OpenAI CEO Sam Altman presented employees last week with three possible approaches to safely developing superintelligent AI. According to ChainCatcher, people familiar with the matter said Altman favored the first: creating a global oversight body modeled on the International Atomic Energy Agency to inspect advanced models and audit safety measures.

The second approach would follow the European Organization for Nuclear Research, or CERN, by bringing major AI labs together in one organization or having them conduct joint research through a shared body rather than racing independently. The third, which Altman called the “Eliezer plan,” would involve one AI company achieving recursive self-improvement first and using overwhelming economic resources to control computing power, preventing other developers from catching up. Altman described this scenario as “melting down every other company’s GPUs” and said it was a very bad idea, included only to lay out the range of possibilities.

The GPU scenario draws on a 2021 thought experiment by AI safety researcher Eliezer Yudkowsky, who imagined superintelligent AI creating self-replicating nanomachines to melt GPUs worldwide. Yudkowsky said the example was intended to illustrate an extreme level of intervention, not to advocate carrying it out.

OpenAI proposed an “AI version of the International Atomic Energy Agency” in 2023, while Google DeepMind head Demis Hassabis has also advocated an “AI version of CERN.” OpenAI, Google and Anthropic plan to establish a joint safety standards organization by the end of this year or early next year, but disagree over who should be responsible for model safety testing. Altman wants limits on the pace of AI development, while saying OpenAI cannot currently give up its competitive advantage to rivals.