The cutting edge, and the cutting edge already on the clock. On the frontier this week: OpenAI shipped its strongest model to ~20 vetted partners, DeepSeek open-sourced the tricks that make models fast, and a new paper squeezed more reasoning out of far less memory. The edge also left the screen: humanoid robots got their first real safety stack, and China gave AI agents ID cards. And it’s already at work: GPT-5 Pro cracked a three-year immunology mystery, six of the world’s top-10 banks bet $200M on AI fraud detection, and coding agents landed on every phone. The lag between a lab result and a deployed system used to be years. This week it was days.