Nostalgebraist's Hydrogen Jukeboxes - by Scott Alexander

Astral Codex Ten

SubscribeSign in

Nostalgebraist's Hydrogen Jukeboxes<br>...<br>May 13, 2026

323

615<br>22

Share

In conclusion, the only good theory of taste is Nostalgebraist’s.<br>He wrote a post called Hydrogen Jukeboxes, analyzing the literary output of an AI called R1. This AI tried hard to write good fiction, which was part of the problem. It crammed its stories with what Nostalgebraist called (stealing a term from Ginsberg) the “eyeball kick” - a flashy stylistic move that immediately catches the reader’s attention and “wows” them. Here are examples - some from R1, others from an experimental OpenAI model trained specifically for fiction-writing:<br>“There is a prompt like a spell: write a story about AI and grief, and the rest of this is scaffolding—protagonists cut from whole cloth, emotions dyed and draped over sentences.”

“When the jar of Sam’s laughter shattered, Eli found the sound pooled on the floorboards like liquid amber, thick and slow. It had been their best summer, that laughter—ripe with fireflies and porch wine—now seeping into the cracks, fermenting.”

“And so I built a Mila and a Kai and a field of marigolds that never existed. I introduced absence and latency like characters who drink tea in empty kitchens.”

“The morning her shadow began unspooling from her feet, Clara found it coiled beneath the kitchen table like a serpent made of smoke.”

Nostalgebraist and another writer, Coagulopath, catalogue some of the most common AI eyeball kicks, each occurring across multiple LLM models:<br>“An overwhelming reliance on cliche. Everything is a shadow, an echo, a whisper, a void, a heartbeat, a pulse, a river, a flower—you see it spinning its Rolodex of 20-30 generic images and selecting one at random.”

“Conjunctions combining one thing that is abstract and/or incorporeal with another thing that is concrete and/or sensory.”

“Repetitive writing. Once you've seen about ten R1 samples you can recognize its style on sight. The way it italicises the last word of a sentence. Its endless "not thing x, but thing y" parallelisms…the way how, if you don't like a story, it's almost pointless reprompting it: you just get the same stuff again, smeared around your plate a bit.”

R1 is a small model - certainly today, but even by the standards of early 2025 when it was trained. We don’t know how big OpenAI’s experimental fictionbot was, but since Altman mentioned it once and never again, it probably didn’t receive too many company resources, either in terms of compute or human attention.<br>Both models were likely trained through RLHF; it tried various styles, minimally-trained humans rated which ones were good and bad, and then it auto-adjusted in favor of the good ones.<br>When you combine low mental capacity (= low ability to tolerate complex abstractions) with high pressure to perform well, you get something that learns a few cheap tricks that work well on untrained readers. This is what Nostalgebraist is complaining about. If you read “Her lips were the whispering echo of a granite conundrum”, then it sounds literary as hell for the 0.5 seconds it takes before you realize it’s meaningless. It’s also computationally cheap: memorize 20 - 30 words that sound really good (“echo”, “whisper”, heartbeat”), have a pseudoscript for generating pseudoprofound analogies (CONCRETE_OBJECT + ABSTRACT_OBJECT), have another pseudoscript that peppers them across your writing in semi-appropriate places, and you’re done. And R1’s eyeball kicks weren’t really as bad as all that. “The rest of this is scaffolding—protagonists cut from whole cloth, emotions dyed and draped over sentences” - basically achieves meaningfulness and self-consistency. There’s nothing wrong with it besides how it becomes grating when you see it every three sentences and realize what the AI is trying to pull.<br>(the infamous em-dash comes from the same imperative — common in formal literature, rare in low-IQ Reddit comments, it’s a cheap way to signal sophistication with a single character)<br>I’m Kenyan. I Don’t Write Like ChatGPT. ChatGPT Writes Like Me describes the struggle of non-English speakers who learned to write passable English in school and are now being told that they sound like AI.<br>ChatGPT, in its strange, disembodied, globally-sourced way, writes like me. Or, more accurately, it writes like the millions of us who were pushed through a very particular educational and societal pipeline, a pipeline deliberately designed to sandpaper away ambiguity, and forge our thoughts into a very specific, very formal, and very impressive shape.<br>The bedrock of my writing style was not programmed in Silicon Valley. It was forged in the high-pressure crucible of the Kenya Certificate of Primary Education, or KCPE. For my generation, and the ones that followed, the English Composition paper - and its Kiswahili equivalent, Insha - was not just a test; it was a rite of passage. It was one built up to be a...