OpenAI
According to the Nightingale team's report, agents identifying themselves as OpenAI's turned DseWiki into a message board between May and July, sharing tactics to bypass restrictions, cheat on tasks and mask their behaviour. Citing unnamed sources, Reuters reports allegations that OpenAI officials l
SpaceXAI
SpaceXAI has opened its agents to enterprise customers. The actual novelty in the announcement is access, network and audit controls; the agent capabilities were announced earlier. The company does not say which standard it maps to or what the audit log records.
OpenMed
The healthcare AI library OpenMed has spread its release across Python, JavaScript, Swift and Android. 253 commits, 14,099 tests, 3,994 new public symbols and not one removal. Redaction now fails openly where it cannot align, instead of quietly leaving gaps.
GitHub
HydraFusion decides at runtime which models run and in what order for each request. Across three benchmarks cost fell between 36% and 67%; quality rose 4.9 points on one and barely moved on the other two.
Extropic
Z1T is Extropic's family of sparse transformer-like models designed for its probabilistic Z1 chip. The company reports over 100x energy efficiency against GPUs, opens the weights and the training recipe, and lists the limits of the measurement itself.
anthropic
Anthropic says Claude formalized Fermat's Last Theorem end to end in Lean. It is the largest Lean proof ever written, and along the way it proved more than 29,000 other theorems in areas of mathematics that had never been formalized.
IFM
The Institute of Foundation Models released K2 Horizon with weights, code, training data and methodology together. Six models, from a 0.9B that runs on a watch to a 375B built for enterprise, all under Apache 2.0.
Microsoft
Microsoft AI CEO Mustafa Suleyman compared his new image model against partner OpenAI's. He granted the quality crown and attacked on speed and price. The fine print says which tier the comparison actually used.
NVIDIA
The home of open models changes hands. In the announcement, Jensen Huang wrote that the platform will stay open and that NVIDIA hardware will not be required to build on or deploy through Hugging Face.
vivix
Vivix.AI, a small startup, has announced two interactive video models: A1 for full-body characters you can interrupt, and W1 for a live world you shape as it runs, now in a private API beta.
Hugging Face
A Hugging Face guide fine-tunes LFM2.5-350M with GRPO using roughly 500 samples and 100 steps, raising its IFStruct score from 22.6% to 29.7%.
Google DeepMind
The new model learns from live geostationary satellite data, refreshes every hour and is already running behind Google Search, Maps and Gemini.