“Sander Schulhoff advises that explicitly using thought-inducing phrases like "chain of thought" is still necessary to ensure robust performance at scale for models like GPT-4 and GPT-4o.”