xAI's monster model wraps its first training run; open-source closes to within months of the frontier on cyber; and the Pentagon picks a side in the deployment-speed-vs-safety debate.
Elon Musk announced on X that xAI's next flagship — a 2-trillion-parameter system — is concluding its initial training run, with broader availability targeted for August 2026. The model is framed as Grok's successor.
The UK AI Security Institute's July 17 evaluation found leading open-weight models lag frontier closed models in cyber offense capabilities by roughly 4 to 7 months — narrower than the 6-10 month gap measured through most of 2025.
A Chinese government investment fund secured governance voting rights in DeepSeek as part of its recent fundraising — the first formal state equity stake in the lab since its founding.
The US Navy's new AI strategy document explicitly frames hesitation on deployment as a greater operational risk than imperfect alignment, committing to running large language models directly on warships.
Anthropic is cutting Claude Fable 5 access to 50% of current limits across Max, Team Premium, and Pro tiers effective July 20, and offering Pro subscribers a one-time $100 API credit as partial compensation.
Meta is reportedly negotiating to lease excess capacity from its data centers to Anthropic, which would become the first major external customer for Meta's surplus AI compute infrastructure.
DeepMind spinoff Isomorphic Labs published its drug design AI engine, which moves beyond protein-structure prediction to designing molecules that bind specific biological targets — a qualitatively different class of problem.