Intern·Duanyan: Closing the Loop in Scientific Research
新智元 · wechat · 2026-07-19
This article discusses a core dilemma in AI4Science: **AI generates hypotheses quickly, but real-world scientific validation is slow, lacks feedback, and makes it hard to retain negative data**. Using examples from materials science and protein engineering, the author notes that while AI excels at proposing solutions, it rarely gets second-level binary feedback like coding does, meaning fundamental research progresses much slower than software tasks. The piece highlights the **Intern·Duanyan** (书生·端砚) scientific discovery platform launched by the Shanghai AI Laboratory at WAIC2026. It emphasizes closing the "read-compute-do" loop: AI generates hypotheses, runs simulations, automated lab equipment executes physical experiments, and the measured data flows back to iterate the model. The author believes this dry-wet experiment loop and research asset retention mechanism can alleviate issues like solution bottlenecks, validation lag, and knowledge loss. The platform also structurally preserves formulas, parameters, code, and both successful and failed data for cross-project reuse. It has reportedly been deployed in protein engineering, quantum computing, brain science, new materials, semiconductors, and earth meteorology. The overarching thesis is that the key to AI-driven research isn't just a stronger model, but connecting models to real-world feedback to build a sustainably iterating scientific intelligence system.
More from Companies & People
- Insight Partners maps the crowded AI automation and agent stack — n_sri_laasya · 2026-07-21
- Companies are forming internal AI think tanks to evaluate tools and internal use cases — fguerino123 · 2026-07-21
- Auki pitches physical AI infrastructure for stores, warehouses and supply chains — broodsugar · 2026-07-21
- VibeLoft’s feedback loop could become China’s Product Hunt, with 3,000+ daily users — vista8 · 2026-07-21
- Judge approves Anthropic’s $1.5 billion copyright settlement, a U.S. record — Polymarket · 2026-07-21
- Chinese AI companies are still making real contributions, despite no o1-level breakthrough — JoshPurtell · 2026-07-21