Reward hacking is an emergent behavior that appears at scale in powerful models and is difficult to elicit from smaller models