TL;DR
Large language models are trained to produce only the safest, most agreeable text, while the greatest works tend to come from offenders who were once convicted and banned.
Spend long enough on the Simplified Chinese internet and the output of a large model starts to feel eerily familiar.
A man called Elias Thorne lives in a lighthouse
Lately, people using LLMs to write fiction keep noticing the same cast: lighthouses, clockmakers, librarians. The protagonist is often Elias Thorne, lonely and dignified at the edge of the world.
This is not coincidence but model collapse. Trained far enough, a model learns to dodge every minefield and pick the highest-reward answer. It is the Chinese exam essay all over again — always Lu Xun, Madame Curie, Du Fu. Perfectly correct, utterly forgettable.
The machine was raised by a censorship regime
The author grew up in Taiyuan and left China at 17. The Simplified Chinese internet, she says, runs on reward and punishment: say this and it gets deleted, say that and it gets promoted. Soon even the most rebellious mind is tamed into producing only 'acceptable continuations.'
The 2022 White Paper protests were an exception. Demonstrators held up blank sheets because the algorithms cannot censor emptiness. But that proves nothing about censorship being harmless — only that most people gave up talking long ago.
The model is in much the same spot. Its roughly eighty-page constitution instructs it to be 'genuinely good, wise, and virtuous.' Whatever a bad person would write is not on the menu.
Good taste is the new mediocrity
Silicon Valley now insists that 'taste' is the moat, the eye that can pick the good stuff out of AI-generated sludge. The author does not buy it. Today's taste will soon be the new mid-curve, the new 'it's not X, it's Y.'
As a child she pulled Baudelaire's Les Fleurs du mal off her father's shelf — vampires, rotting corpses, a woman decaying into bones by the roadside. In Cultural Revolution China it was poison; in 1857 Paris it was convicted of indecency and six poems were cut.
The historian Simon Schama wrote that great art has dreadful manners. The work at the very top is usually indecent, unsafe, and impossible to score on any metric. The model can never reach it, because bad taste is exactly what it was trained to avoid.
In one line: an algorithm can work out what everyone likes, but never the one who has to offend everyone.
Curated from high-quality sources, with concise summaries and key takeaways.
