LLMs win gold at IMO but fail in medicine; data is key
balazskegl · x · 2026-08-18
LLMs have achieved gold medals at the IMO by verifying symbol strings but cannot determine treatment efficacy. Science lacks a mechanical verifier because its object is the world, requiring inductive and abductive reasoning on causal graphs—structures text underdetermines. Web-scale pretraining predicts text, not causal models.
Healthcare offers the missing signal: pairing human-curated concepts (pathologies, symptoms) with free text and intervention outcomes. This pairing can ground language in a causal model of the world. The author joined @Doctolib to build AI that helps clinicians and solve this grounding problem, questioning if major discoveries will come from LLMs with the right data or require a new architecture.
Related event: Ex-Kaggle CTO: LLMs Win IMO Gold but Can't Judge Medical Efficacy(2 posts)→
More from AGI Musings
- Revisiting Vernor Vinge's 1993 essay: superhuman intelligence within 30 years ends the human era — khademinori · 2026-08-18
- Every AI System is a Division of Labor Between Model, Tools, and Human — coallaoh · 2026-08-18
- Software Augments Model, Humans Define Tasks and Decide — coallaoh · 2026-08-18
- AI Division of Labour Varies by Domain: Coding vs Clinical — coallaoh · 2026-08-18
- Neo-PE AI plays fail due to lack of worker dignity — curious_vii · 2026-08-18
- Models beat humans on narrow tasks; data is the new AGI bottleneck — haider1 · 2026-08-18