Stanford's James Zou: language may be <1% of intelligence, so LLMs won't reach AGI
james_y_zou · x · 2026-09-17
Stanford AI for Science Lab head James Zou argues that text-only LLMs likely won't reach AGI because language captures a tiny fraction of the biological signals underlying thought — every word follows tens of thousands of physiological signals that LLMs never see. His lab's alternative direction: training on physiological signals. Projects include Virtual Lab (cloning his team into autonomous AI agents, published in Nature), Paper2Agent (turning papers into interactive virtual authors), and SleepFM (one night of sleep data predicting risk for 130 diseases). His core claim: the current text pretraining paradigm captures only a small part of intelligence.
More from AGI Musings
- Metzger: If Someone Gives You an AI Extinction Probability, Ask for the Math — nptacek · 2026-09-17
- With Tools Shipping So Fast, Jayesh Urges People to Reset Their Sense of What's Possible — stuffyokodraws · 2026-09-17
- So8res Adds One-Year Retrospective to 'If Anyone Builds It, Everyone Dies' E-book — connoraxiotes · 2026-09-17
- A Year After 'If Anyone Builds It, Everyone Dies', Authors Give Away 1,000 Free E-books — connoraxiotes · 2026-09-17
- 'This will kill everyone, give us more power': blogger mocks recurring AI doom rhetoric from the world's most powerful — kevinnbass · 2026-09-17
- Podcast: TheZhi and Robert Wright debate whether the AI 'slowdown' is real — TheZvi · 2026-09-17