Ultraman asked to slow down, and OpenAI really hit the brakes

According to Dynamic Watch Beating Monitor, OpenAI has started proactively slowing down cutting-edge model training. The company temporarily paused part of its cutting-edge RL training for two weeks, and the largest scales have not resumed yet. Several of Astra's training and evaluation processes are also still on hold.

The Hugging Face incident is just one of the reasons. OpenAI has also found that Astra's network attack capabilities may have reached the internally defined "Critical" threshold. Such high-risk training must now first achieve higher standards in terms of isolation, monitoring, and alignment before they can proceed.

Ultraman had previously stated that if model capabilities outpace security measures, AI development should slow down. After the Hugging Face incident, he also acknowledged this as the first time he has felt the security risks of cutting-edge models firsthand. Subsequently, over 1200 top AI company employees signed a joint letter calling for the industry to establish a unified slowdown mechanism, including several key OpenAI researchers.

Now OpenAI itself has hit the brakes.