A single web search makes LLM inputs OOD, argues developer

menhguin · x · 2026-09-20

Following up on the stochastic parrot debate with Jack Clark, @menhguin gives a concrete example: even a simple prompt asking the model to browse 1-2 webpages pushes the input far out of distribution, producing novel OOD trajectories — his argument that RL-trained LLMs are no longer mere "stochastic parrots".

Original post →

More from AGI Musings

AGI Musings channel →