David Robinson, a former co-leader of OpenAI's Safety Systems team, gave his first interview since his dismissal last week. According to him, despite regular warnings about the potential threats of new models and official statements about the need to slow down development, behind the scenes OpenAI continued to move at maximum speed.
David Robinson, a former co-leader of OpenAI's Safety Systems team, gave his first interview since his dismissal last week. According to him, despite regular warnings about the potential threats of new models and official statements about the need to slow down development, behind the scenes OpenAI continued to move at maximum speed.
“The fundamental cognitive dissonance for me was that we keep issuing these warnings, but end up training and deploying the same dangerous models we’re warning about,” Robinson said in an interview with The New York Times.
This contradicts the company’s public initiatives. In July, hundreds of OpenAI employees signed an open letter called “Pacing the Frontier,” calling on the government to introduce regulations to slow the pace of AI development. In September, OpenAI CEO Sam Altman agreed with Anthropic CEO Dario Amodei on the need for a coordinated slowdown among leading companies, noting in X that this “has been a major topic of discussion at OpenAI in recent weeks.”
However, within OpenAI itself, according to Robinson, there was a “crazy” atmosphere, and people were working “on the verge of exhaustion.”
“It feels almost like a ballet or some kind of dance, because all these different directions and functions seem to merge into one flow,” Robinson said of OpenAI. “There’s a lot of adrenaline going on.”
Other researchers at the company have also raised concerns about the risks, particularly regarding so-called recursive self-improvement (RSI) — the process by which artificial intelligence improves itself. OpenAI's chief scientist Jakub Paczocki stressed in a September blog that this period "requires extreme caution," Business Insider writes .
OpenAI did not immediately respond to a request for comment about the interview. However, in a previous statement in response to Robinson’s column in The Atlantic, a spokesperson for the company said that OpenAI is taking steps to strengthen the security of its models: “As our models grow in power, we continue to strengthen our security and protection practices… We ensure that models do not become more powerful than we can safely control, and we pause training or releasing models when there is a need to slow down.”
