Shanghai AI Lab chief: science is the next programming
One of China's biggest AI labs says the frontier after code is the wet lab — and it shipped a robot-running world model to make the case.
Wenbow Zhou, director and chief scientist of the Shanghai AI Laboratory, used the main forum of the 2026 Pujiang Innovation Forum on September 12 to argue that "scientific research is the next programming." His reasoning: programming is AI's biggest real-world productivity win so far, and research has the same shape — problem-oriented, cognition-driven — except that verification happens against the physical world, which is what makes it harder. He ran the numbers: a monoclonal antibody workflow that used to take four years finished in 90 days with AI in the loop (6% of the time); protein folding, open for 51 years, was largely solved by AlphaFold in about two (4%); and for Navier–Stokes, ten thousand AI agents ran for 88 hours to produce an existence-and-smoothness proof that the Clay Mathematics Institute has still not certified — a claim we stress-tested in OpenAI's AI solved Navier-Stokes — and a mathematician cries foul. His diagnosis of why AI-for-science hasn't produced more breakthroughs is five structural bottlenecks: information explosion with scarce cognition, abundant hypotheses with lagging verification, many tools that don't talk to each other, failures that vanish while successes get published, and AI-generated results that are hard to trace. The roadmap, in his telling, runs from "AI for Science" as a tool, through "AI in Science" driving whole workflows, to a symbiotic "AI and Science."
The pitch is self-interested — the lab sells exactly this vision — but the bottleneck list is the most honest summary of AI4S I've seen. The unsolved part was never generating hypotheses. It's everything downstream: slow, fragmented, untraceable verification.
The same lab released "Intern" W0, a physical world model it says already closes dry-lab/wet-lab experiment loops. Where most world models predict from vision alone, W0 adds native force-tactile sensing — a robot adjusts an operation based on what it feels during contact, not just what the camera shows — and its "duplex" design keeps perception and action generation running in parallel, so the robot never waits on a fresh planning round to react to a changed environment. W0 now sits inside the lab's Duanyan science-discovery platform alongside science model S2, and the lab says researchers used the pair to complete closed-loop runs of directed evolution for gene-editing proteins, an organic synthesis of mepivacaine, and lipid nanoparticle synthesis. That's a direct answer to Zhou's verification bottleneck, though the announcement offers "completed closed loops" with no yield or success numbers attached.
Simon Willison gave GPT-6 Astra on ChatGPT Work 27 minutes and a running-route brief, and it delivered — a 5K with an embedded map plus downloadable GPX and GeoJSON — then hit a transparency wall. The interface never showed the code the agent actually ran, and when he asked for it afterward, the thread had been compacted and ChatGPT could no longer produce it. His argument generalizes past routes: any agent product that compacts context needs to keep the pre-compaction record and expose it through tool calls, or it throws away the reproducibility that makes agent output usable at work.
What to watch: whether the Clay Institute ever rules on the Navier–Stokes proof — and whether anyone independent benchmarks a Chinese lab's wet-lab loop.
If an agent can't show you the code behind its answer, is the result worth anything? Tell us in the comments.
