"Train a frontier model that's actually good at writing" — researchers slam LLM prose
MichaelTontchev · x · 2026-10-17
Researcher Michael Tontchev calls for someone to "train a frontier LLM that is actually good at LANGUAGE—specifically writing," arguing current flagship models produce horrible prose and "seem to be getting worse with every version."
The thread drew in Anshul Kundaje, who said writing an abstract from a manuscript takes 5+ iterations, with models only decent at compression (500 words → 150 words). The core point: frontier models race ahead on code and reasoning while natural-language writing quality stagnates or regresses.