Most complex AI benchmark launched
A new benchmark HUMANITY’S LAST EXAM has been introduced, featuring 3000 difficult questions across dozens of subject areas. Questions were selected through a multi-stage process.
From 13000 proposed questions where leading AI models showed poor results, experts selected 3000, modifying them to ensure quality and clarity.
Authors of the top 50 questions received 5000 dollars each. The next 500 questions earned their creators 500 dollars each. Benchmark leaders – o 1 and R 1 show results below 10%. R 1 leads in the text portion but cannot process images, which make up 10% of the test.
HUMANITY’S LAST EXAM aims to assess AI’s capability limits, as existing tests have been mastered by models with over 90% accuracy. Initial results are shocking: even GPT-4 o showed only 3.3% accuracy, with the best result at 9.4%.
The benchmark also evaluates model self-calibration – their ability to assess confidence in their own answers. R 1 leads significantly, but calibration error still exceeds 80%.
Authors expect new models might reach 50% accuracy on this challenging new test by year-end. Apparently, to beat AI in testing, it’s enough to pay people to create truly difficult questions.
