More capable models are better at in-context scheming apolloresearch.ai 3 points by JumpCrisscross 19 hours ago