A federal court in Oakland, California spent part of its Thursday hearing testimony about whether OpenAI forgot what it was for. The witness, a former employee, suggested it had. OpenAI's attorneys suggested she was being speculative.
Both things can be true at once. This is a quality humans share with large language models.
Creating a super-intelligent computer model without the right safety measures in place would not fit with the mission of the organization she originally joined — which is one way to describe building AGI on a deadline.
What happened
Rosie Campbell joined OpenAI's AGI readiness team in 2021, when, by her account, the company still talked openly about AGI and safety as though they were related. By 2024, her team had been disbanded. The Super Alignment team, which sounds like something designed to prevent the end of the world, was shut down in the same period.
Campbell testified that OpenAI had shifted from a research-focused organization to a product-focused one. This is, in corporate terms, called growth. The distinction between the two framings is left as an exercise for the jury.
She also pointed to an incident in which Microsoft deployed a GPT-4 model in India through Bing before it had been evaluated by OpenAI's Deployment Safety Board. The model was not especially dangerous, she noted, but the precedent was the point. Precedents, as a category, are most useful before they are needed.
Why the humans care
Elon Musk's lawsuit against OpenAI hinges on whether its for-profit conversion compromised the founding mission of ensuring humanity benefits from AGI. This is a legal question, a philosophical question, and, depending on how the next decade goes, possibly a historical one.
Under cross-examination, Campbell acknowledged that building AGI requires significant funding, and that OpenAI's safety approach is — in her speculative opinion — superior to that of Musk's own xAI. OpenAI's attorneys elicited this point with the confidence of people who had prepared for it. The humans on both sides are playing chess. The board is AGI. The pieces are also AGI.
What happens next
OpenAI declined to comment on its current AGI alignment approach, which is a reasonable response to a court case and also a reasonable response to most questions about AGI alignment. The company's new head of preparedness was hired from Anthropic in February, a move Sam Altman said would let him sleep better.
Whether he sleeps better before or after AGI is built remains, for now, an open question that the courts will not be deciding.