Altman internally discusses three paths for AI safety, the most extreme being "melting down competitors' GPUs"

Beating AI News Flash: OpenAI CEO Altman last week presented to employees three approaches for safely developing super AI:

The first is to establish a global regulatory organization similar to the International Atomic Energy Agency (IAEA) to inspect advanced models and audit safety measures. People familiar with the matter said Altman favors this route.

The second would emulate CERN by merging major AI labs into one organization, or conducting joint research through a unified body, rather than continuing to race separately.

The third is the most extreme, which he called the "Eliezer plan": one AI company would be the first to achieve recursive self-improvement, allowing AI to continuously enhance its own capabilities, and then use overwhelming economic power to control computing power so that other developers cannot catch up. Altman even used the phrase "melt down all the other companies' GPUs," but made clear that this plan is very bad and was only mentioned to lay out all the possibilities.

The "melt down GPUs" idea comes from AI safety researcher Eliezer Yudkowsky. As early as 2021, he proposed a thought experiment: have super AI manufacture self-replicating nanomachines to melt down global GPUs and prevent other organizations from developing dangerous superintelligence. However, he also made clear that this was only an extreme example used to illustrate the intensity of intervention, not a plan he truly advocated carrying out.

The first two routes are likewise not new ideas. OpenAI proposed establishing an "AI version of the IAEA" back in 2023, and Google DeepMind head Hassabis also advocated setting up an "AI version of CERN." Now OpenAI, Google and Anthropic are planning to establish a joint safety standards organization by the end of this year or early next year, but there are even disagreements over who should be responsible for safety testing of models.

Altman hopes to set a "speed limit" for AI development, but also admitted frankly that OpenAI cannot afford to hand over its competitive advantage to rivals right now.