Demis Hassabis warns A.I.’s uneven reasoning limits progress toward AGI, despite breakthroughs like AlphaFold and Olympiad-level math performance.