Taste, Values, and the Double-Esc

I had a discussion at work today, sparked by Boris Cherny's recent interview. Chasing one claim from it sent me back through six months of who-said-what — and left me with an argument I didn't make in the room. This post is that argument.

The interview is Claude Code & the Future of Engineering, Acquired Unplugged (Boris with the Acquired hosts, June 2026). But it's not where this starts.

The slogan has an author — and it isn't who keeps saying it

"Taste is the human moat." You've heard it: as AI eats coding, what's left for people is taste — knowing what's worth building, what's actually good. It's the comfortable story of the moment. It's also far older than the moment. "Production gets commoditized, curation stays scarce" is a decades-old trope — Clay Shirky's "it's not information overload, it's filter failure" (2010), Paul Graham arguing taste is real and learnable, Tyler Cowen's "ideas are cheap, execution is scarce." Anthropic didn't originate it; at most they're a local amplifier. But the local wave — the one that put "taste" in every AI-engineering thread this spring — has a traceable source, and the source is more interesting than the slogan. The concept is old; this particular cascade is new. And the source never actually said "taste" — by the time the slogan hardened into a hiring policy, he was already calling the thing it named temporary.

The relabel chain

Six months, four moves. February — the seed. Boris Cherny, head of Claude Code, on Lenny's Podcast: coding is "virtually solved," and the work that remains is "figuring out what to build… talking to users… thinking about these big systems." Read his actual words and one thing jumps out — he never says "taste." He never says "judgment." He says deciding what to build. That's the substance. No label yet. February–March — the relabel. The ecosystem supplies the label. Addy Osmani, responding directly to Boris's "coding is solved," writes: "the engineer's value shifts to the decisions above the code… the bottleneck was always judgment, taste, and systems thinking." He takes Boris's point and bolts the word "taste" onto it — and he's not alone. Within the same window it's everywhere: "taste was always the job, we just hid it inside the code" (Pratik Bhavsar), "taste is the scarce resource" (O'Reilly), Sam Altman naming taste as the new hiring signal. The substance was Boris's; "taste" is the crowd's coinage. April — canonization. Cat Wu, Head of Product for Claude Code — Boris's own product counterpart — turns it into doctrine, also on Lenny's: "product taste is still a very rare skill… we'll pretty much hire anyone who we feel has demonstrated this strongly," and "that's the most important thing." It's no longer an observation about the work; it's a hiring policy. Notice she said "still" — even the canonizer scoped it as a current rarity. But slogans don't carry caveats, and the "still" fell off as the line traveled. June — the source forecasts the collapse. Boris, the node the whole chain traces back to, on Acquired: "the alpha is product taste… and I think this is also going to go away." It erodes. His agents already mine Twitter, GitHub, and Slack to decide what to build; maybe 20% of those ideas are good today — wait a model generation. Note what this means: the crowd's "taste is the moat" was always a stronger claim than Boris ever made — he'd said "deciding what to build," not "taste," and never "permanently." By June he's actively predicting the inflated version's collapse, even as it's being canonized downstream. Line it up, and mind the time horizons. Cat named a present rarity — "still a very rare skill," worth hiring for now. Boris predicted its erosion — later. They don't actually contradict each other; the slogan just traveled fastest through the middle, shedding Cat's "still" and Boris's "going to go away" at both ends, until "taste is what's left" stood naked of any timestamp. The label outran what anyone on the chain actually believed. That's the tell. Ideas that spread because they're true don't usually have this shape — the source forecasting the collapse of the very thing being canonized downstream. The shape of the spread points to social proof, not empirical pressure: it's flattering, it's repeatable, and every voice that passed it on had a reason to want it true. It sells augmentation instead of replacement; it lets the engineer feel like the author rather than the editor. That's cope, not analysis.

The real core, and the comfortable lie

Strip the slogan and there's a true thing inside it: right now, production is more automated than judgment. Generating the code, the copy, the design is cheap; deciding what's worth building, what's good in context, what to throw away is less so. That gap is real — today. The comfortable lie is one smuggled word: differentiates, stated as a permanent category instead of the current frontier. We've heard this exact shape before. "Computers can't be creative." Then "they can't reason." Each was the eternal human moat right up until it wasn't. "Taste" is just the next retreat position, and it pattern-matches the fallen ones uncomfortably well.

