OpenAI CEO Altman proposes three endgames for AI: from IAEA-style global regulation to a single giant "rendering rivals' GPUs obsolete"

OpenAI CEO Sam Altman recently briefed employees on three superintelligence safety development paths under discussion in the industry. The first route is to establish an international organization similar to the International Atomic Energy Agency to uniformly formulate and enforce AI safety standards, which Altman relatively supports; the second references the CERN model, integrating all AI laboratories into a single entity to jointly conduct safety R&D; the third is an extreme option that Altman explicitly does not endorse, namely that one AI company would be the first to achieve recursive AI self-improvement and, relying on strong economic power, cut off competitors' computing power supply, "melt down all rivals' GPUs," and prevent other teams from catching up. This option is also called the "Eliezer plan," and Altman said outright that this idea is very bad and was included in the discussion only for completeness. Altman said internally that there is currently a lack of a feasible AI development "speed limit" plan, and it is necessary to preserve human control over AI and avoid excessive concentration of power in the industry. At present, OpenAI, Google, and Anthropic are preparing to establish a frontier AI safety institute, but the parties still have disagreements on issues such as assessment rights and responsibilities.