New Chat Sign up
提问NewsTech, finance and policy — daily must-reads
简体繁體EN

AI Moves Into the Lab and the Workflow: From Autonomous Experiments to Agent Deployment

AIDailyArchive →

Highlights

  • Shanghai Jiao Tong University's AutoDNA appears in Nature; AI autonomously performs molecular experiments
  • Tencent showcases a full Agent lineup; Yao Shunyu says the moat lies at the product layer
  • AMD data center revenue nearly doubles; China Mobile Cloud teams up with Zhipu on domestic compute

AI-native lab: from bolt-on to full experimental workflow

A team led by Chen Haibo and Wang Zhaoguo at the IPADS Lab of Shanghai Jiao Tong University's School of Computer Science, working with Fan Chunhai and Wang Fei of the School of Chemistry and Chemical Engineering and the Frontiers Science Center for Transformative Molecules, has built AutoDNA (Siyuan-1), an AI-native science lab that performs autonomous molecular biology experiments. The results were published in Nature and featured as the lead story on the journal's homepage.

The key to the system is not "plugging in a large model" but the architecture. Existing AI autonomous science labs, according to the source material, typically have experts pre-configure instruments, prescribe workflows and set optimization rules, with AI connected only to isolated steps such as literature search, protocol generation or data analysis — a "bolt-on" architecture. AutoDNA instead uses co-design across model, experiment and instrument: multi-agent systems break complex tasks into literature search, reagent selection, protocol design, code generation, hardware execution, results analysis and iterative optimization; instrument capabilities are distilled into "atomic services" that models can understand and invoke; and micron-level positioning and microliter liquid-handling hardware ensure precision. The platform also integrates a control terminal built on OpenHarmony and has been validated in scenarios including nucleic acid testing, enzymatic DNA synthesis and cancer detection, demonstrating an end-to-end DNA storage workflow of encoding, writing, preservation, reading and decoding. (Source: ITHome)

The significance of this line of work: the bottleneck for AI in scientific experiments is shifting from "can the model think" to "can the system reliably act." Whether AutoDNA can be reproduced by other labs and extended to more experiment types is more worth tracking than any single result.

Agent deployment: applications are just the entry point; the gaps are in data and engineering

Tencent's Agent lineup is more complete than the WorkBuddy that outsiders know. Dowson Tong, senior executive vice president of Tencent and CEO of the Cloud and Smart Industries Group, told the Global Digital Ecosystem Summit that the biggest obstacle to enterprise Agent deployment is not that models are insufficiently powerful but that there are "too many breakpoints": AI can write proposals and do analysis, yet cannot connect to business systems or access enterprise data. On that judgment, Tencent proposed building an Agent productivity system that is "full-scenario, high-efficiency and more reassuring," with more than 40 launches over two days spanning applications, data, security and engineering.

At the application layer, Tencent launched Buddy One, a cross-application collaboration system that uses WorkBuddy to connect Tencent Docs, Tencent Drive, Tencent Lexiang and design and development tools, while bringing Agents into business execution scenarios through Tencent Cloud ADP, Yuanbao search, the Tianlai voice model and MAGIC AI. But the source material stresses that applications are only the entry point; for Agents to enter the enterprise, three basic questions must be answered: can they read and make good use of enterprise data, can they act within clear security boundaries, and can they run reliably at scale. (Source: Leiphone)

At the same time, Tencent chief AI scientist Yao Shunyu offered a more direct commercial judgment in a conversation at the 2026 Future Science Prize: AI moats lie more at the product layer than the model layer. Citing WeChat, he argued that what is truly hard to replace is an AI product that understands user context, leverages proprietary data and context, and earns user trust — not simply "the smartest model." Zhang Ya-qin joked that by this logic Anthropic might not be worth $1 trillion or even $2 trillion in the future, because models may one day become commoditized. (Source: ITHome)

Read together, Tencent's Agent narrative and Yao's moat judgment follow the same logic: the focus of competition is shifting from model parameters to workflows, data and trust. That judgment can be overturned by later facts — if the capability gap between models widens again and directly determines product experience, the case for a product-layer moat weakens.

Compute and industry: data centers first, domestic compute alliance

Demand for compute is still rising. AMD CEO Lisa Su, in an interview with Yahoo Finance, described the company as "data center first," just days after its fiscal 2026 third-quarter results showed data center revenue nearly doubling from a year earlier. She said AI adoption is outpacing many companies' expectations, and firms that once set investment plans annually are adjusting budgets more frequently to meet compute demand; AMD is responding by raising supply capacity and working with customers on larger-scale system deployments. (Source: ITHome)

Domestically, China Mobile Cloud and Zhipu Huazhang signed a cooperation agreement on Nov 6 to deepen collaboration in Token technology innovation, model services, product development and industry solutions. The two will focus on adapting domestic compute and tuning models, jointly build a Token development center to improve Token production efficiency, and connect Zhipu's GLM series models to China Mobile Cloud's MoMA model service platform. They also plan to jointly build an intelligent product matrix, co-create AI Coding integrated products, and deepen work in key industries such as finance and government affairs to build integrated "compute + model + application" solutions. (Source: ITHome)

Separately, the open-source physical intelligence platform Hitch Open was launched at the China International Import Expo. A team led by Professor Gao Yue of Shanghai Jiao Tong University and Shanghai Innovation Institute, working with the platform, completed 142 consecutive rallies of AI autonomous-decision robot table tennis, setting a world record for the project and providing an extreme test scenario for the "perception-decision-action" loop of embodied intelligence. (Source: ITHome)