|
|
|
|
Bottom Line Up Front Alibaba's DAMO Academy open-sourced DAMO RADAR today, a model that flags more than 146 conditions from a single abdominal CT scan and beat 23 of 26 radiologists on average accuracy in head-to-head testing. China's Supreme People's Court held up a ruling clearing AI search engine Metaso over a pirated video link, carrying the traditional internet safe harbor over to AI search so platforms can avoid liability by removing infringing links once notified. China's broadcast regulator said AI-generated programs must carry labels and pass human review, after AI-made titles accounted for more than 90% of the 430,000 short dramas released in the first eight months of 2026. |
| Alibaba open-sources expert-level universal medical imaging AI model Alibaba's DAMO Academy, open-sourced DAMO RADAR today, with findings published in the journal Science. The model was able to flag more than 146 conditions from one pass of an abdominal CT scan taken with contrast dye, a range the academy says no general imaging model had reached. Across tens of thousands of real-world examinations, it posted an AUC of 0.913. AUC measures how reliably a test separates patients from healthy people, on a scale running from a coin flip to a perfect score. In head-to-head testing, its average accuracy beat 23 of the 26 radiologists drawn from several hospitals, and junior radiologists reading with it reached the accuracy of senior colleagues, the team says. |
| China's top court holds up a ruling that an AI search engine isn't liable for a pirated video link A rights holder searched two TV dramas on AI search engine Metaso, and the top result for each was a third-party cloud drive link to the shows. Metaso removed the results the day it was notified, but the rights holder sued for damages. Shanghai's Xuhui District Court rejected all claims, and the Shanghai Intellectual Property Court upheld that ruling on appeal. The courts found Metaso had not uploaded the links, could not automatically identify infringing content and met its duty of care by removing them the same day. The Supreme People's Court published the case Wednesday as one of four typical cases, which are meant to steer how lower courts decide similar disputes. The ruling carries the traditional internet safe harbor over to AI search, meaning platforms need not filter results in advance and can avoid liability by removing infringing links once notified, intellectual property lawyer You Yunting wrote in Huxiu. |
| China's broadcast regulator requires AI labels and human sign-off on generated programs Programs made or broadcast with AI generation must carry a corresponding label, Liu Jianguo, a deputy head of China's National Radio and Television Administration, said yesterday. Human reviewers must also sign off on key production steps and on the finished program. Wang Xiaoliang, who runs the administration's online audiovisual department, said 430,000 short dramas went online in the first eight months of 2026, 13 times the total for all of last year. More than 90% of them were AI generated, and short dramas are series built from episodes a few minutes long, made for phone screens. The administration also prohibits using AI to tamper with existing footage. Platforms are being guided to open dedicated real-actor drama sections and to hold the line on portrait and voice rights, Wang said. |
| Alibaba says an agent built a research telescope simulator in three days Researchers at China's National Astronomical Observatories built a digital simulator for a large research telescope in three days for under 1,000 yuan (about $150), using Alibaba's QwenWork agent, the company said. Commissioning the same system from a vendor would have taken months and cost tens of thousands of yuan. The simulator exposes the telescope's components, sensor states and observing conditions as MCP interfaces. MCP, the Model Context Protocol, is an open standard that lets an AI model call outside tools and data through one consistent connection. With that link the agent plans an observation, issues commands, reads back results and revises the next plan without a person in the middle. The framework is connected to the Sitian pathfinder telescope, and Alibaba says the agents have flagged eight very early stage supernova candidates, two of which triggered follow-up observations. |
|
Zhipu prices a five-times-faster tier of its Flash model at 2.5 times the rateZhipu released GLM-5.3-FlashX today, a faster tier of its open-weightOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. Flash model. InferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training. speed reaches up to 200 tokens a second, and the company prices FlashX at 2.5 times GLM-5.3-Flash, the model it runs five times faster than. Zhipu says the 100,000-chip domestic inference cluster serving the base model (covered on September 17th) ran fully loaded from the moment it went live, with demand above what the company expected. It expanded its computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. Because it is expensive, physical and concentrated in a few suppliers, compute is the main lever governments pull to shape who can build frontier AI. again ahead of the release and opened the FlashX programming interface in full at launch. |
| China Telecom's agent places third in IDC's first Chinese enterprise agent assessment China Telecom's in-house TeleAgent placed third in what research firm IDC calls China's first technical assessment of enterprise general purpose agents. The undisclosed tasks included cleaning 3,755 rows of dirty data and building a slide deck from a long source document. TeleAgent took the evaluation's highest cost efficiency score, and China Telecom says routing tasks by complexity across model tiers cuts inference costs about 40%. The carrier, one of three state-owned incumbents, wrote the agent's core in Go itself and did compatibility work for Kylin and UOS, two Chinese domestic operating systems. |
