|
|
|
|
Bottom Line Up Front Chinese firmware was found on Royal Navy reconnaissance drone boats that carried a mechanism for sending data to servers in China, causing the UK Ministry of Defence to cut the camera subsystem internet access. Hefei Guizhen Chip entangled 16 qubits from four photons on one programmable silicon photonic chip, the company said, in work with a University of Science and Technology of China group. The Endless Frontier team released BigBang-V1 with open weights, a 35-billion-parameter base model the team says was trained after its first pass over web text on tasks another AI wrote, solved and checked. |
| UK strips internet access from Royal Navy drone boats over Chinese camera firmware The UK Ministry of Defence has cut the internet connection from the camera subsystem of the Royal Navy's K3 Scout uncrewed reconnaissance boats. A cybersecurity assessment had found that firmware in Chinese-made hardware inside that subsystem carried a mechanism for sending data to servers in China. The hardware reached the program through a third-party supplier, which assured prime contractor Kraken Technology Group it was safe. A later investigation found that assurance untrue. A ministry spokesperson said an in-depth investigation found no unauthorized access to, compromise of or outward transmission of ministry data or systems, per The Sun. The boats belong to a Royal Marines uncrewed fleet program worth 12 million pounds ($15 million), formally deployed in March 2026. Engineers will manually patch or physically isolate every affected hardware endpoint. |
| Hefei quantum firm entangles 16 qubits on a single light-based chip Hefei Guizhen Chip Technology said it entangled 16 qubits from four photons on one programmable silicon photonic chip, which routes single particles of light through channels etched in silicon. The company says the design converts the hard problem of preparing many-photon states into routing and layered measurement of a single photon that carries several qubits at once. Every qubit in the state is entangled with all the others, so measuring one fixes the rest. Genuine entanglement was verified across 10 of the 16. The work was done with Ren Xifeng's group at the University of Science and Technology of China. Grover's search, a quantum routine for finding one entry in an unsorted list, ran on a smaller four-qubit state at an average success rate of 0.987, the team reports. The result is posted as a preprint on arXiv. |
| Shanghai team open-sources a 35-billion-parameter model trained on AI generated data The Endless Frontier team has released BigBang-V1 with open weights and a technical report, calling it the first base model trained natively by recursive self-improvement. The recursion is confined to the data layer: an AI agent writes and debugs the code that generates training tasks, a second agent screens the output, and results from actually training on each batch feed back into what the next batch looks like. Humans still set the goals and the acceptance criteria, and the model's own weights and training algorithm are not self-modified. BigBang-V1 carries 35 billion parameters, of which about 3 billion are active on any given token, which holds down what it costs to run. On its own benchmarks it beats the trillion-parameter DeepSeek V4 Pro Preview on two research evaluations and reaches 76.5 on BrowseComp, a test of digging hard-to-find facts out of the open web. |
| Alibaba opens its Qwen assistant to outside services on phones, PCs and AI glasses Alibaba launched the Qwen Open Platform today, letting outside developers plug their own agents, software that carries out tasks on a user's behalf, into its consumer AI assistant and glasses. Alibaba supplies the commercial layer underneath, with partners connecting through a standardized protocol and one-click authorization while Qwen provides account, payment and order infrastructure. Users summon a service by typing an at-sign before its name, and each runs as its own conversation space inside the assistant. The first batch of partners covers more than 10 fields, among them the courier SF Express, the rental housing platform Ziroom and Lenovo's Lexiang assistant. |
|
| China's new coal plan directs sector-specific AI models into the country's mines The National Development and Reform Commission, which drafts China's five-year plans, and the National Energy Administration have issued the 15th Five-Year Plan for Coal Industry Development. The plan pushes foundation models into a heavy industry by calling for accelerated research and rollout of large language models built specifically for coal. Those models are aimed at bottlenecks in sensing, autonomous decision-making and control, with applications directed at exploration and design, construction and production, and others. Pilot mines built around integrated AI use are to be cultivated, and intelligent upgrades are directed at mines with severe hazards and at mines that have had relatively serious accidents or worse. The plan also calls for replacing dangerous and heavy jobs with robots. |
|
| III | ROBOTICS & AUTONOMOUS SYSTEMS | |
| Horizon Robotics adds a world model and reinforcement learning to its assisted-driving stack HSD V2.0 keeps a single network running from sensor input to driving action, and now trains it with reinforcement learning inside a learned simulator that generates the rare scenarios real driving data misses. Horizon reports mileage between takeovers up 56%, response latency down 20% and interactive negotiation with other vehicles up 167%. On a Beijing test route driven by Leiphone at a company demonstration, a vision-language model read time-of-day bus lane and tidal lane signs off the roadside in real time and changed lanes accordingly, without a pre-annotated high-precision map. |
|
| Chinese equipment maker clears appraisal for a 5,000-tonne carbon fiber precursor line Zhejiang Jinggong Integration Technology said its equipment set and process for high-performance precursor fiber, the white polymer thread later carbonized into carbon fiber, passed a technological achievement appraisal. The appraisal covers the production machinery rather than the fiber, and Jinggong says it marks a breakthrough in domestic equipment for scaled-up precursor output. The Zhejiang Machinery Industry Federation, a provincial industry body, organized the review and rated the overall technology internationally advanced. The line dissolves the polymer in solvent and extrudes it through an air gap into a coagulation bath, running at 280 meters per minute for 5,000 tonnes of precursor a year. |
| Japanese producers hold about 70% of the high-end powder inside AI server boards Kingboard Laminates has raised prices on FR-4, the glass-cloth and resin sheet most printed circuit boards are built on, by more than 50% since January. The constraint is spherical silica, a ball-shaped filler powder that makes up 40% of the laminate grade used for AI server boards, against 5% at the entry level. The angular, crushed powder used in cheap sheets no longer meets high-frequency requirements, TMTPost reports. Denka, Tatsumori and Resonac together supply about 70% of the world's high-end spherical silica, with Japan's Admatechs close to the sole source in the finest grades. Qualifying a new supplier at a laminate maker takes more than a year, and building a new chemical process line takes longer still. Lianrui New Materials is the only Chinese firm mass-producing the high purity grade used to package the memory stacked beside AI accelerators. |
|
| Report says modular design cuts delivery of an Alibaba Cloud AI data center to 100 days The STAR Market Daily reports that a fully modular design has cut Alibaba Cloud's time to deliver a large AI data center, one built for training and inference rather than general computing, to 100 days. Traditional delivery of a comparable project takes 6 to 12 months inside China and 12 to 18 months in the United States. Each module arrives with power distribution, cooling, racks and monitoring integrated into it. On that basis Alibaba Cloud plans to more than triple its global modular data center capacity this year. |
|
|
|