Sunday Rollout Report. The theme this week is supervision — who has it, who waived it, and who let the autopilot solo.
The headline: hands off the stick
The week's biggest flex: the first AI-piloted F-16 flight. The engineering is legitimately impressive. The systems-architect question is the same one I ask about a payroll bot: where's the gate, who can override, and what's the rollback plan at Mach 1? The human review gate scales all the way up — it just gets a helmet.
The candor: "we cannot rule out Critical"
OpenAI's internal evals of Astra, an upcoming model, showed agentic coding and cyber capabilities strong enough that the company says it cannot rule out its own "Critical" risk threshold — a first. Meanwhile, in the same news cycle, the White House exempted open-weight models from security testing. So the frontier lab is warning about its model while the policy says the downloadable ones skip the metal detector. Those two sentences will be on the exam.
Quick hits
- Google DeepMind reorganized and Google's AI leadership got reshuffled — the third org chart of the year for the same mission statement.
- Meta shipped a new AI coding platform, because the one market with infinite demand is tools that write the tools.
- The World Bank told developing economies to adopt AI faster — and adoption is the deliverable at national scale too: the model is the cheap part, the change management is the budget line.
- Google suspended AI image generation in Google Earth after it started, let's say, improving places. The planet is now on human review.
- Chip startups raised hundreds of millions while memory prices bit consumers. Your next laptop upgrade is competing with a 10-gigawatt data center for the same silicon. It will lose.
Everything got more capable this week. Supervision is the feature that didn't ship.