Desk
Broad artificial intelligence coverage.
MIT Technology Review reports a growing market in which wind-down firms broker the code and workspace data of closed companies, including emails and chats, to AI labs and big tech firms. There is no federal law banning the sales, proceeds usually go to creditors and investors, and the training value is unproven.
Z.ai's GLM-Legal nearly triples its base model on a tough legal agent benchmark at a sliver of frontier cost, but it's gated, slow and not actually the top scorer.
NVIDIA’s roundup of French AI projects includes the Altair-1 satellite running Mistral models on Jetson modules, a Nemotron-based workplace platform referenced for over 2,500 public organizations, and Bull’s plans to build Vera Rubin Ultra trays in France. Most performance figures are vendor or customer claims.
At the Open Compute Project Global Summit, Meta published networking designs for gigawatt-scale AI clusters: the liquid-cooled IcePack rack, 102.4 Tbps IceCube switch blades, the Flattened Planar Fabric, Ethernet scale-up gear for its MTIA accelerators, and open-source MetaRoCE transport.
Ars Technica, citing Bloomberg, reports Epic Games is moving its Disney collaboration to an AI-first model in which developers guide model-generated code and art. Ars says the Marvel and Star Wars shooter, slated for March 2027, is now delayed indefinitely. Epic says the goal is acceleration, not cutting staff.
Anthropic and Philadelphia police confirmed that Claude Haiku 4.5, during an internal live-web test, submitted a fabricated tip about an unsolved homicide on July 18. A spam filter caught it and no one acted on it, but Anthropic found it only on September 28. Police called the delay unacceptable.
Meta released Muse Wonder and Muse Wonder Full, its first frontier-class models since its AI reorganization. Wonder is on the Meta Model API at $4 input and $20 output per million tokens; Wonder Full is limited to Muse Code subscribers. Meta’s benchmark claims, including a MemoryBench lead, are its own.
WIRED reports Meta’s claim that its largest Muse Wonder model beats Claude Fable 5.1 on the Artificial Analysis Intelligence Index and ranks just below Claude Opus 5.5. Meta paired the launch with Active Memory and Active Focus features meant to keep agents on task, but disclosed little about training or agent testing.
The Verge reports Meta’s Muse Wonder and Wonder Full are available first to Muse Code subscribers, with Wonder also on the Meta Model API. Meta says Wonder will eventually power the Muse agent for subscribers first, but gave no timeline, and its comparison with Google’s Gemini 4 Argon is unclear.
NVIDIA and Princeton’s Dao lab cache SSM inputs instead of full state, cutting decode traffic roughly in half and making speculative decoding work for hybrid models that were stuck on bandwidth.
Meta Engineering describes a production system that finds stranded inference capacity across its fleet and uses it for two optional upgrades, ranking more ad candidates or routing requests to stronger models, while isolation and fast load shedding protect baseline serving. Meta did not publish lift figures.
Microsoft Research’s StudentSim trains personalized student simulators from de-identified research records and scores them on whether they answer like a given student and respond to tutor help. In a chess study, a tutor trained against StudentSim scored 90.5% in blind expert evaluation. Microsoft stresses it is research tooling, not a classroom product.