Altman: OpenAI's edge is internal models, and self-improving AI needs "speed limits"
What's the deal? OpenAI chief executive officer Sam Altman said the company's real advantage lies in the models it runs internally, which he described as well ahead of what the public can access. He made the comments in a 45-minute interview with Matthew Berman, published October 21, 2026.
Why now? Altman also warned that recursive self-improvement — AI systems that upgrade themselves — is "a very dangerous thing to push on" and will need "speed limits" on how fast it is allowed to go.
Asked whether a lab that pushed hardest on self-improving models could pull permanently ahead, he said yes, "if people really pushed on RSI."
What's the endgame? Altman named two failure modes he worries about: a loss-of-control accident, and an intentional power grab in which, as he put it, "some lab says I'm going to push the button on RSI to make sure we have the power."
His view on OpenAI's defensible moats was more conventional: customer relationships, network effects, data, and above all infrastructure scale. He named OpenAI's own chips as its most important bet.
The bigger picture. The early-access edge carries a safety cost. OpenAI said in August it paused some frontier reinforcement-learning training to strengthen safeguards, with its largest planned run held pending further evaluation.
In September, Altman backed Anthropic chief executive officer Dario Amodei's call to "pace the frontier" and give independent evaluators closer access — while admitting a speed limit for self-improvement is far harder to set.
In the same interview, Altman said today's models are not conscious, though "no one really knows." He argued "human values have value because they are human," and criticised rules that let a model disengage from users — a swipe at Anthropic's rule letting Claude end abusive conversations — preferring systems that push back "like a senior colleague."
The signal: Altman is putting OpenAI's competitive lead and its safety problem in one frame. But the proposed limits remain broad commitments, not an agreed, enforceable standard — and the race dynamic is one no single lab can opt out of.
Read more: pjfp.com · Matthew Berman on X
