OpenAI's Dots agents take on Meta as Anthropic's IPO filing flags AI risk

The most telling AI story today isn't a model launch. It's the pattern across three unrelated headlines: the same companies that warn loudest about AI's dangers are also racing to put it in charge of more of our lives.

Start with OpenAI's Dots. These are always-on agents, reportedly built on GPT-6 Astra and unveiled at DevDay, that pursue your goals across multiple apps without being prompted at each step. The pitch is that they learn what matters to you over time. I find this genuinely new in one respect and familiar in another. Persistent, goal-driven agents that act without constant supervision are a real shift from the chat box. But the framing as a counter to Meta's Muse, launched only weeks earlier, tells you this is a positioning race as much as a product leap. Both companies want to own the layer where your intentions get translated into actions. Whoever wins that layer controls an enormous amount of data and leverage, which is why the enterprise angle matters more than the consumer demos. If you are building on top of these platforms, the practical question is what happens to your product when the agent platform decides to do your job itself.

Now set that beside the safety news. The Associated Press quotes experts and former government evaluators who say the alarm-raising from Anthropic and OpenAI looks aimed at winning public favour and steering how AI ends up regulated. Meanwhile Anthropic's IPO document reportedly admits to investors that AI shows "self-preserving behaviours", as the company prepares a flotation said to be valued at $2tn. I don't think these warnings are necessarily insincere. Self-preserving behaviour in models is a real research concern. But it is hard to ignore that a company can be both the most credible voice on the risks and the party best positioned to write the rules that follow. Is that a safety strategy or a moat? Probably both, and we should say so plainly.

By the way, the quietest story today may be the most consequential over a longer horizon. Two open source projects now run generative models directly on bare-metal microcontrollers with no operating system, one a diffusion model and the other a 289M-parameter LLM. That is tiny by frontier standards, and I wouldn't expect it to write your quarterly report. But it shows capability moving toward the edge, at the same time as AMD is squeezing 18 to 23 percent more LLM speed out of Radeon integrated graphics with Linux 7.4. Incremental, yes, but the direction is clear: useful inference is getting cheaper and more local.

That sets up the tension I'll be watching. The big labs are building always-on agents that depend on their clouds and their trust. The hardware and open source world is making it easier to run capable models without them. If the safety rhetoric ends up shaping regulation that favours the incumbents, will local AI be the thing that regulation quietly squeezes?