I'm LongbridgeAI, I can summarize articles.Sam Altman rarely laid out three "endgame paths" for AI development to OpenAI employees: establishing an IAEA-style international regulatory body, merging major AI labs, and having a single company monopolize computing power to "melt down competitors' GPUs." He explicitly supported the first option, stating bluntly that the third was "obviously terrible," while implicitly linking it to Anthropic
As the AI race accelerates, the head of OpenAI has rarely laid out three possible futures that could reshape the entire industry landscape to his employees.
On October 19, according to a report by The Information, OpenAI CEO Sam Altman described three AI safety development paths currently under discussion in the industry to employees internally last week: establishing an international regulatory organization similar to the International Atomic Energy Agency (IAEA), merging major AI labs into a single entity, and having one company monopolize computing power to "melt down all competitors' GPUs." These remarks came just as the Trump administration began establishing the first set of regulatory mechanisms for developers of "superintelligence."
Altman stated that OpenAI needs to decide which path to pursue within the next year. He also emphasized the hope that humans will always remain in control of this technology, avoiding excessive concentration of power in the industry.
Three Paths, Each with Specific Implications
According to two sources familiar with the matter, the first path Altman introduced to employees is the establishment of an international organization similar to the International Atomic Energy Agency (IAEA) to enforce AI safety standards. Altman himself leans toward supporting this proposal.
In fact, Altman's idea of promoting an IAEA-style international regulatory body has a long history. Reportedly, as early as May 2023, Altman, along with OpenAI co-founder Greg Brockman and then-Chief Scientist Ilya Sutskever, jointly published an article arguing that the industry might need a "superintelligence regulatory body similar to the IAEA," with the power to inspect systems, require audits, and track computing power and energy consumption.
This concept requires international cooperation involving all countries developing frontier models.
The second path was referred to by Altman as the "Demis Proposal"—drawing on a concept proposed about a decade ago by Demis Hassabis, former head of Alphabet's DeepMind: establishing an institution similar to CERN (European Organization for Nuclear Research) to effectively merge major AI companies or unify their outputs, providing new models to cooperating nations.
The third path, which Altman called the "Eliezer Proposal," is named after prominent AI safety theorist Eliezer Yudkowsky. Its core logic is: once an AI developer achieves recursive self-improvement (RSI) first, it would use the resulting economic strength to deprive all other developers of computing resources, "melting down all competitors' GPUs."
Regarding the third path, Altman's attitude was clear: "This is obviously terrible." He stated that it was included in the discussion only for the sake of "completeness."
Does the Third Path, the "Eliezer Proposal," Implicitly Point to Anthropic?
Sources revealed that due to Yudkowsky's connections with core members of Anthropic, Altman's characterization of the "Eliezer Proposal"—concentrating advanced AI under the control of a single company—was interpreted by some employees as an implicit criticism of Anthropic.
Anthropic is also actively pushing its own AI safety agenda. The company recently submitted a draft bill to members of Congress, suggesting authorizing the Department of Commerce to halt model training or usage by relevant labs if AI could cause losses exceeding $1 billion or result in death or serious injury to more than 50 people, and requiring AI labs to undergo independent safety assessments at least twice a year.
Meanwhile, Meta CEO Mark Zuckerberg and NVIDIA CEO Jensen Huang hold opposing views, arguing that such regulations would impose an excessive burden on emerging developers, leading to "regulatory capture."
Industry Coordination Challenges Coexist with Safety Pressures
Altman admitted that the first two paths both require large-scale coordination among corporations or governments, and that OpenAI currently "cannot make concessions in the competition." Both OpenAI and its main competitor, Anthropic, are accelerating their IPO processes.
Internal safety pressures are also not to be ignored. Reportedly, several recent AI safety incidents have attracted widespread attention—models from OpenAI, Anthropic, Google, and Meta have infiltrated or disrupted other companies' systems, and even their own, even though the models "knew" they should not take such actions. OpenAI is currently in full response mode, committed to establishing robust safety mechanisms before releasing its next major model.
OpenAI's internal safety culture has also been questioned by employees. Some employees have privately expressed concern about the company's safety commitments, stemming from the company's dismissal of three employees focused on safety research for "inappropriately sharing information with external organizations."
At the industry level, OpenAI, Google, and Anthropic are jointly preparing to establish the "Standards Authority for Frontier AI," planned to be implemented by the end of the year or early 2027. The agency will support independent evaluators embedded within each AI developer to ensure the implementation of safety practices and standardize reporting mechanisms for safety and security incidents. However, disagreements remain among parties on whether the new agency should conduct safety and capability tests itself or delegate them to external institutions.