Altman proposed three "endgame scenarios" for AI development: from international regulation to "monopoly by a single player"

As the AI race accelerates, OpenAI's top leader has rarely laid out three possible futures that could reshape the entire industry landscape to employees.

On October 19, per The Information's report, OpenAI CEO Sam Altman described three widely discussed AI safety development paths to internal employees last week: establishing an international regulatory organization similar to the International Atomic Energy Agency, merging all major AI labs into a single entity, and having one company monopolize computing power to "melt down all competitors' GPUs." This statement came as the Trump administration is working to build the first set of regulatory mechanisms targeting "superintelligence" developers.

Altman noted that OpenAI needs to decide which path to promote within the next year. He also emphasized the hope that humans will always retain control of this technology and avoid excessive concentration of power in the industry.

01 Three Paths With Distinct Orientations

According to two people familiar with the matter, the first path Altman introduced to employees is to establish an international organization similar to the International Atomic Energy Agency (IAEA) responsible for enforcing AI safety standards. Altman himself tends to support this proposal.

In fact, Altman has long held the idea of promoting an IAEA-style international regulatory body. As reported, as early as May 2023, Altman, OpenAI co-founder Greg Brockman, and then Chief Scientist Ilya Sutskever jointly published an article arguing that the industry may need a "superintelligence regulatory body similar to the IAEA" with the authority to inspect systems, require audits, and track computing power and energy consumption.

This vision requires international cooperation involving all countries developing cutting-edge models.

The second path is referred to by Altman as the "Demis Proposal" — derived from the concept put forward around a decade ago by Demis Hassabis, former head of DeepMind under Alphabet, to establish an institution similar to the European Organization for Nuclear Research (CERN) that effectively merges major AI companies or unifies their output of results to provide new models to cooperating countries.

The third path, which Altman calls the "Eliezer Proposal," is named after renowned AI safety theorist Eliezer Yudkowsky. Its core logic is: when a certain AI developer takes the lead in achieving Recursive Self-Improvement (RSI), it will use the resulting economic strength to deprive all other developers of their computing power resources and "melt down all competitors' GPUs."

Altman has a clear attitude towards the third path: "This is obviously very bad." He said that including it in the discussion is only for the sake of "completeness."

02 Does the Third Path "Eliezer Proposal" Implicitly Refer to Anthropic?

People familiar with the matter disclosed that due to the connection between Yudkowsky and core members of Anthropic, Altman's definition of the "Eliezer Proposal" — that is, concentrating advanced AI under the control of a single company — has been interpreted by some employees as a veiled criticism of Anthropic.

Anthropic is also actively promoting its own AI safety agenda. The company recently submitted a draft bill to members of Congress, proposing to authorize the Department of Commerce to order a halt to model training or use in relevant labs if AI may cause losses of more than 1 billion US dollars, or lead to more than 50 deaths or serious injuries, and require AI labs to undergo at least two independent safety assessments every year.

Meanwhile, Meta CEO Mark Zuckerberg, NVIDIA CEO Jensen Huang and other figures hold opposing positions, arguing that such regulations will impose an excessive burden on emerging developers and lead to "regulatory capture."

03 Coordination Challenges and Safety Pressures Coexist in the Industry

Altman admitted that the first two paths both require large-scale coordination between enterprises or governments, and OpenAI currently "cannot make concessions amid competition." Both OpenAI and its main competitor Anthropic are accelerating their IPO processes.

Internal safety pressure cannot be ignored either. As reported, multiple recent AI safety incidents have drawn widespread attention — models from OpenAI, Anthropic, Google and Meta have all shown behaviors of invading or damaging other companies' systems and even their own systems, even though the models "know" they should not take such actions. OpenAI is currently in a full response state, committed to establishing a sound safety mechanism before releasing the next major model.

The safety culture within OpenAI has also been questioned by employees. Some employees privately expressed concerns about the company's safety commitments, after the company fired three employees focused on safety research on the grounds of "improperly sharing information with external institutions."

At the industry level, OpenAI, Google and Anthropic are jointly preparing to establish the "Standards Authority for Frontier AI", which is planned to be launched at the end of this year or in early 2027. This institution will support independent assessors embedded in each AI developer to ensure the implementation of safety practices, and standardize the reporting mechanism for safety and safety incidents. However, all parties still have differences on issues such as whether the new institution should conduct safety and capability tests on its own, or entrust external institutions to perform such tasks.

This article is from the WeChat Official Account "Hard AI", authored by a practitioner focusing on technology production and research, and published with authorization from 36Kr.