such that the presence of some set of them will warrant the application of an aesthetic term.” No feature-vector entails the verdict — “cannot be derived from data,” written in 1959. Mothersill hands us the puzzle: there are truths of taste and no laws of taste, because beauty is judged of this particular. And Schmitt, held at arm’s length, gives the extreme: the genuine decision cannot be deduced from any norm — it is sovereign precisely where the rule runs out. Hume keeps it from collapsing into relativism: the standard is the joint verdict of qualified critics, converging over time.

Taste is empirically real. Klein’s firefighters and nurses make excellent judgments without comparing options — they recognize a prototype and the right action arrives. Simon’s deflation: “intuition is nothing more than recognition.” That is the strongest pro-machine voice — if taste is compressed experience, a big enough pattern-matcher might absorb it.

But taste is conditional. Kahneman and Klein (2009): intuitive expertise is trustworthy only in a high-validity, regular environment with rapid, reliable feedback. Where those hold — chess, firefighting, wine, attribution, code review — taste is earned. Where they do not, confident “taste” is illusion and a simple algorithm wins; Tetlock’s confident storyteller is usually wrong. And taste is fallible, both ways: on the Getty kouros, the connoisseur’s eye beat fourteen months of lab science in two seconds — yet the same faculty let the world’s top Vermeer scholar certify a van Meegeren forgery, precisely because it confirmed his own prior. Real, reliable under conditions, not infallible. That combination is what “cannot be fully derived from data” honestly means — and it sharpens the genuine open question: is taste compressed experience a large model absorbs (Simon), or an embodied, social capacity that resists capture (Dreyfus: at the expert level there are no rules being followed at all)?

We already built machines that hold taste. There is even neural warrant for a learnable scalar: Zeki’s field A1 in the orbitofrontal cortex fires parametrically for beauty across vision, music, and mathematics. And it is deployed — the LAION-Aesthetics predictor, one CLIP-based scalar trained on a few thousand ratings, chose which billion images trained Stable Diffusion.

But watch how it fails, because the failure is the essay. These predictors reward processing fluency — the easily-processed, the symmetric, the prototypical — and collapse toward the homogenized “AI look.” The same flattening shows wherever generation outruns selection: Doshi and Hauser (2024) found AI makes individual stories more creative while making the corpus 10.7% more similar — the tails collapse. Shumailov (2024) proved the limit case, model collapse: train generation on its own output and “the tails of the distribution disappear,” irreversibly. Generation cannot bootstrap its own quality. The parts models miss are the ones that matter most: the fragile liking under the robust wanting; the expert judgment that lives below the rating; and the Knightian value-call over genuine novelty — uncertainty with no probability distribution, out-of-distribution by construction. The boss decides under Knightian uncertainty; the slave optimizes under measurable risk.

You feel like the boss some hours and the slave others, and you experience it as instability. It is not. It is how taste stays alive. You descend into the slave-work — make, execute, struggle, ground — to earn the calibration; then you ascend to boss — judge, select, refuse — to spend it. A boss who never descends goes blind. This is The Duty Cycle re-cut: not exploit/explore but execute/decide, slave/boss — and the same law holds, that a mind jammed in one stroke self-destructs.

And there is the second slavery, which is not about AI at all: being the slave to someone else’s taste — the metric, the platform, the request you did not author. That is the loss of sovereignty over selection. The whole future skill is the move From Navigators to Authors wrote at civilizational scale, now at the scale of a single working life: migrating up the stack from executor-of-others’-objectives to author-of-your-own. That migration is owning your taste.

The deepest form of taste is the one almost no one writes about: the refusal. Generation has no native stop — a generator generates; that is all it does. Taste is the only faculty that can look at something technically achievable, data-backed, and defensible, and decide it should not exist. Acemoglu’s “so-so technologies” give it an economic form; Jonas’s heuristic of fear a moral one; Bill Joy’s relinquishment the hardest one; Postman’s prior question — “for what problem is this a solution?” — the cultural one; Taleb’s via negativa an epistemology; and Chesterton’s fence the oldest rule: do not remove what you do not yet understand.

The single internal critic the essay must carry — so negative taste does not collapse into reflexive technophobia — is Sunstein: a blanket precautionary principle is incoherent, because risks exist on all sides. That is the point. Negative taste is judgment, not a rule. It is precisely the faculty that knows when refusal is wisdom and when it is mere fear — and, like all taste, it cannot be reduced to a procedure, because if it could it would not need taste.

The strong version is true. Across evolution, optimization theory, RL from preference, and the verifier systems, value enters at the selection step, and selection is a human-defined objective the generator cannot supply for itself. No Free Lunch makes it formal; InstructGPT makes it measurable; AlphaEvolve makes it the lived experience of the frontier labs. As the generators approach free, taste becomes the scarcest input in the system.

The weak version is false. Taste is not magic, not unconditioned, not infallible, not beyond all modeling. It is real under conditions, partly modelable, reliably fallible. Anyone who claims taste is a permanent untouchable human essence makes the symmetric overreach to those who claim it is just data. It is a grounded, conditional, correctable faculty whose successes and failures both live below the level of any stated rule — which is why it cannot be fully derived, and why it can be partly automated, at the same time.

And the deepest reading is the one the boss/slave image carries: the all-slave-no-boss future is the Master stagnating, the Emissary usurping, conception severed from the execution that grounded it. The human future is not to ascend permanently to pure judgment. It is to run the oscillation — to keep descending into the forming so the judging stays honest — and to refuse the surrender of your selection function to someone else’s metric. The machine wins generation. The skill is owning what you select, refusing what should not exist, and never commanding from so high that you forget how the thing is made.