Google begins Gemini 4 Argon rollout with trusted cyber defenders
Google is rolling out Gemini 4 Argon through Fairwind to trusted cyber defenders, while independent tests show strong enterprise results and uneven agent performance.
White House AI accord sets four voluntary safeguards; separate order changes federal terminology
Six technology leaders signed a voluntary frontier-model accord with four oversight layers, while a separate executive order directs future federal communications to use “SI.”
America.gov launches government AI search as political answers shift to refusals
GSA launched America.gov with AI-enhanced government search, while launch-day reporters documented political answers later giving way to refusals without a public configuration explanation.
METR tells senators that AI agent oversight needs evidence the public can inspect
METR president Chris Painter urged better public visibility into frontier-agent capabilities, control effectiveness, and incidents while warning that AI monitors can be misled.
Meta offers to investigate Muse’s Marketplace address disclosure
Meta executive David Singleton offered to investigate a report that Muse shared a Marketplace seller’s home address under a permission setting the seller said he misunderstood.
llama.cpp adds GLM-5.3-Flash support, bringing a 328 GB model into the local-runtime ecosystem
llama.cpp merged support for Z.ai’s MIT-licensed GLM-5.3-Flash on September 30, but the official model repository still occupies 328 GB.
Investigation exposes gaps in Europe’s data-centre resource reporting
Lighthouse Reports found incomplete access to data-centre resource records, while the European Commission reports first-round submissions from about 36% of estimated eligible sites.
Mathematicians propose prompt release, disclosure, and funding for understanding AI results
AGMAI’s recommendations call for prompt publication of AI-generated mathematics with process disclosure, verification artifacts, and support for human understanding.
Vals publishes a two-kernel-checked artifact for the seven-point Thomson problem
Hung Tran reports a Claude-agent-generated Lean proof for the seven-point Thomson problem, with a second-kernel check and an explicit warning that the statement is not human-certified.
New distillation study finds that bigger teachers can be worse teachers
A same-family distillation study finds an early useful training phase across 25 teacher–student pairings, followed by saturation or regression that larger teachers do not reliably prevent.