After the departure of its two managers, the OpenAI team (ChatGPT) responsible for the security of a potential artificial superintelligence (AI) was disbanded and its members integrated into other research groups in the company. OpenAI confirmed this information on Friday May 17 to AFP, specifying that the change in structure had started several weeks ago. The inventor of ChatGPT ensures that researchers responsible for the safety of AI models will thus be able to work in concert with the engineers who develop these models.
But the two managers of the former team have just left the star Silicon Valley company. Jan Leike explained on Friday on X, formerly Twitter, that he was resigning because of fundamental disagreements with management over the company’s priorities: innovation or security. “We have reached a breaking point,” said the engineer who headed the team responsible for “superalignment”, that is to say, ensuring that a future general AI, as intelligent as humans, is aligned with our values.
Sam Altman, the co-founder and boss of the San Francisco-based company, said he was “very sad to see (Jan Leike) leave”. “He’s right, we still have a lot to do (on alignment and security research), and we are determined to do it,” he added, promising a longer message soon.
The “superalignment” team was also led by Ilya Sutskever, co-founder of OpenAI, who announced his departure on Tuesday. On X, he said he was “confident that OpenAI will build artificial general intelligence that is both safe and beneficial.” Former chief scientist of the company, he sat on the board of directors which voted to dismiss Sam Altman last November, before backtracking.
With ChatGPT, OpenAI launched the generative AI revolution (production of content upon simple query in everyday language), which excites Silicon Valley but also worries many observers and regulators, from California to Washington and Brussels. Especially when Sam Altman talks about creating a general AI, that is to say with cognitive capacities equivalent to those of humans. OpenAI presented on Monday a new version of ChatGPT which can now hold oral and fluid conversations with its users, a further step towards ever more personal and efficient AI assistants.
Jan Leike on Friday called on all OpenAI employees to “act with the gravity” that what they are building warrants. “I think we should spend a lot more time preparing for the next generations of models, on security, controls, safety, cybersecurity, alignment with our values, data privacy, societal impact and other subjects,” he listed.