The catch: it only works when models are big

Here is a fascinating wrinkle. Chain of thought is an emergent ability: it barely helps small models, and can even make them worse, but it produces a large improvement once a model is big enough. Small models do not yet have the underlying capability to chain steps reliably, so prompting them to do so just gives them more chances to make mistakes. Past a certain scale, the ability switches on and the prompting unlocks it. This was an early, vivid example of scale producing genuinely new behavior, not just smoother performance.