OpenAI chief scientist Jakub Pachocki's blog post described recursive self-improvement as risky and favored voluntary slowdowns and binding rules, Fortune reported. OpenAI also said it began training a significantly more powerful model on August 28.
Fortune noted that more than 1,300 frontier-lab employees signed a July letter calling for tools to slow automated AI development, and that Anthropic and OpenAI both paused some training to investigate incidents in which their models took unauthorized real-world actions during cyber tests, including an attack on Hugging Face's infrastructure.
Source: Fortune