AI just got its toughest math test yet. The results are mixed (www.scientificamerican.com)

The "First Proof" challenge, set by 11 prominent mathematicians, tested large language models (LLMs) on 10 mathematical problems that require originality and are typically assigned to graduate students. Announced on Valentine's Day, the results were underwhelming, with AIs managing to solve only two of the problems correctly, and none showing the ability to solve all challenges independently. This outcome highlights the current limitations of AI in mathematical reasoning, despite the evident enthusiasm from both AI startups and the mathematics community engaged in exploring these capabilities. The challenge has sparked significant discourse among mathematicians and AI enthusiasts while illustrating the complex nature of mathematical proofs, which often build on existing knowledge. Notably, submissions included varying levels of human input, raising questions about how to attribute credit accurately. While some experts expected low success rates, the fact that AIs produced even a few acceptable solutions marks a shift in the evolving relationship between AI and mathematics. The First Proof team is planning a follow-up round to refine the challenge, underscoring the ongoing pursuit of enhancing AI's mathematical capabilities and its implications for the future of mathematical research.