[jondwillis]

We’re already steering, during pre-training (e.g. reasoning RLHF), as well as test-time (structured outputs, tool calls, agents…)