Claude Speeds Up 30+ Biomolecular Models ~4x in Four Weeks; ScienceIDE Turns Scientific Code Into Agent Training Grounds
机器之心 · wechat · 2026-09-18
Anthropic published research showing Claude, supervised by two staff members without GPU kernel backgrounds, optimized 30+ open-source biomolecular models (AlphaFold3, Boltz-2, Chai-1, ESMFold2, RFdiffusion) in under four weeks — 4x average speedup with minor numerical tolerance, 2x for exact-match outputs. Code is open-sourced. The report also admits its limit test ran but produced incorrect predictions: passing isn't scientifically right.
Enter ScienceIDE, an open-source project from nonprofit AItonomy Foundation (sponsored by PhAILabs and Alibaba Qwen):
- Idea: turn scientific codebases into executable, verifiable agent-learnable environments; domain experts define test cases and acceptance criteria, validated interaction traces feed evaluation, SFT and RL.
- Scale: 64 environments from 27 codebases, 2,812 tasks, 1,076 executable scientific checks spanning magnetohydrodynamics, ocean climate, quantum, immunology and more.
- PLUTO experiment: AI's first GPU port gave <5x over a 128-core CPU node; another 1-2 days of joint tuning added 3x — one-shot generation leaves multiples of performance on the table.
- Benchmark: on 85 hard ScienceIDE-Hard tasks, the best model (Claude Fable 5.1) scored 67.1%; over half of frontier models stayed under 30%.
- Training value: SFT on scientific trajectories produced open PhAI-IDE-4B/9B/72B models that also improved on general benchmarks (e.g. 72B LiveCodeBench Execution 0.539→0.594), suggesting science as a training ground for intelligence itself. The team plans to wrap Anthropic's 36 open-sourced optimization packages into environments.
More from Research
- WASABI: Blei lab trains diffusion models to sample Boltzmann distributions with 3.3x fewer energy evaluations — skoularidou · 2026-11-02
- A complete math roadmap for machine learning: linear algebra, calculus, probability — TivadarDanka · 2026-11-02
- A base LLM plus 11 lines of code is all you need for a decision model — helloiamleonie · 2026-11-02
- Editor fully resigns from JMLR, arguing AI reviews beat rushed humans and journals need a rethink — roydanroy · 2026-11-02
- flowl: a probabilistic programming language built on optimal transport, not distributions — burny_tech · 2026-11-02
- Asynchronous Flow Map Language Models Generate All Tokens in a Single Forward Pass — kalyan_kpl · 2026-11-02