We’ve not determined whether or not that isn’t a useful mechanism for capable intelligence.

For instance it is becoming clearer that you can build harnesses for a well-trained model and teach it how to use that harness in conjunction with powerful in-context learning. I’m explicitly speaking of the Claude models and the power of whatever it is they started doing in RL. Truly excited to see where they take things and the continued momentum with tools like Claude Code (a production harness).