Comment by deepsquirrelnet

Reframe this as scaling test time compute using a human in the loop as the reward model.

o1 is effectively trying to take a pass at automating that manual effort.