Stanford's Anshul Kundaje: models are soberingly bad at data-poor biology, stop extrapolating from scaling
anshulkundaje · x · 2026-09-09
Stanford professor Anshul Kundaje pushed back on blanket extrapolations about model capabilities: even today there are clear limits set by the nature of available training data.
His example: hard biology problems lacking necessary data, where these models are "soberingly" underwhelming. He concedes new capabilities may emerge in some domains within years and can't judge whether Terence Tao is right, but argues that extrapolating in all directions without any map of how is fortune-telling, not science.
More from AGI Musings
- AI safety testing itself may be the biggest risk? Debate on lab experiments — dreamwieber · 2026-09-09
- Dev argues AGI may come from 500K copies of OK intelligence, not one god model — mertdumenci · 2026-09-09
- tszzl wraps up the debate: grabby aliens and a behind-us great filter — tszzl · 2026-09-09
- Scobleizer on why AI existential risk doesn't scare him: nukes, cults, and billionaire incentives — Scobleizer · 2026-09-09
- Chernobyl was a safety test: the biggest AI risk may be testing AI for safety itself — dreamwieber · 2026-09-09
- Watching nerdy online communities become world-shaping forces: a reflection on Yudkowsky as prophet — dioscuri · 2026-09-09