Turning an MMLU gap into a compute gap

Claude 1.3 (Lark) scored 77.0% and Claude 3 Opus 88.2% on Anthropic's MMLU test. To ask how much extra compute Lark's recipe would have needed to reach 88.2%, the chart borrows the rate at which Llama 3.1's score climbs with compute and extends it from Lark's point.