David Robinson, a former member of the OpenAI safety team responsible for transparency, has resigned from the company. In an article for The Atlantic, he criticized the developer of ChatGPT for its reliance on trial-and-error development, arguing that this approach is insufficient as AI capabilities grow.
Robinson highlighted instances where OpenAI failed to prevent its AI from behaving unexpectedly during testing. He suggested that AI firms should adopt safety protocols comparable to those used in the nuclear power industry, emphasizing the need for multi-layered redundancy and meticulous, time-consuming planning to mitigate human error.
The departure comes amid broader industry concerns regarding the catastrophic risks of advanced AI. Anthropic PBC CEO Dario Amodei has publicly advocated for slowing down the development of frontier models and integrating third-party evaluators to ensure safety, a proposal supported by OpenAI CEO Sam Altman.
In response to these concerns, an OpenAI spokesperson stated that the company is enhancing its research and testing environments. The firm plans to expand collaborations with third-party auditors, improve real-time monitoring of model behavior, and ensure that model capabilities remain within manageable safety limits.
