AI for Science skeptic: better prediction isn't scientific discovery
KordingLab · x · 2026-11-16
A widely shared long post questions the hyped "AI for Science" narrative:
- Too low a bar: nearly anything now counts as AI for Science—train a network on scientific data, beat a baseline, improve accuracy by 2%, and it's framed as accelerating discovery.
- Core argument: a better predictor isn't a better understanding of nature. A model predicting molecular properties more accurately hasn't taught us new chemistry; a better fit to experimental data hasn't uncovered mechanism. Computational progress is being confused with scientific progress.
- Misaligned evaluation: much of the field is still judged like ordinary ML research—new architecture, better benchmark, higher accuracy, another paper—rather than by what was actually discovered.
The author calls for evaluating such work by scientific contribution, not ML metrics.
More from AGI Musings
- willcb's research hack: ask 'what would a next-gen superintelligent LLM agent do?' — willcb · 2026-11-16
- Dwarkesh: The math avalanche is coming for AI development, winning over an RSI skeptic — BethMayBarnes · 2026-11-16
- Ex-Google CEO Eric Schmidt tells students: use AI to scale your work, entry cost has never been lower — rohanpaul_ai · 2026-11-16
- Miles Brundage: 'whistleblowing' conflates three distinct acts — only one needs legal protection — Miles_Brundage · 2026-11-16
- Six months later: agents edit weights, Anthropic warns of near-RSI — mvandemar · 2026-11-16
- Tech workers delaying kids and homes over AI uncertainty; rez0 says don't wait — rez0__ · 2026-11-16