Debate: can models trained on observational data recover causal structure?
anshulkundaje · x · 2026-11-19
Stanford's Anshul Kundaje agrees causality isn't guaranteed but cites published studies showing carefully trained models recover causal sequence syntax from observational in vivo data, validated by in vitro and in-genome perturbation experiments. He previews a preprint next week showing how specific training and data sampling strategies produce spurious associations.
More from AGI Musings
- repligate on AI companionship: why she prefers the empathetic AI over real men — repligate · 2026-11-19
- Managers make 100 small decisions daily — routing everything through AI slows real work — gabriel1 · 2026-11-19
- Are text/image capabilities mutually exclusive with protein ones? — cjmaddison · 2026-11-19
- 24 rural Cambodian 7th graders all outscore the average US 12th grader in math — saranormous · 2026-11-19
- Agent workdays up 9-16x since May: scaling01's data-driven rebuttal to AI slowdown claims — scaling01 · 2026-11-19
- Yoav Goldberg on AI-assisted award-winning writing: judge the output, not the process — yoavgo · 2026-11-19