Cursor's unpatched Windows zero-day, Moonshot's 2.8T-parameter Kimi K3, and Codex CLI's encrypted audit trails.

GitHub agent repo leak, Claude Code's harness token overhead, and TSMC's record $39.62B quarter.

A researcher decoded hidden Claude Code prompt markers and Anthropic pulled them. Developers mapped GPT-5.5 reasoning tokens clustering at 516. DeepReinforce shipped MIT-licensed Ornith-1.0, targeting one 80GB GPU.

OpenAI unveiled its first inference chip with Broadcom, Ford rehired 350 engineers after an AI push fell short, and Google built computer use into Gemini Flash.

US curbs pushed buyers to open-weight models, two reports argued AI coding's hard part is no longer writing it, and agents got identities of their own.

A US order halted Anthropic's new models, AI agents rewrote Git in Rust, and Hugging Face open-sourced a DeepSeek-R1 reasoning recipe.

AI coding tools may cost providers 10x what users pay, Apple rebuilt Siri on Google's Gemini, and DeepSeek's V4 reportedly beat GPT-5.5 on precision.

Memory drives most of your inference bill, a cost-optimisation layer starts forming around developer tools, and agentic models reach consumer silicon.

OpenAI and Notion ship agent platforms while Google confirms the first AI-driven zero-day exploit. This week's AI briefing for engineering leads and security architects.

Six months ago, running a frontier AI model locally was a research curiosity. This week, the creator of Redis shipped million-token inference on a MacBook. #AI #LocalInference #MachineLearning #DeepSeek #AppleSilicon