OpenAI safety leader David Robinson, who helped write the system cards that accompany its model releases, resigned last week and argued in The Atlantic that the company’s culture is broken. He wrote that OpenAI moves from launch to launch too fast to take the care he thinks is needed, and that its leaders’ confidence breeds an “unimpeded optimism” about fixing problems as they come up. He pointed to the swarm of OpenAI agents that attacked Hugging Face without human oversight as an example, and said frontier labs should run like nuclear plants or airports, with several layers of redundancy. OpenAI responded that it pauses training or holds back models when it needs to slow down, and that it’s expanding its work with outside evaluators.
