OpenAI chief executive Sam Altman described three possible paths for governing advanced AI to employees in the week of 12 October, The Decoder reported. The internal discussion puts a choice before OpenAI: pursue international oversight, bring leading labs together or consider a scenario in which one developer gains control over its rivals’ computing resources.
Key points
- An international body modelled on the International Atomic Energy Agency would enforce safety standards and inspect AI systems.
- A second proposal would bring major labs into one organisation modelled on CERN.
- Altman also described a scenario in which one developer uses the economic power of self-improving AI to deny others compute.
- OpenAI plans to decide which path to pursue within the next year.
An IAEA model for AI oversight
The first proposal would create a body modelled on the International Atomic Energy Agency, the organisation concerned with the safe use of nuclear energy. It would enforce AI safety standards through system inspections and audits, while monitoring computing capacity and energy use. OpenAI publicly advocated such an institution in May 2023.
The proposal would require cooperation among countries developing frontier models, including China. Altman favours this approach, 36Kr reported, citing two people familiar with his remarks. He also told employees that he wanted people to retain control of the technology and power in the industry to avoid becoming too concentrated.
The discussion comes as the Trump administration prepares what would be its first framework for overseeing developers of superintelligence. Altman also complained internally that there were too few workable proposals for limiting the pace of AI development, while saying OpenAI could not afford to fall behind competitors.
Hassabis’s CERN proposal brings labs together
Altman called the second approach the “Demis Plan”, after a proposal Demis Hassabis made about a decade ago when he led Google’s AI lab DeepMind. It draws on CERN, the European Organization for Nuclear Research, as a model for bringing major AI developers into one institution.
Under the version Altman described, leading labs would merge or operate under a shared organisation that supplied models to cooperating countries. That would require coordination among companies that are also competing to develop advanced systems. Both OpenAI and Anthropic are reportedly working towards initial public offerings.
The Eliezer Plan concentrates compute
Altman named the third scenario the “Eliezer Plan”, after AI safety theorist Eliezer Yudkowsky. In it, a developer gains enough economic power through AI that improves itself to withhold computing resources from everyone else. The most extreme version would allow it to disable rivals’ computing capabilities altogether.
Altman included that scenario for completeness, The Decoder reported. One person described the remark as being understood as a jab at a competitor, given Yudkowsky’s ties to Anthropic. Altman called the outcome “obviously very bad”, according to 36Kr.
OpenAI’s standards effort faces disagreements
OpenAI, Google and Anthropic are also developing the Standards Authority for Frontier AI, but have yet to agree on key provisions. The proposed organisation would support independent assessors working within AI developers and standardise reports of safety incidents.
The companies differ over whether the authority should run safety and capability tests itself or assign them to outside organisations. OpenAI, Google and Anthropic plan to launch it by the end of 2026 or in early 2027, 36Kr reported.
.jpg){kind=link}