“The best current AI models score between 10% and 20% on the "Humanity's Last Exam" benchmark, a test designed to measure superhuman capabilities.”