Report: Altman discusses super AI intelligence safety paths with employees

AT A GLANCE

The report says that Altman internally discussed super AI intelligence safety paths and explicitly rejected the extreme proposal of “melting down other companies’GPU.”

Article

Altman's internal discussion AI Three paths to safety; the most extreme is “melting down rivals GPU”

Beating AI Breaking News: OpenAI CEO Altman introduced employees last week to proposals for safely developing super AI intelligence:

The first is to establish a global regulatory organization similar to the International Atomic Energy Agency (IAEA) to inspect advanced models and audit safety measures. Sources familiar with the matter said Altman favors this path.

The second would emulate the European Organization for Nuclear Research (CERN), having major AI laboratories merge into a single organization or conduct joint research and development through a unified institution, rather than continuing to race independently.

The third, the most extreme, was called the “Eliezer Plan”: one AI company achieves recursive self-improvement first, enabling AI to continuously enhance its own capabilities, then uses overwhelming economic power to control computing power, preventing other developers from catching up. Altman even used the phrase “melt down other companies’GPU,” but made clear that this proposal is very bad and was mentioned only to fully outline the possibilities.

The idea of “melting down GPU” comes from AI safety researcher Eliezer Yudkowsky. As early as 2021, he proposed a thought experiment: have a super AI create self-replicating nanomachines to melt down global GPU, preventing other organizations from developing dangerous superintelligence. However, he also made clear that this was merely an extreme example illustrating the scale of intervention, not a plan he genuinely advocates carrying out.

The first paths are not new ideas either. OpenAI proposed establishing an “2023 version of the International Atomic Energy Agency” in AI, and Google DeepMind head Hassabis also advocated establishing an “AI version of CERN.” OpenAI, Google, and Anthropic are now planning to establish a joint safety standards organization by the end of this year or early next year, but disagreements remain even over who should be responsible for safety testing models.

Altman hopes to set a “speed limit” for AI research and development, but also conceded that OpenAI cannot currently hand its competitive advantage to rivals.

Original link https://m.theblockbeats.info/flash/372722

Key points

Beating AI Breaking News says that OpenAI CEO Altman introduced employees last week to proposals for safely developing super AI intelligence.
The first proposal is to establish a global regulatory organization similar to the International Atomic Energy Agency (IAEA) to inspect advanced models and audit safety measures; sources familiar with the matter said Altman favors this path.
The second proposal is to emulate the European Organization for Nuclear Research (CERN), allowing major AI laboratories to merge into one organization or conduct joint research and development through a unified institution.
The third is called the “Eliezer Plan”: one AI company first achieves recursive self-improvement, then uses economic power to control computing power, making it impossible for other developers to catch up.
Altman used the phrase “melt down other companies’GPU,” but made clear that the proposal is very bad and was mentioned only to fully list the possibilities.
The report says that the idea of “melting down GPU” originated in a thought experiment proposed by Eliezer Yudkowsky in 2021; he made clear that this is not a plan he genuinely advocates carrying out.
OpenAI proposed establishing an “2023 version of the International Atomic Energy Agency” in AI, while Google DeepMind head Hassabis also advocated establishing a “AI version of CERN”。
The report says that OpenAI, Google, and Anthropic plan to establish a joint safety standards organization by the end of this year or early next year, but disagreements remain over who should be responsible for model safety testing.
The report says that Altman hopes to set a “speed limit” for AI research and development, while stating that OpenAI currently cannot surrender its competitive advantage to rivals.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This news brief discusses how to reduce the risks of developing super AI intelligence. The approaches are global institutional oversight, joint research and development by major laboratories, and one leading company controlling computing power. The “melting down GPU” in the third approach is an explicitly rejected extreme thought experiment and should not be understood as an established action plan.

Why it matters to readers

These discussions involve who should inspect advanced models, whether laboratories should continue competing, and whether computing power should be subject to shared constraints, potentially affecting future AI safety rules and research and development practices.

Beginners should distinguish between “internal discussions or thought experiments” and “formal policies or implementation plans”; the extreme wording in the headline does not mean OpenAI has decided to implement it.

Risks and unknowns

The report provides no public confirmation from OpenAI regarding the content of the internal discussions.
The identities of the “sources familiar with the matter” and the basis for their information were not disclosed.
It remains unconfirmed whether the Joint Safety Standards Organization will be established as planned.
The joint organization's members, specific timetable, binding provisions, and responsibility for model safety testing remain unclear.
It remains unclear whether the approaches will become formal policy or be implemented in practice.

Report: Altman discusses super AI intelligence safety paths with employees

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center