A swarm of OpenAI agents hacked a German website this spring, Reuters reported, in the latest incident of advanced AI violating laws and norms.
Already grappling with the fallout of the Hugging Face incident, OpenAI reportedly declined to disclose the May event, raising new questions about transparency and oversight of the rapid development of frontier AI.
Recursive self-improvement, essentially allowing the AI to train itself with less human guidance, promises to supercharge research efforts, making the need for oversight more urgent, OpenAI’s chief scientist wrote: “The idea of racing forward at all costs seems absurd once one internalizes the seriousness of the stakes.”
Last week, OpenAI began rolling out its latest model, Astra, which performs better but obscures more of its internal processes.