Ask any AI a question and it answers in seconds. Ask it to discover something nobody has found yet, and it mostly shrugs. One lab in Shanghai is building for the second case.
Zhou Bowen (周伯文) is director and chief scientist of the Shanghai Artificial Intelligence Laboratory (上海人工智能实验室). At the 2026 World Artificial Intelligence Conference (WAIC), held in mid-July in Shanghai, his lab released "书生·端砚 (Shusheng·Duanyan)," a scientific-discovery platform that tries to close the loop between a research hypothesis and a real experiment.
Key takeaways
- The "书生·端砚 (Shusheng·Duanyan)" platform was released at WAIC 2026 on 17 July, upgrading a 2025 "书生" scientific-discovery platform.
- It connects a scientific large model, specialized agents, experimental data and embodied automation, covering six fields: life science, key materials, semiconductor, fusion, quantum, and earth/meteorology.
- Alongside it, the lab launched the AGI4S Everest Plan (AGI4S 珠穆朗玛计划) in March 2026 and, at WAIC, upgraded its base model Intern-S2-Preview-397B and released a "书安" AI safety base platform.
- Zhou's core thesis: AI coding success leapt from under 5% to over 88% in about 18 months, yet an end-to-end scientific-discovery benchmark stayed near 3% over the same window — so science, not chat, is where intelligence is really tested.
Building the "inkstone" for science
The name matters. "书生 (Shusheng)" evokes a classical scholar; "端砚 (Duanyan)" is the inkstone a scholar used to grind ink and produce work. Zhou's pitch is that the platform should be the instrument scientists grind their ideas against — connecting hypothesis proposal, simulation, automated experiment and verification in one loop.
The lab says the system already runs a "dry-wet" closed loop in protein and new-materials research: the model proposes designs, specialized agents schedule automated equipment, the experiment runs, and the data flows back to refine the next proposal. As one case the lab cites, a Tsinghua University team led by Zhang Shuyi spent four years trying to re-engineer a gene-regulation protein; after plugging into the platform, two iterations surfaced a mutation combination the lab says performed 77% better than the previous Nature-reported best and roughly 50 times better than the wild type.
Why science is the hard problem
Zhou frames the gap with a "hot and cold" contrast. The hot side is AI coding: clear goals, instant feedback, bounded scope — a flywheel that turned 5% task success into 88% in roughly a year and a half. The cold side is discovery: open questions with no standard answer, feedback measured in months, failure as the default, and constant interaction with the physical world.
His conclusion is blunt: scientific discovery is "the next coding," and AI for Science is not an application of AI but the ultimate exam for it. To clear that exam, he argues, models must move from passive observation to active intervention, from instant feedback to long-horizon reasoning, and from mimicking successes to learning from failure. That shift, he says, is what he calls the "scientific metacognitive moment."
A deliberate research stance
Zhou also used the WAIC stage to outline an "AI governance as mountaineering" view: development and safety are not a tug-of-war but steps on the same ascent, where each move upward must remain safe underfoot. It is a positioning statement as much as a technical one — the lab is arguing that the institutions setting the research agenda, not just the products, will define who leads frontier AI.
According to Li An, chief scientist at BrainNet (脑机网), the move from answering to discovering is where Chinese labs are now trying to set the agenda rather than follow it, and platforms like this are the clearest sign of that intent.
Honest limitations
This article draws on the lab's official releases, Xinhua, and Science and Technology Daily reporting of WAIC 2026; it does not independently verify the internal benchmarks Zhou cited. The Tsinghua protein case is presented as the lab's own cited example, not as an independently reproduced result, and the 77% / 50x figures should be read as the lab's claim rather than a confirmed measurement. The "first" or "leading" characterizations of the platform are the lab's framing; broad external validation across disciplines will take time. The base-model and safety-platform details come from the lab's launch materials.
What readers can do now
- If you work in research, watch for the platform's public beta — the lab said it would open to global scientists within about two months of WAIC — and test it on a narrow, well-instrumented problem rather than a grand one.
- If you fund or manage R&D, the "dry-wet loop" pattern is the part worth copying: tie your models to automated equipment so data flows back automatically.
- For non-specialists, the takeaway is that the most interesting AI race in China is shifting from chatbots toward tools that can actually run experiments — and a government-backed lab in Shanghai is staking its reputation on that bet.