Alexander Mattick discusses the evolution and challenges of probabilistic inference methods in machine learning, emphasizing the limitations of current models, the importance of incorporating constraints for safety and reliability, and the need for more rigorous, testable theories to better understand deep learning. He critiques optimistic assumptions like “reward is enough,” highlights the practical difficulties in modeling complex environments, and underscores that more data alone cannot replace the necessity of structured priors and constraints for effective, robust AI systems.