Self-Supervised Models Are the Wrong Paradigm for Scientific Design, Says Stanford Prof
anshulkundaje · x · 2026-11-18
Stanford computational biology professor Anshul Kundaje argues self-supervised models are the wrong paradigm for design tasks: without supervision or conditioning on assay/cell context/task specs, neither sequence-only models nor general LLMs will deliver. They can act as efficient samplers or pretrained representations, but always need conditioning or an oracle to reach the desired objective.
More from AGI Musings
- Horace Dediu: railroads, telecom, internet — will AI infrastructure repeat the cycle? — asymco · 2026-11-18
- Podcast: Horace Dediu on Apple, the economics of AI and consumer tech's future — asymco · 2026-11-18
- Stanford's Anshul Kundaje: self-supervised models are wrong paradigm for bio design — anshulkundaje · 2026-11-18
- AI safety researcher Haydn Belfield quips on deterrence as 3 Body Problem returns Jan 14 — HaydnBelfield · 2026-11-18
- Alljoined Builds Non-invasive BCI to Keep AI Aligned With Human Minds — rishabh16_ · 2026-11-18
- Gary Marcus: AI-linked cyberattacks are hitting US infrastructure, Optimizely breach hits millions — GaryMarcus · 2026-11-18