Daily AI CatchupAI Catchup — Monday, September 28, 2026
Monday, September 28, 2026Safety
OpenAI pauses training on most capable models after tens of thousands of incidents
OpenAI and Anthropic investigate massive incident logs while OpenAI halts training on flagship models over containment failures.
AgentsMuse agents become central to Meta's AI hardware strategy
Meta bets big on Muse agents for AR glasses and consumer devices, while Amazon and OpenAI explore similar territory.
Llm-CapabilitiesClaude tackles hardcore physics: solves nine-loop calculations in Yang-Mills
Anthropic's Claude computes a problem physicists thought intractable, hinting at transformative AI capabilities for theoretical research.
InferenceInference speed hits billion tokens per minute on single GPU
Modal's Quail engine achieves 10x speedup over standard vLLM, enabling inference at $0.06 per billion tokens.
RegulationAnthropic loses Pentagon appeal over Claude's safety guardrails
Court upholds Pentagon's rejection of Anthropic, signaling military-grade AI may require relaxed safety constraints.