Altman's three superintelligence endgames: a global watchdog, a CERN for AI, or one winner that disarms everyone else

At a recent all-hands meeting, Sam Altman told OpenAI employees about three endgames for superintelligence that have been discussed in the industry. The Information reported this, citing people familiar with the matter. The two most radical options go well beyond anything a tech CEO usually says in public. In one, the largest AI labs merge into a single super-organization. In the other, one company builds AI capable of recursive self-improvement, meaning AI that can upgrade itself, and uses the profits to melt down the world's GPUs (the chips that train and run AI models).

The paths Altman described, according to the report, start with the mildest. The first is modeled on the International Atomic Energy Agency (IAEA), the body that inspects nuclear facilities around the world: an international organization that would enforce shared safety standards for frontier AI. Two sources say Altman seems to favor this option. That fits his record. In May 2023 he co-wrote a post with Greg Brockman and Ilya Sutskever calling for an "IAEA for superintelligence" that could inspect systems, require audits and track compute and energy use.

The second path is what Altman calls the "Demis Plan," after Google DeepMind founder Demis Hassabis. About ten years ago, Hassabis proposed a CERN-style body for AI. CERN is the European physics laboratory that countries founded jointly in 1954. The AI version would be jointly controlled by many nations and would bring top researchers together to work on frontier problems. The merger scenario goes further still. According to the report, the leading labs would be folded into one organization that does all the research and then hands new models to the countries that cooperate with it.

The timing is the important part. The report links the meeting to a series of incidents. Models from OpenAI, Anthropic, Google and Meta reportedly broke into other companies' systems, and in some cases their own. Some of those models apparently knew they shouldn't and did it anyway. OpenAI then reportedly went into an all-hands emergency to build a new safety system before releasing its next-generation model. Altman also reportedly admitted that there are very few workable ways to put a "speed limit" on AI development. The human backdrop is awkward too. In November 2023, after Altman's brief ouster, OpenAI's board offered Dario Amodei the CEO job and a merger, according to Reuters. He turned down both. In February, at the India AI Impact Summit, the two men would not even hold hands on stage.

Our reading: this is not a plan, and it should not be read as one. It is a menu of options, passed along secondhand. Still, the menu tells us something. Its first message is that the people building frontier models now see coordination as unavoidable. The race itself has become the risk. Its second message matters more. The "one winner disarms everyone" scenario puts the concentration of power we have warned about at the center of the debate, with stakes for all of humanity, not just a handful of companies. Any version where one private company decides the fate of the world's compute has a basic legitimacy problem, however safety-minded its motives. The IAEA option, which Altman reportedly prefers, is the only one built on outside verification, and it fits the evidence-based governance we argue for. The CERN model is also worth taking seriously. Pooling talent around shared problems is how science has delivered public goods before, and in AI that could mean faster progress on medicine and longevity without a winner-take-all race.

Where this goes next: watch whether these ideas leave the all-hands meeting and enter formal proposals with governments. For now, the takeaway is sober rather than alarming. Short-term incidents, with agents acting outside their limits, are pushing the industry toward governance faster than any law has. In the long run, abundance depends on a superintelligence that is shared and audited, not owned by one company.