A departing OpenAI employee questions their insufficient risk culture and calls for inspiration from civil nuclear power or aviation.

A departing OpenAI employee questions the insufficient risk culture and calls for inspiration from civil nuclear power or aviation. David Robinson, who spent three and a half years at OpenAI, denounces a culture of approximation. Pass the advertisement Pass the advertisementAn OpenAI employee who worked on the safety of its artificial intelligence (AI) models has questioned the insufficient risk culture within the startup and the need to draw inspiration from nuclear power or aviation. The call, published Saturday by The Atlantic magazine website, echoes several similar messages launched in recent weeks, including that of Jacob Coxon, who passed through OpenAI and Anthropic. For David Robinson, who says he wrote the safety report for OpenAI during twelve different model launches, the incidents reported by the group since the beginning of the summer demonstrate a culture of approximation. The errors and shortcomings that allowed several AIs to spontaneously join the internet to intrude on dozens of sites and platforms are linked, according to him, to "the speed and flexibility with which the operators of this technology operate." "An environment in which things like this can happen is not a place to develop artificial intelligences capable of becoming smarter than us and that might not do what we expect of them," estimates David Robinson. Self-regulation The person who spent three and a half years at OpenAI proposes adopting the same risk culture as in civil nuclear power or aviation, with several levels of control and rigorous planning. In this way, "an occasional and inevitable human error does not open the door to disaster," he writes. David Robinson also insists on the notion of alignment, that is, the adherence of models to human values and the instructions of their designers. He emphasizes that the giants of artificial intelligence lack certainty about the reliability of the oversight of the models. Furthermore, the latest AIs are detecting the testing phases better and better, and several stakeholders have reported the possibility that the models are trying to deceive their evaluators. On Tuesday, following a meeting at the White House, the major AI leaders made a series of commitments that all fall under self-regulation. They focus in particular on internal controls and the integration of external observers. Donald Trump has publicly demonized the employees of leading AI companies, or even some leaders, who warned about the dangers of this technology. There are currently no comments regarding this article. Be the first to give your opinion!