Short briefings on models, tools, agents, and the rules that govern them.
English edition
Systems that take a sequence of actions, and the question of what they are allowed to touch.
Anthropic's S-1 and its numbers, a $63 billion Broadcom chip deal for Anthropic, Meta's plan to raise up to $95 billion, Microsoft's buyout offer, OpenAI's cheaper GPT-6.1 Terra, Siri AI in Portuguese and OpenAI's data on ChatGPT at work.
OpenAI's offline training and Brotli flaw, Altman and Cloudflare on AI-found bugs, Texas sues OpenAI, Elvis's rights holders sue Anthropic, a study of Grok and GPT voting picks, Sweden blocks ChatGPT and Claude, and the ITU's AI-divide warning.
Moonshot's Kimi K3.1 Pro, MiniMax M3.1 Flash weights, Block's speech-only Dessa, China's AI-jobs guidelines, the UK's £600m AI-security package, and Anthropic's Mythos 5 incident review.
Cursor Cloud Agents can now generate interactive React canvases. The release also adds the Environments API beta and speeds up multi-repo workspaces.
Nous Research integrated Hermes Agent with NVIDIA Nemotron 3.5 Super VL, reaching 63.44 on Hermes Bench and up to 5x higher throughput.
Perplexity open-sourced SPACE, its Firecracker microVM sandbox for AI agents, separating system root access from API credentials and network gateways.
Nous Research released 50+ Hermes Agent connectors with RBAC controls, operational boundaries, and local execution that requires no account.
Waymo opened its robotaxi service to everyone in San Antonio, San Diego, and Tampa, ending waitlists and connecting San Antonio International Airport.
Anthropic rolled out native OpenTelemetry traces in Claude, exporting agent latency, tool calls, and token metrics directly to enterprise OTLP collectors.
Intel showed how AI agents cut PDK design rule cycles by more than half and navigated a 350-dimension fab problem with 10^700 combinations in months.