Oxford paper argues LLMs are mathematically incapable of true invention

blaizedsouza · x · 2026-09-19

A widely shared thread summarizes an Oxford paper claiming LLMs can never truly invent anything, on mathematical grounds: a next-token predictor can never believe something the existing data says is wrong, while every historic breakthrough began with exactly such a belief — as when all prediction models in 1903 put human flight millions of years away, nine weeks before the Wright brothers flew. The gap, the author argues, is that LLMs mirror the past while humans reason toward a future that doesn't yet exist. Note this is secondhand paraphrase; the paper's actual rigor is unverified.

Related event: Oxford Scholars Argue LLMs Mathematically Cannot Truly Invent(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →