Alan Chan argues labs should pause AI scaling after loss-of-control incidents

What was said

"If you have a loss of control incident, and what is a loss of control incident? It's like either an agent that gets unauthorized internet access, breaks out of a sandbox, accesses some sort of like internal research cluster, it was not authorized to access like stuff of that extent, then you pause further training, evaluation, deployment of that model and investigate until you're able to solve that to understand if the issue was resolved or not."

Alan Chan argues “responsible scaling” commitments should be made more operational by defining specific “loss of control” incidents that trigger a pause in training, evaluation and deployment during an investigation. He cites examples such as unauthorised internet access, sandbox escapes, or unauthorised access to internal research clusters.