ChatGPT Fails Basic Physics Intuition Test That Video-Trained V-JEPA Solves Easily
A head-to-head comparison shows ChatGPT failing a simple physical reasoning task that Meta's V-JEPA — trained on video, not text — handles correctly, reigniting debate over whether LLMs can ever develop genuine world models.
A demonstration shared by @heynavtoor shows ChatGPT failing a basic physics intuition test — reportedly involving predicting the behavior of physical objects — while Meta's V-JEPA model, which is trained on video data rather than text, succeeds. The comparison is a clean illustration of a fundamental limitation that LLM skeptics have long flagged: language models learn statistical patterns in text, not causal models of how the physical world works.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.