Scaling fuzzy n-grams works as well as LLMs did years ago, researcher argues

RexDouglass · x · 2026-09-25

Rex Douglass argues that a perfectly good LLM metaphor has landed in a new social context of people wanting to feel they understand SOTA better than others, without better scientific understanding — just new memes for identity clubs. He adds that prior work scaling fuzzy n-grams to webscale corpora works surprisingly well: intentionally stochastic parrots can do much of what commercial LLMs were praised for a few years ago.

Related event: Researcher Criticizes LLM Community for Identity-Driven Metaphors Over Mechanistic Inquiry(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →