Open weights had a big week: Moonshot released Kimi K3's weights and technical report just as Anthropic published a policy post insisting it never called for banning open-weight models outright, only for guardrails on frontier-capability releases. NVIDIA is trying to get ahead of regulation with a new Open Secure AI Alliance, while Ilya Sutskever's Safe Superintelligence quietly struck a compute partnership with the same company. Underneath all the model news, the agent-tooling stack is maturing fast - NVIDIA shipped an open-source agentic RL framework and harness-design guidance, GitHub extended Copilot governance controls, and a researcher published a sandbox-escape technique against Claude's Cowork environment. Meanwhile The Register flags that this year's roughly $1 trillion AI infrastructure spend is starting to show up on customer invoices, and DHH's new essay pushes back hard on the rush to hand agents broad autonomy.

• Kimi K3 model weights released, raising the open-weight bar again
• Anthropic says it never asked for a blanket ban on open-weight models
• NVIDIA forms Open Secure AI Alliance; also backs Ilya Sutskever's SSI with compute
• This year's ~$1T AI infra spend is getting passed to customers
• DHH's "I'm sorry, Dave" warns against over-delegating judgment to agents