A successful AI system for scientific problems requires three characteristics: substantial real data (supplemented where needed with verified synthetic data to maintain distribution matching), a clear metric that the system can optimize against (such as minimizing free energy in a system), and a massive combinatorial search space too large for brute force methods.