“High scores on reasoning benchmarks like the "Humanity's last exam" do not necessarily mean an AI model will be capable of performing real-world science.”