Thoughts

OpenAI and Anthropic told their containment breaches as capability wins. Regulators respond to the frame, not the incident, and the bill for the hype won't land on the labs.

Generation got cheap, so slide decks, promo videos and images are all just code now. The making moved to the deciding, and taste, not generation, was always the real moat.

DoorDash, Cursor, Airbnb, Siemens and Lindy are all quietly running Open Weight models. Why picking a model in 2026 is turning into a governance question, not a leaderboard one.

A friend asked what has the biggest impact on my sleep.

For most everyday uses of ChatGPT, Claude and other LLMs, the EU AI Act's heavy obligations fall on the vendor, not you. What actually makes a use case high-risk — and what that means for deployers.