Comment by DoctorOetker

Is there a reason we can't use thinking completions to train non-thinking? i.e. gradient descent towards what thinking would have answered?