Altman warns of AI power concentration as 400 scientists urge binding rules

Sam Altman is now warning that the AI race could end with a single company holding all the power and cutting everyone else off from resources. I find that remarkable coming from the CEO of one of the three or four firms best positioned to become exactly that company. Is it a sincere worry, a regulatory pitch, or both? I suspect both, and it matters less than what we do with it.

The same week, roughly 400 researchers from 34 countries called for an international body to oversee advanced AI, arguing that voluntary commitments are not enough. I think they have a point about the voluntary part. Voluntary frameworks tend to hold only until the first competitive threat appears. But I also think the signatories underestimate how hard it is to build a binding regime when the leading developers sit in countries that see AI as a strategic asset. Nobody hands over oversight of their most valuable technology to a body their rivals also sit on. Altman's warning and the scientists' letter point at the same problem from opposite directions: concentration of power is the risk, and no one has a credible mechanism for preventing it.

The counterweight, at least on the technical side, is quietly arriving. The llm-d project has shipped version 1.0, giving organisations an open-source path to run LLM inference at production scale. This is not a headline-grabber, but it is the kind of infrastructure that determines whether the future looks like Altman's nightmare or something more distributed. If serving large models well is something any competent team can do on its own hardware, the leverage of a single provider shrinks. By the way, this is where the practical question for builders lies: how much of your stack are you comfortable renting from a company that might one day decide what you can access?

Two other stories show how quickly AI is moving from advice to action. Microsoft has begun rolling out Copilot Code, a coding agent that builds apps from plain-language descriptions, and Ripple released a starter kit that lets AI agents pay autonomously in XRP and RLUSD through the x402 protocol. Put those side by side and you get software that writes itself and then pays its own bills. I am not sure we have thought through what accountability looks like when an agent both builds and spends.

And accountability is exactly what the hantavirus study raises. Experts rated 900 answers from ChatGPT, Gemini and Claude on an outbreak and found 88.9% correct or mostly correct, with 11.1% potentially misleading. Many will read that as a good score. I read it differently: in a public health emergency, roughly one answer in nine being misleading is a serious failure rate, especially when people rarely check.

My closing thought is a simple one. We are giving these systems more autonomy faster than we are building the institutions, technical or political, to hold them to account. Which one catches up first?