Former OpenAI Safety Systems leader David Robinson said OpenAI continued training and deploying models while warning about their capabilities, first reported by Businessinsider. His account juxtaposes OpenAI's public warnings and calls to pace AI development with his description of continued training, deployment and frenetic activity inside the company during that period.
Robinson spoke with Ezra Klein on The New York Times’ “The Ezra Klein Show” in his first interview since quitting OpenAI last week. He told Klein that OpenAI had repeatedly warned about new models’ capabilities and expressed interest in slowing AI development, even as it moved as fast as ever behind the scenes. Robinson described the mismatch as cognitive dissonance.
In July, hundreds of OpenAI employees signed an open letter titled “Pacing the Frontier,” urging the government to impose regulations that would slow AI development. In September, CEO Sam Altman said he agreed with Anthropic CEO Dario Amodei, whose essay advocated that AI companies coordinate a slowdown; Altman also said pacing the frontier had been a primary topic in recent discussions at OpenAI.
The public calls to pace AI development sit alongside Robinson’s account of continuing internal activity. He described activity inside OpenAI’s office as frenetic, saying people were “running on fumes” as different functions streamed together and adrenaline ran high.
OpenAI chief scientist Jakub Pachocki called for extreme caution in a September blog post about recursive self-improvement, or RSI, the process of AI building itself.
In response to an essay Robinson wrote in The Atlantic, an OpenAI spokesperson previously said the company was strengthening safety and security practices as its models became more capable. The spokesperson said those practices address current risks, include safety cases to anticipate risks from future advances, and are intended to keep models from becoming more capable than OpenAI can safely manage and secure. The spokesperson also said OpenAI pauses training or holds back models when it needs to slow down.