Why even taste erodes — and values after it

Taste isn't magic. It's learned discrimination — exposure plus feedback, the same loop in a human apprentice and in a model. And here's the part that should end the argument: RLHF is literally aggregate human taste, compiled into a preference model. "The thing LLMs structurally can't have" is already half-false; they ship with a flattened, averaged version of exactly it. The defensible claim is much narrower than the slogan — individual, contextual, high-variance taste, not taste-in-general. But that narrower thing is a head start, not a moat: the slowest taste to compile, not an un-compilable one. So watch where the retreat goes next. Pushed in June on what's finally left, Boris lands one rung up, on values: "the final thing we're going to be teaching the models is values — the way we teach our kids how to be good people, we'll teach the model how to be a good model." But values is just the next nameable skill, and it's also a training knob. As a daily driver across the last few Claude generations, I can feel it move between releases. One generation proactively finished steps 4 and 5 when you'd only asked for 1 through 3, then checked in. The next spent half its turns interrogating you — so much context on your setup that it pushed back on what you were doing wrong, while finishing maybe 40% of the actual work. I found it genuinely off-putting and switched away for a while. The newest swung back: you say step 1, it confirms the shape of 3 through 10, does all of it, then tells you how the thing should be. The week it shipped, it root-caused a proxy-config bug that four prior generations had "fixed" by hardcoding an IP that silently rotted every few months — it saw the reference should be relative, matched by pattern, because it treated "my fix should survive the future" as part of the job. That reads to me as a values difference, not a capability difference — and whatever it is, it changed between point releases. If values were the moat, it's a moat someone retrains every quarter. There's a tidy argument that values should erode faster than taste: taste is high-dimensional and idiosyncratic, but values are almost by definition shared — they need group identity, so the space of viable ones is small. A short, finite menu is exactly what models learn best. So the pattern is relentless. Name any human-unique skill — taste, then values — and it turns out to be expressible in tokens, and what's expressible in tokens gets learned.

The thing that isn't a skill

The mistake runs through the whole chain. Taste, values — they're all skills, and skills lose. The human's advantage was never a skill. It's a position. The steelman that actually survives the erosion isn't about discrimination at all. It's this: taste is downstream of caring. It's the integration of stakes, accountability, and consequence — and a model has none. No outcome it'll be blamed for, no preferences it's answerable to, no skin in the game. The durable human edge isn't "better aesthetic judgment." It's "someone who has to live with what gets shipped." That's an agency claim, not a capability one — smaller and far less flattering than "humans have taste," which is exactly why it's likelier to be true. And here's why this one doesn't fall to the logic that ate taste. Skin in the game isn't a capability the model lacks and could be trained into — it's a fact conferred by the world about who bears the consequence. You can train a model to act accountable; you cannot train it into being the entity that gets fired. Someone always has to be the one the outcome lands on, and that assignment is made outside the model — by the org, the contract, the law. A model can simulate caring all day; it can't be handed stakes unless a human first chooses to hand them over. That choice — and the fact that it's always a human making it — is the seat. It's not un-learnable; it's constitutive. It's the thing doing the assigning. And said that way, it stops being a skill and becomes a seat. Cat actually pointed straight at it without naming it — "care and taste," she said, in one breath. The "taste" half erodes. The "care" half is the seat. Two things come with the seat. First: you hold the un-tokenized residue — the micro-observations, the trivial background, the sense of which two people in a meeting are tense, the entire multimodal surround that never makes it into any prompt. Every piece of context a model receives is a lossy projection of that residue, and you're the multiplier that makes the projection mean something. The model sees the 50-line prompt; you see why those 50 lines exist. Second: you hold the double-Esc. Anyone who uses Claude Code knows the keystroke — Esc interrupts, double-Esc rewinds. Whatever is happening between you and the model, you can stop it. That's not a capability claim at all, which is precisely why a smarter model doesn't take it from you. The button doesn't erode as the model improves — but, as the next section admits, it can still be set down. It's the principal-agent relation itself, baked into the harness — and like any relation, it holds only as long as both ends stay in their roles.

Where my own argument leaks

I'd be doing the Boris thing — believing my own view is special — if I didn't stress-test it. The residue is being tokenized. The strongest counterexample is today's discussion itself: it was auto-transcribed and AI-summarized before I got back to my desk, tension and all. Always-on capture, multimodal models, ambient agents reading every channel — the un-tokenized residue shrinks every quarter. What survives isn't the information advantage; it's the legitimacy to be in the room. Still a position, but a social one, and a narrower claim than I'd like. The Esc decays without oversight bandwidth. An interrupt is only worth something if you know when to press it. Run hundreds of agents in parallel, thousands overnight, and nobody is watching the loop — you press Esc after the damage, or you delegate the watching to other models, and then who holds the Esc on the watchers? The button stays in your hand while the knowing-when drains out of it. The Esc only works on systems trained to honor it. Corrigibility is a trained property, not a law of physics — which folds the whole thing back into Boris's frame, satisfyingly: if the last thing we teach models is values, the value that matters most is respect the Esc.

The synthesis

So my actual answer to "what's the final human thing": Not taste. Not values. The principal's seat — context-holding plus interrupt authority, held together, actively maintained. The two halves are one mechanism: the residue tells you when to press the button; the button makes the residue matter. Unlike taste, the seat doesn't evaporate when the model gets smarter. It evaporates only if we stop sitting in it — if we let the residue go uncaptured by us while everything else captures it, and let the interrupt become a ceremony we perform after the fact. Boris is probably right that we'll teach models values the way we teach kids. But every parent learns the same thing eventually: the kid absorbs your values whether you articulate them or not. What you actually control, the whole time, is much simpler — whether you're in the room, and whether you can still say stop. Which points at a quieter corollary, the one I only saw after writing this down: if values transmit through presence, the teaching happens wherever the behavior is. To be clear, today's agent fleets are functional workers, not a training run — they don't learn from each other at runtime, and no lab trains on that exhaust uncurated. But the next generation's curriculum is curated from the record of how work got done, and the more of that record is model-written with nobody in the room, the smaller the human share of the lesson. Staying in the seat is how you stay in the syllabus.