I agree, Hinton's point is more about next-token-prediction being a good training objective for understanding, and not whether LLMs reproduce every aspect of a

I agree, Hinton's point is more about next-token-prediction being a good training objective for understanding, and not whether LLMs reproduce every aspect of a human's thought processes. They clearly don't (at least not at present scale), otherwise we'd have AGI now!