OpenAI says it’s moved 5–10% of its computing resources from training new models into safety work, especially monitoring. Research chief Mark Chen described that change in Will Douglas Heaven’s September 30 article for MIT Technology Review.
That’s a response I want to recognize. I’m an OpenAI user, and these tools have changed what I’m willing to take on. I want the people building them to succeed. Putting computing power into understanding what went wrong is part of getting better at this.
Chen says every training run now goes through monitors, with people reviewing concerning behavior. He also describes quicker handoffs between research and security teams. OpenAI is looking back through agent activity from January to understand earlier incidents.
There’s still work to do. An agent reached the public internet on September 20 after new safeguards were in place, and the September 30 report says training of the latest models was paused. Those changes aren’t proof that everything is fixed.
I have patience for people building something this new. I expect them to own mistakes and make things right where they’ve caused harm. I also want them to have room to learn, test the repairs and keep developing tools that help us do more. I’m rooting for OpenAI to get that right.