Reading up on Humanity's Last Exam

Parallel evaluates models' web-search ability by combining DeepSearchQA, Humanity's Last Exam, and its WISER benchmark, measuring accuracy, lift, cost, speed, and Pareto efficiency.

AI output quality converges to the mean due to structural limits in next-token prediction, RLHF, and model collapse, making domain expertise more valuable as humans catch AI's subtle errors.