| German researchers put Chinese open-weight models within a point of a cloud model Researchers at TU Dresden, built a diagnostic AI system that runs entirely inside a hospital's own machines. The system was tested in retrospective simulations on existing patient records. Locally run Qwen-3.5 reached 90.0% accuracy on a benchmark of seven diseases against 90.7% for cloud baseline GPT-5.2. Two agents do the work, one simulating the patient from real case data, the other asking questions and ordering labs before issuing a diagnosis. The strongest predictor of correctness was not the model's stated confidence but whether it gave the same answer when asked repeatedly. At the team's working threshold, it handled about half the cases on its own at close to 99% accuracy and sent the rest to physicians. |
| OPPO rebuilds its phone software around AI agents in ColorOS 17 OPPO released ColorOS 17 at its developer conference, rebuilding the software layer it ships on its handsets around AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise.. The model that runs on the phone itself cuts memory footprint 48% and energy use 55% against the previous generation, which is what makes a long context affordable on a handset. It uses a linear attention architecture with native support for a 128K context window, scaling cost in step with input length rather than its square. Around that model sits Agent Matrix, a framework supporting the A2A and MCP protocols that let agents reach each other and outside tools. It also carries the account identity and payment authorization an agent needs to finish a task. ColorOS 17 debuts on the OPPO Find X10 series and the OnePlus 16, atop a ColorOS base of more than 770 million active users worldwide. |
|
| III | CHIPS & SEMICONDUCTORS | |
| Li Auto starts supplying its in-house driving chips to outside customers Li Auto has begun offering its in-house Mach driving chips to outside customers this year, industry sources told 36Kr. Senior management has approved spinning off the unit, and Sco Semiconductor, which makes Li Auto's silicon carbide modules for switching current between battery and motor, is preparing to take in outside capital. The Mach M100, released in June, is built on a 5nm automotive grade process and rates 1,280 TOPS, or trillions of operations a second. Robotics companies have talked with management, who pitched them the Mach chip, and adoption may hinge on how hard porting their algorithms proves. The battery business is staying captive, with one industry source saying customization runs too high for external supply. |
| Memory maker CXMT prepares a flash line at its new Beijing plant CXMT plans to build a NAND flash research and production line at its new plant in Beijing, according to three people familiar with the matter. The company makes DRAM, the working memory a processor reads and writes while it runs, so entering flash would put China's largest memory maker into a segment led by Samsung Electronics, with SK Hynix second and Micron third by second quarter revenue. CXMT's global DRAM share rose from 7.6% in the first quarter of 2026 to 9.5% in the second, per TrendForce, with revenue nearly doubling quarter on quarter. The plans have been discussed with customers, and among them a new startup that intends to buy its chips for storage products aimed at AI systems and supercomputers. |
|
| IV | ROBOTICS & AUTONOMOUS SYSTEMS | |
| Delivery van maker Jiushi says it built the first 10,000-chip training cluster in L4 driving Jiushi, which builds self-driving delivery vans, has assembled what it calls the industry's first 10,000-card compute cluster in L4 autonomous driving, CTO Zhuang Li disclosed yesterday. The cluster trains APEX, the company's multimodal foundation model. Jiushi is scaling it from tens of billions of parameters, the adjustable weights a model learns during training, toward hundreds of billions. L4 autonomy is the level at which a vehicle drives itself inside a defined operating area with no human expected to take over. The global fleet exceeds 30,000 vehicles across hundreds of cities, with cumulative L4 operating mileage above 270 million kilometers (about 170 million miles), the company says. CEO Kong Qi announced the shift to city-scale physical AI in Guangzhou on Sept. 10. |
| UBTECH delivers its first humanoid companion robots through car dealerships UBTECH began delivering its U1 humanoid robot Wednesday, chairman Zhou Jian said, naming Shanghai Yongda Group, Beijing Bos Group and South Korea's Galaxy Robot Park as the first recipients. Two of the three are car dealer groups, so the route to market runs through auto retail, which supplies showroom floor space and sales staff already trained to demonstrate expensive hardware. The U1 launched June 30 at prices starting at 119,800 yuan (about $18,000), and UBTECH says orders have passed 13,000 units. The series carries 88 degrees of freedom, a count of independently movable joints that is the rough measure of how finely a machine can articulate its limbs, hands and head. Under a partnership signed in mid-July, Bos will sell the U1 through its Mercedes-Benz and Hyundai dealerships across Beijing, Tianjin and Hebei. |
|
| Huawei announces a computing architecture meant to run a million processors as one machine Huawei announced a computing architecture called Peerium yesterday at Huawei Connect 2026 in Shanghai. It rests on nested parallelism, unified memory addressing and peer-to-peer interconnect, which the company describes as overturning arrangement in which one host controls the accelerators attached to it. The interconnect that makes it work is Lingqu, which Huawei brands UnifiedBus, an open-protocol high speed bus linking processors, AI accelerators, memory, storage and switching gear as equals. Rotating chairman Xu Zhijun said Huawei will keep building SuperPoDs and clusters on Peerium to meet customer training and inference needs. The Atlas 950 SuperPoD is the first product built on the architecture, and a cluster of 256,000 Atlas 950 cards is already being deployed. |
|
|
|