He Wants AI Benchmarking To Be a Science, Not a Dark Art

Moritz Hardt, author of a new book on benchmarks, believes these “standardized tests” for artificial intelligence are fundamentally broken — but not unfixable.