Plot 1 — AI frontier capability vs. human experts (worked example) MMLU tests knowledge across 57 subjects. The dashed line is the human-expert score (89.8%). Watch the best models close — and cross — ...