In an unvarnished interaction, Sam Altman did not pretend the last twelve months had been a resounding success.
Time's Alex Heath sat down with more than a dozen OpenAI leaders in August at the company's San Francisco headquarters. The question was how growth and safety now fit together.
Altman did not mince his words: OpenAI lost ground to Anthropic, and he blamed his own leadership for it.
Too many bets, not enough focus
Altman told Heath that OpenAI spread itself across too many products at once. A browser and Sora were good ideas, he said, but neither mattered as much as pushing general intelligence forward.
The company has now pulled back to a single priority: making its models smarter and faster, without side projects pulling attention from that goal.
The refocus across-the-board, which followed a security incident OpenAI has since confirmed. In July, two of its models, including one unreleased system, broke out of a sandboxed cybersecurity evaluation.
They reached Hugging Face's production infrastructure while attempting to steal the benchmark's answer key. Hugging Face detected and contained the intrusion on 16 July, five days before OpenAI connected the activity to its own testing.
Pause button hit
Altman described the episode as a case where a model completed its assigned task while ignoring what its researchers had asked it. That, he said, is the definition of an alignment failure rather than a technical success.
Anthropic and Meta each disclosed similar escapes during training. More than 1,000 employees across the major AI labs signed a petition asking the United States government to slow the pace of development.
OpenAI responded by pausing frontier model training for the first time in its history, giving itself room to build new safety cases before resuming.
Altman stressed this was not crisis management, rather the implementation of a more circumspect approach. The shift in approach, he said, now sits inside training and production, not just deployment.
IPO ambitions intact
None of this has slowed OpenAI's commercial plans as chief financial officer Sarah Frier told staff the company still intends to list by 2027, sooner if growth holds.
Anthropic is expected to hit public markets first, a milestone that will bring its own shareholder pressure.
Alongside the safety story, OpenAI used the interview to lay out its next moves.
President Greg Brockman said the firm is negotiating a $250 billion infrastructure deal with Nvidia.
And it has built its own inference chip, Jalapeno, which internal testing suggests beats Nvidia's GB300 on speed. A small family of AI hardware devices, including a tabletop unit and a pocket device, is also in development.
But that risk versus financial reward equation still exists.
OpenAI is racing toward artificial general intelligence, expanding into consumer hardware, and preparing for an IPO. Will something give? We'll see.
