Discovery at Scale: AI for Science bottlenecks are moving upstream, from compute to problem selection
bravo_abad · x · 2026-09-24
This week's Discovery at Scale argues the most convincing AI for Science work is shifting from replacing scientific computation to other parts of the loop.
- Upstream bottlenecks: Faster molecular dynamics doesn't tell you which configurations deserve simulation; better interatomic potentials don't automatically construct meaningful reaction paths. AI now decides what to compute, constructs problems before the solver sees them, and makes published methods easier to reuse.
- Gen-COMPAS (Tang et al., Nature): a diffusion model proposes plausible intermediate molecular configurations, with short MD trajectories determining what they actually do — avoiding brute-force waste on rare transitions.
- Weak predictors win too: an AUC 0.62 classifier cut 65,578 candidates to 145 DFT calculations and still yielded six discoveries.
- Also covers agents that turn papers into executable tools and method-reusability efforts.
More from Research
- Autonomous agent edits just 1,024 weights to lift Qwen3.5 0.8B from 7.4% to 37.4% on MATH-500 — VictorSanh · 2026-11-12
- Gait dataset labeled by encounter, spanning pre-op to 3 months post-TKA — zawthet · 2026-11-12
- Walking alignment measured 2.5-3.1° more neutral than X-ray, iPad gait study finds — zawthet · 2026-11-12
- MIT's Tegmark coins 'formal-washing': Signal protocol proof is far from complete — tegmark · 2026-11-12
- Signal formal verification hits 80%, Tegmark says Wired overstated completion — tegmark · 2026-11-12
- Category-theoretic framework asks: when does a physical system actually compute? — neurovium · 2026-11-12