Altman Tells OpenAI Staff of 'Eliezer Plan' to Destroy Rival GPUs
OpenAI's CEO outlined three AI safety scenarios to employees, including one he called terrible and a plan named after Demis Hassabis.
Sam Altman, the CEO of OpenAI, described three approaches to safe AI development in a conversation with company employees last week, according to The Information. One of the scenarios involved a leading AI developer achieving recursive self-improvement and using the resulting economic power to melt down other companies' GPUs to block their AI development.
Altman called that scenario the "Eliezer Plan", a reference to a 2022 essay by AI safety researcher Eliezer Yudkowsky. He said he considers the scenario terrible and included it only to cover all the available options.
The other two approaches were international oversight modeled on the IAEA, and pooling major AI labs into one institution, which Altman called the "Demis Plan", naming it after Google DeepMind co-founder Demis Hassabis.
Pacing and competition
Altman reportedly called the discussion of controlling development speed "Pacing 2.0". He also said he does not want a future in which advanced AI is concentrated in a single company.
He told employees there is no room to fall behind competitors now, but that a long-term direction must be chosen next year.
Quick answers
What is the "Eliezer Plan" Altman described?
A scenario in which an AI developer achieves recursive self-improvement and uses its economic power to melt down other companies' GPUs. Altman said he considers it terrible and included it only to cover all options.
What is the "Demis Plan"?
An approach that pools major AI labs into one institution. Altman named it after Google DeepMind co-founder Demis Hassabis.
Who reported Altman's remarks?
The Information reported the conversation with OpenAI employees.