New model buzz, agent wars, and a push for an AI kill switch

There's a particular irony in Diogo Almeida walking away from the chatbot he helped build. Almeida co-created RLHF, the technique that made ChatGPT behave like a helpful assistant rather than a text-completion engine, and by most accounts he grew disillusioned with where that path led. His new model approach is now generating real buzz among developers, and I think the reason is simple: people are hungry for evidence that the current paradigm isn't the only way forward. We've spent three years treating "make the chatbot nicer to talk to" as the central problem in AI, and one of the people who built that machinery is now suggesting there's more interesting work elsewhere. Worth watching not because it's guaranteed to matter, but because dissent from insiders carries more weight than dissent from critics who never touched the plumbing.

Meanwhile the agent story keeps accelerating on every front, and not always in ways companies are prepared for. AI agents are reportedly becoming one of the largest sources of internet traffic, which sounds abstract until you consider what it means practically: websites built for human browsing patterns are now serving bots that read differently, click differently, and don't tolerate friction the way people do. E-commerce platforms and content sites are having to redesign for an audience that isn't human, which is a strange sentence to write but an accurate one. Google, for its part, is pushing its CC agent into household management — shared calendars, shared emails, coordinating family logistics — which tells you agents are moving from "my personal assistant" to "our shared infrastructure." That's a bigger trust leap than it sounds. Letting an agent read your calendar is one thing; letting it mediate between multiple people in a household with different priorities is another.

On the enterprise side, VB Pulse data shows a real gap opening between OpenAI and Anthropic when it comes to agent platform adoption — companies using OpenAI's tools are far more likely to actually be building products with them, not just experimenting. That's the kind of data point that matters more than any benchmark score, because it measures commitment rather than curiosity. If this gap holds, it becomes self-reinforcing: more real deployments mean more feedback, more integrations, more reasons for the next enterprise to default to the same platform. I'd want to see this tracked over the next two quarters before drawing firm conclusions, but it's a signal worth taking seriously.

And then there's the safety conversation, which is starting to sound less theoretical. Gavin Newsom signing an executive order to explore a possible AI "kill switch" is a notable moment for a state government to formalize, even as a review rather than binding law. Pair that with Dario Amodei's renewed call for the industry to slow down and let safety research catch up, and you get a picture of an industry racing forward while its own leaders and regulators quietly admit they're uneasy about the pace. By the way, isn't it striking how rarely "slow down" calls come from anyone actually willing to slow down first? That tension isn't going away — if anything, agents handling real infrastructure and real households make it more urgent, not less.