Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Nostalgebraist's Hydrogen Jukeboxes - By Scott Alexander
TL;DR
Scott Alexander’s core claim is brutal and simple: taste is mostly nostalgia — what we call “bad taste” is usually just overused, easy-to-like tricks like R1’s “echo, whisper, heartbeat” prose that sophisticated audiences have grown tired of.
The AI fiction critique starts with the ‘eyeball kick’ — Nostalgebraist’s term for flashy, instantly impressive moves like “a serpent made of smoke” or “liquid amber,” which feel literary for about half a second before collapsing into cliché.
Scott’s explanation for AI slop is a capacity-and-pressure mismatch — small RLHF-trained models like R1, and likely OpenAI’s shelved fiction bot, learn cheap scripts that impress untrained readers because they can’t sustain deeper abstraction under evaluation pressure.
He makes the argument personal with a Kenyan essay anecdote — in ‘I’m Kenyan. I don’t write like ChatGPT. ChatGPT writes like me,’ formal exam-driven English training rewarded proverbs, ‘wow words,’ and polished sentence structures in ways that look eerily like AI style.
Toddlers become his laboratory for understanding taste — Lisa Frank art, ‘Choo Choo Train,’ and orange juice are all maximalist bundles of bright, sugary, evolution-friendly rewards, which makes ‘cheap tricks’ feel less like failure and more like the default way humans enjoy things.
His most provocative move is turning the anti-slop argument back on elites — if you ban every easy pleasure until only ‘10 geniuses’ can make art and ‘10 connoisseurs’ can enjoy it, you may end up defending a culture where almost everything is ugly to almost everyone.
The Breakdown
The AI fiction problem: too much style, not enough meaning
The video opens on Nostalgebraist’s essay about “Hydrogen Jukeboxes,” where AI fiction like R1 tries so hard to sound literary that it stuffs every paragraph with what he calls the “eyeball kick.” The examples are perfect specimens: shadows unspooling, laughter shattering, absence and latency drinking tea — lines that hit with a quick flash before you notice they’re mostly scaffolding.
How AI learned the same 20 to 30 fake-deep moves
Nostalgebraist and coagulopath’s diagnosis is that these systems overuse a tiny bag of tricks: clichés, abstract-plus-concrete pairings, repeated parallelisms, and flourishes like italicizing the last word. Scott’s gloss is sharp: when a low-capacity model is under heavy pressure to perform via RLHF, it learns the cheapest possible signals of sophistication — “echo,” “whisper,” “heartbeat,” and the infamous em dash — because they work on first contact.
The Kenya essay detour that makes the AI point land harder
Then comes a memorable pivot to the essay “I’m Kenyan. I don’t write like ChatGPT. ChatGPT writes like me.” The writer describes KCPE composition exams as a 40-minute life-defining sprint where students were drilled to open with proverbs, deploy big vocabulary, and sound maximally polished for graders chasing a score out of 40 — basically the human version of optimizing for reward under narrow formal constraints.
Bad taste as overused effectiveness
Scott connects that school-training story back to AI with a broader definition: bad taste is what happens when easy, high-yield techniques get overused until experienced audiences can’t stand them. “It’s not X, it’s Y” is his go-to example — a genuinely good rhetorical move in moderation that AI turns into a script and repeats until your eyes bleed.
Toddlers, Lisa Frank, and the economics of cheap delight
The essay gets funnier and more grounded when Scott says he lives with “the dumbest and most gullible audience of all: small toddlers.” Lisa Frank art becomes his case study in maximal cheap tricks — rainbow colors, sparkles, smiling animals, flowers, busy scenes — all bundled together with zero restraint because each element is an easy button for positive feeling.
Choo Choo Train, orange juice, and what sophistication strips away
He extends the analogy to music and food: his daughter loves “Choo Choo Train,” his twins love “If You’re Happy and You Know It,” and their favorite drink is basically sugar with orange-adjacent chemicals. Against that, he jokes, sophisticated adults champion melody-less tonal music and molecular-foam cuisine that “tastes slightly like the color chartreuse,” as if refinement often means systematically removing the things ordinary people obviously enjoy.
The real thesis: good taste may just be exhausted appetite
From there Scott states the big idea outright: poor taste overuses the simplest innate pleasures, while good taste deliberately avoids those “blaring claxons” so attention can settle on subtler patterns. But he immediately complicates it with poetry, architecture, and modernism, admitting that ornate dragons and decorative roofs still work on him — which raises the uncomfortable possibility that taste is partly just what you haven’t become bored by yet.
The elitism problem at the end
The closing stretch is the most provocative: if we keep banning cheap tricks, then medium-cost tricks, we end with art only a handful of geniuses can make and only a handful of connoisseurs can enjoy. Scott pushes that logic to absurdity with a superintelligence generating “1 million schematics for reimagined houses” and asks why we should accept a worldview where nearly everything becomes “ugly to everyone,” especially when his daughter seems to get more joy from “Choo Choo Train” than most tasteful adults get from high art.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
