Thursday, 18 June 2026

When Noam Shazeer walked out of Google this week to join OpenAI, he didn't just change jobs—he crystallized something we've been watching unfold for months: the talent migration that follows perceived momentum shifts in AI.

Shazeer is no junior engineer. He co-led Gemini, one of Google's flagship models, and held the title of VP of engineering. His departure matters not because any one person is indispensable—they aren't—but because it signals something about where serious builders think the most interesting work is happening right now. OpenAI clearly sees him as valuable enough to justify poaching him from a company with vastly deeper pockets. Google must see it that way too, or they would have fought harder. By the way, this kind of churn at the top is often a lagging indicator. We should be asking what Shazeer saw that made him jump, and what Google's leadership is feeling that they let him go without apparent negotiation.

Meanwhile, the real geopolitical story isn't the personality shuffle—it's Microsoft quietly building a substantial business selling OpenAI's models to Chinese companies. This caught my attention because it represents a crack in the US-China AI rivalry that everyone assumed would be airtight. Microsoft has found a workaround in the regulatory ambiguity, and they're making money from it. The irony is sharp: the same models that the US government is theoretically worried about China accessing are being sold to them through a licensed partnership. This doesn't resolve the tension; it just relocates it. At some point, someone in Washington will notice, and the arrangement will either become formalized or it will blow up.

On the applied side, the announcements from AWS this week around agents—web search integration in Bedrock, new tools for orchestration—suggest that the industry is finally moving past the chatbot phase. These are incremental improvements, but they're the right kind of incremental. Agents that can ground themselves in current information and coordinate across tasks are closer to something genuinely useful than another fine-tuned chat interface. The question that still lingers is whether companies can deploy these at scale without the machinery becoming too expensive or brittle to maintain.

There's also growing noise about government oversight of frontier models, with Rep. Gottheimer pushing for mandatory reviews of advanced systems. I find myself ambivalent here. The risks he's citing—cybersecurity, bioweapon design—are real enough that they merit serious attention. But the conversation around "frontier models" is still fuzzy. What's a frontier model? Who decides? And how does a regulatory review actually reduce risk versus just creating compliance theater? These questions need answering before the rules calcify.