"This I think is the fundamental difference between reasoning models and the 4.1s and 4 Os of the past where these models could not think. You kind of had to put them in boxes with a predefined set of state transitions whereas here we have the model, the harness, be the whole box."