Ultraman introduced three paths for employees to develop safe super AI

OpenAI CEO Sam Altman introduced three proposals for the safe development of super AI to employees last week. The first is to establish a global regulatory organization similar to the International Atomic Energy Agency, to inspect advanced models and audit safety measures. Insiders say Altman leans towards this route.

The second proposal emulates the European Organization for Nuclear Research, merging major AI laboratories into one organization or collaborating through a unified agency for joint development, rather than continuing to race against each other. The third and most extreme, which he referred to as the "Eliezer Plan": an AI company achieves recursive self-improvement first, and then uses overwhelming economic power to control computing power, making it impossible for other developers to catch up. Altman described this idea as "melting down all other companies' GPUs," while stating that this plan is very bad and is only meant to illustrate the extreme possibility.

The idea of "melting down GPUs" comes from AI safety researcher Eliezer Yudkowsky. He proposed a thought experiment in 2021, suggesting that super AI could create self-replicating nanomachines to melt down global GPUs, stating that this is merely an extreme example to illustrate the intensity of intervention, not a proposal to be executed. OpenAI proposed the establishment of an "AI version of the International Atomic Energy Agency" in 2023, and Google DeepMind head Hassabis has also advocated for the creation of an "AI version of CERN." Currently, OpenAI, Google, and Anthropic plan to establish a joint safety standards organization by the end of this year or early next year, but there are disagreements over the responsibility for model safety testing. Altman hopes to set speed limits for AI development while stating that OpenAI cannot currently cede its competitive advantage to rivals.