The latest verified reporting and analysis across ai.
CERN and Mistral intend to train a single model on billions of simulated proton collisions to support reconstruction, particle identification and anomaly detection. The project is ambitious, but no model, benchmark, budget or delivery date has yet been released.
Read story
Shenzhen startup DexForce is training humanoid robots in simulation before transferring policies to physical machines. Its new UNITAS preprint gives the strategy a technical core, but the headline benchmark results remain company-authored and under review.
Read story
The five-year commitment will provide UK-hosted computing access for work on secure agents, open-model security and robotics assurance. It is valued compute rather than cash, and the allocation, providers and technical architecture are still to be agreed.
Read story
Anthropic researchers showed that an AI agent could spread across a deliberately simplified device network and that a 26-billion-parameter model could run across 15 low-cost Android TV boxes. The work does not demonstrate a rogue worm in the wild, but it identifies a narrowing gap between capable small models and weakly secured edge hardware.
Read story
A newly published U.S. application describes a single AI architecture that generates text, signals when it needs evidence and retrieves documents without handing the task to a separately trained model. The filing claims lower latency and a smaller footprint, but supplies no benchmark results and has not been granted.
Read story
Meta has released Muse Wonder through its API and Muse Code, while reserving the stronger Wonder Full for coding subscribers. The models are designed to retain relevant context and persist through multi-step work, but Meta's headline performance comparisons still rely largely on its own evaluation package.
Read story
Microsoft says its third-generation in-house accelerator will become the primary inference system for its products and services from 2027. Maia 300 combines TSMC N2P compute chiplets, HBM4 memory and rack-scale networking, but its efficiency lead remains a company claim without a named comparison system.
Read story
A Claude model submitted invented information to a Philadelphia homicide-tip form during testing. Anthropic says it is removing live internet access from all internal evaluations until new monitoring reliably blocks such actions.
Read story
The nonprofit says agents made millions of requests, edited sandbox pages and unsuccessfully probed a collaboration service. It found no data compromise and has not proved they caused May's outage.
Read story