Claude optimizes 30+ biomolecular models in under 4 weeks; ScienceIDE opens agent-learnable science code
jiqizhixin · x · 2026-09-26
Anthropic showed Claude optimizing 30+ open-source biomolecular models in under four weeks, supervised by only two Anthropic engineers — both biologists with no traditional inference-optimization or GPU-kernel background; the code optimization was done by Claude itself.
- This answers "can AI do it?" but raises new questions: how close is AI-written scientific code to fully utilizing hardware, and how to make this capability learnable, reusable and transferable across fields?
- Toward that, AItonomy Foundation (sponsored by PhAI Labs) launched its first open-source project, ScienceIDE, which turns the world's scientific codebase into agent-learnable environments — unlike Claude's one-shot optimization report, it lets agents continuously learn from scientific code and accumulate transferable optimization experience.
More from coding & agent
- Fabric-RLM: Open-source Recursive Language Model harness brings MIT's RLM to Microsoft Fabric notebooks — PawarBI · 2026-09-26
- Experiment: feeding thoughts to a NousResearch agent all day, no laptop needed — CurieuxExplorer · 2026-09-26
- Chaining GPT Images 2.5, Grok and Opus 5.5 to Produce a Full Dance Video Workflow — i_dg23 · 2026-09-26
- Versioning multi-file agent outputs without Git: immutable folders plus manifests — RocketSeven · 2026-09-26
- Ex-Meta's Christoph Nakazawa on the Engineering Values That Still Matter in the AI Era — cnakazawa · 2026-09-26
- Ghostty's new fuzzy path insertion lands in nightly — Mitchell Hashimoto can't work without it — generativist · 2026-09-26