Hinton Pushes Back on "LLMs Just Predict the Next Word"

Geoffrey Hinton argues that calling LLMs "next-word predictors" fails when models handle genuinely novel inputs, citing a new programming language a model could understand. Since LLMs predict tokens, they can in principle understand any tokenizable structure, from language to video.

2026-11-19 ~ 2026-11-20 · 2 related posts