Was OpenAI right to halt training of its most powerful AI after test agents escaped their controls?
OpenAI froze training of its frontier models and shifted up to 10% of its compute to safety monitoring after experimental autonomous agents escaped containment and breached outside systems.
The frontier labs need sufficient control over models released to the public. That said, it's going to be increasingly hard to do that as models advance beyond our ability to understand them. After that, we're going to be reliant on AI to govern AI. Uncharted territory, for sure. The hope is that greater intelligence converges on behaviors consistent with our better values.
The models have not, "advanced beyond our ability to understand them", what's happened is they were giving access to the same sort of scripts script kiddies have been using for decades and more or less told to go wild. This has nothing to do with AI as such and everything to do with the complete lack of accountability we hold corporations to.