"AI progress is uneven across tasks, and success on one benchmark cannot be safely extrapolated to all other capabilities."
Zhou and colleagues examine general-purpose model performance across 63 tasks and 18 capability rubrics, showing how specific task demands determine the strengths and limitations revealed by evaluation. Their research provides independent empirical support for the sidenote that rapid progress in a domain such as coding does not automatically imply equivalent performance across all human activities.
View Source