brwilder: anthropomorphism alternatives fade as LLMs drift from pretraining

brwilder · x · 2026-09-05

AI researcher brwilder argues that alternative mental models of LLMs ("next token predictor", "cultural technology") have become progressively less useful for predicting behavior as models move beyond pretraining/SFT. He proposes treating LLMs as engineered systems and attributing behavior to parts of the training pipeline, while conceding anthropomorphic explanations can still be wrong since LLMs acquire beliefs/desires in non-human-like ways. Building real precision, he says, is a job for science.

Related event: Rethinking LLMs: Psychology Language Over 'Stochastic Parrots'(3 posts)→

Original post →

More from AGI Musings

AGI Musings channel →