it just makes it too easy to distill the reasoning into a separate model I guess. though I feel like o3 shows useful things about the reasoning while it's happening