OpenAI CEO Sam Altman described three ways the race for superintelligence could end to employees at a company meeting last week. According to The Information, one scenario ends with a single company “melting everyone else’s graphics processors. Altman spoke about this option with clear distaste. According to him, however, the company has roughly a year to decide which path to take.
The report is based on anonymous sources, but it comes at a time when the debate over artificial intelligence has reached its peak and the shape of new products is being influenced not only by developments in laboratories but also by Washington. OpenAI has not publicly commented on the matter.
International Agency for AI Oversight
The first option is an international body modeled on the International Atomic Energy Agency, IAEA, which would enforce safety standards among the most advanced laboratories. Two people familiar with Altman’s remarks told The Information that the OpenAI CEO clearly favors this option.
This is not a new idea. Back in May 2023, Altman, together with OpenAI President Greg Brockman and then-chief scientist Ilya Sutskever, wrote that the world would probably need something like the IAEA for the development of superintelligence. Such an institution could inspect systems, mandate audits, and monitor how much computing power and energy companies consume. At the time, the authors proposed that companies voluntarily begin complying with some elements of future oversight, with individual countries then incorporating them into their own rules.
Demis’s plan envisions one large CERN
Altman named the second path after Google DeepMind CEO Demis Hassabis. It is based on a proposal Hassabis put forward roughly ten years ago. In Altman’s version, the major laboratories would merge into a single entity or operate under its umbrella, which would then provide models to participating countries.
The model is CERN, the international center for particle physics research. Organizations such as OpenAI, Google DeepMind, and Anthropic would thus be brought under a single shared system. According to reports, Altman sees this option as a way to prevent an uncoordinated arms race.
Eliezer’s plan and melted chips
The third scenario attracted the most attention. Altman named it after artificial intelligence safety theorist Eliezer Yudkowsky. The developer that first achieves recursive self-improvement—that is, artificial intelligence capable of independently improving its own abilities—would use its economic power to cut off all rivals’ access to computing power. This is where the remark about melting competitors’ graphics chips was made. Yudkowsky is known for writings about the so-called decisive act, in which the leading artificial intelligence system would be used to ensure that no one else could build a dangerous model.
Of course, I think this option is terrible, Altman reportedly said, adding that he mentioned it only for the sake of completeness. According to reports, some employees saw the label as a veiled reference to the company’s main rival because Yudkowsky is said to have ties to important people at Anthropic. However, this is the employees’ interpretation, not Altman’s own words.
Washington wants oversight
Altman spoke about the future at a time when the rules for artificial intelligence development are still taking shape. The Trump administration is preparing the first-ever comprehensive set of rules for overseeing superintelligence developers. Under the current approach, which stems from a June executive order, developers are expected to voluntarily provide the government with access to selected models up to 30 days before their release. However, as of October 13, the text of this framework had not yet been published.
Companies have already experienced what such oversight looks like in practice. In June, at the direction of the U.S. government, Anthropic disabled access to its Fable 5 and Mythos 5 models for all customers. The company expressed its disagreement, stating that, in its view, the narrowly defined risk of safeguards potentially being bypassed was not sufficient grounds to withdraw a commercial model. Two weeks later, at the government’s request, OpenAI limited the rollout of GPT 5.6 models to a small group of partners and stated that similar restrictions should not become standard practice.
At the end of September, the heads of Anthropic, Google, Meta, OpenAI, Nvidia, and xAI signed a voluntary agreement that Trump described as morally binding. However, it does not impose any legal liability. According to the report, OpenAI, Google, and Anthropic are also working on a U.S. group called the Standards Authority for Frontier AI, but they disagree over who should test new models. Critics, including Meta’s Mark Zuckerberg and Nvidia’s Jensen Huang, argue that the new rules would merely entrench the position of the largest developers.
It will not work without China
An international agency would be impossible without Beijing. On September 24, Chinese President Xi Jinping told Trump that both countries have the capabilities and responsibility to oversee artificial intelligence so that it remains under human control. However, the U.S. proposal for countries to notify one another about artificial intelligence incidents threatening national security is still only an outline, not a finalized agreement.
According to some people in the industry, there is not much time left. Anthropic co-founder Jack Clark said at a meeting of the International Monetary Fund and the World Bank that only a few years remain for negotiations.
A year to decide
Altman left employees in no doubt that even the ideal plan must not slow OpenAI down. According to the report, he told them that the first two options require extensive cooperation and that OpenAI cannot currently afford to lose its position, especially as both the company and Anthropic prepare to go public. Meanwhile, things are not entirely calm within the company. Following the dismissal of three researchers focused on safety, some employees began to question the company’s approach to safety.
Meanwhile, the rivalry between the two laboratories is also spilling over into finance. According to reports, OpenAI is seeking to raise roughly $30 billion before Anthropic goes public.
Sources: forbes.com and benzinga.com
