Tuesday, 23 June 2026

The talent drain from Google's AI division continues unabated. Noam Shazeer, who spent years building foundational systems at Google before his recent departure, is now joining OpenAI. It's worth noting what this signals—not just that OpenAI can attract top researchers, but that Google's ability to retain them seems to be slipping. This matters because the people building these systems often matter more than the compute or capital behind them. When a figure like Shazeer moves, it's a vote on where he thinks the most interesting problems are being solved.

By the way, the competition for multiagent orchestration is heating up in unexpected places. Sakana AI, a Japanese startup, just released Fugu—a system that dynamically coordinates multiple language models to behave like a single, more capable system. The results are striking enough to be worth taking seriously. Fugu is matching or exceeding performance from Anthropic's latest models on standard benchmarks. This is the kind of news that doesn't dominate headlines but should occupy space in any founder's head. The insight is simple but profound: you don't always need a bigger model. Sometimes you need a smarter way to combine smaller ones. For teams building production systems, that's potentially liberating.

What's less liberating is the safety pressure now manifesting in real regulation. Anthropic's Dario Amodei warned publicly that their Mythos model posed national security risks. The government listened, and Anthropic shut it down. I find this genuinely concerning—not because safety scrutiny is wrong, but because the process here remains opaque and reactive. The precedent matters more than this particular case. When does a capability become a threat? Who decides? And how do we avoid a scenario where safety warnings become a blunt instrument for suppressing competition? These questions are now live in ways they weren't six months ago.

The developer tooling layer is also shifting. Microsoft's Copilot ecosystem is expanding—Copilot Cowork is now generally available, and JetBrains IDEs now support Claude as an agent provider. Meanwhile, GitHub released a spending meter in VS Code after developers complained about unexpected billing. These moves suggest two things: first, that AI-assisted development is becoming mainstream enough to require financial transparency and governance controls. Second, that the IDE wars are being won and lost in the margins of agentic AI integration. When your development environment can coordinate multiple AI models and decide when to call which one, the choice of IDE suddenly carries more weight than syntax highlighting ever did.