Simons Institute for the Theory of Computing

The speaker argues that scaling inference-time compute (i.e., "thinking") is a more powerful and efficient path to advanced AI reasoning than simply scaling pre-training compute and model size. Dra...