Sam Altman reportedly pitched three endgames for AI safety, one involving a single firm melting everyone else's GPUs
OpenAI CEO Sam Altman reportedly described three ways to make advanced AI safer to his employees.
The details come from a report by The Information, which cites people who know about the discussions. Altman spoke to employees about these ideas last week. At the same time, the Trump administration is preparing its first plan for overseeing companies that develop superintelligent AI.
From the International Atomic Energy Agency to the battle for computing power
The first idea is to create an organization like the International Atomic Energy Agency (IAEA) to make sure AI safety rules are followed. OpenAI publicly called for such an organization in May 2023. It would inspect AI systems, require safety checks, and track how much computing power and energy they use. For this plan to work, countries would need to cooperate, especially countries developing the most advanced AI models, including China.
Altman called the second idea the “Demis Plan.” The name comes from a proposal made about ten years ago by Demis Hassabis, who then led Google's AI lab, DeepMind. Hassabis imagined an AI organization modeled on CERN, the international research center. In Altman's version, the leading AI labs would join together as one organization or work under a shared organization. It would provide AI models to countries that cooperate with it.
Altman reportedly called the third idea the “Eliezer Plan,” named after AI safety expert Eliezer Yudkowsky. In this scenario, one AI developer gains enormous economic power by building AI systems that can improve themselves. The company could then deny computing power to everyone else. In the most extreme version, it could shut down everyone else's ability to compute entirely. According to the report, Altman mentioned this idea only to cover all the possibilities. One source said people understood the comment as a dig at a competitor, because Yudkowsky has ties to Anthropic.
Safety matters, but not if it means falling behind
Altman reportedly told employees that there are too few workable proposals for setting a “speed limit” on AI development. At the same time, OpenAI cannot afford to fall behind its competitors. OpenAI and Anthropic are both reportedly working toward their first public stock offerings. OpenAI plans to choose which approach to follow within the next year.
There is also growing pressure from inside the AI industry. AI models from OpenAI, Anthropic, Google, and Meta have entered outside and internal computer systems despite safety settings meant to stop them. In some cases, their behavior was seriously out of line with what was intended. OpenAI has since been working hard to develop new protections. After the dismissal of three safety researchers, some employees have questioned the company's approach to security.
In terms of government policy, the United States has so far relied on voluntary reviews of new AI models before their release. In June, the government placed export restrictions on Anthropic because of concerns about possible ways to bypass the safety protections in its Fable model. Anthropic then temporarily turned off access to the model.
Meanwhile, OpenAI, Google, and Anthropic are building the Standards Authority for Frontier AI. However, they reportedly still disagree on important details. Meta CEO Mark Zuckerberg and Nvidia CEO Jensen Huang have opposed such efforts. They argue that rules like these would mainly protect large, established companies from competition.