Closed-loop AI scientists: researchers map the path to autonomous biomedical discovery
marinkazitnik · x · 2026-09-10
A new preprint led by Ada Fang with Ayush Noori and Marinka Zitnik asks what it would take for AI to make Navier-Stokes-level breakthroughs in life sciences. Their core argument: math and programming allow cheap, massive-scale evaluation of AI-generated candidates, but in biomedicine the loop of hypothesis → experiment → revised hypothesis remains the central bottleneck for AI-driven discovery.
Their vision for autonomous AI scientists:
- Reason over scientific knowledge, multimodal data, competing hypotheses, and uncertainty across multi-loop campaigns lasting days to months;
- Learn under sparse and delayed feedback, combining soft verifiers (simulations, predictive models, biological world models) with hard verifiers (wet-lab assays, robotic experiments, organoids, clinical studies);
- Allocate limited experimental budgets to the highest-value tests.
The authors note several AI scientists have already produced findings labs later confirmed — the challenge is generalizing that closed-loop validation.
More from AGI Musings
- Philosopher: I thought philosophers were like LLMs — now I'm more optimistic about LLMs — FlorianGallwitz · 2026-09-10
- DeepMind's Nando de Freitas amplifies 'if you don't own your model, the model owns you' — GregoryDiamos · 2026-09-10
- Photographer: when AI makes things up about you, you can't prove a negative — "endlessly bleak" — zemotion · 2026-09-10
- David Patterson predicts robots and AI will 10X economic output by 2035 via UHI — davidpattersonx · 2026-09-10
- Kindergarten teachers push back: relational jobs in the AI age aren't meaningless — binarybits · 2026-09-10
- Irving: emergency safety patching pressure crowds out superintelligence-proof alignment work — geoffreyirving · 2026-09-10