"The 'First Proof' project tested four AI systems on ten research-level math problems. The best-performing model, from ETH Zurich, solved six out of ten problems. The test was designed to prevent models from relying on training data by using unpublished problems."