The signal story from Anthropic's London developer event is stark: half of active developers are already shipping unreviewed AI-generated code. Paired with Claude Code's new 'dreaming' feature — which lets agents self-document and learn across tasks without human input — the oversight gap isn't theoretical. It's compounding in production right now.

OpenAI's general-purpose reasoning model independently disproved a significant mathematical conjecture this week, with no specialised training and no human guiding the proof. It's demonstrated capability, not projected potential — and the direction is clear.

At Google I/O, the Gemini for Science package signals a deliberate pivot from deep specialisation toward autonomous agent-driven discovery. WeatherNext's advance warning on Hurricane Melissa's Jamaica landfall shows what that shift looks like when it produces real-world impact.

On the infrastructure side, Alibaba's T-Head division has produced 560,000 Zhenwu AI chips in total. Nvidia ships three to four million GPUs annually. US semiconductor export controls are working as designed — but the domestic capability gap shapes every ambition in Chinese AI infrastructure.

Finally, third-party involvement in security breaches has doubled, from 15% to 30%, with AI-driven attack vectors — model poisoning, prompt injection, supply chain manipulation — as the documented driver.

The thread across every story today: automation is accelerating, and human validation is retreating. The metric to watch is whether the industry builds review infrastructure that keeps pace with deployment speed. Right now, it isn't.

This episode includes AI-generated content.