"Convert out-of-distribution gaps into training coverage Hashimoto offers a unifying principle for why AI models fail on edge cases: 'Whatever is out of distribution, just make in distribution — make sure that you make all the out of distribution in distribution. This is how generative AI works. This is how even self-driving cars work.' He contrasts this with human learning, where we handle novel situations without needing prior examples — calling it 'the real mystery of intelligence.' The actionable takeaway for anyone building a model, process, or training program: systematically identify the scenarios your system has never seen, and deliberately create exposure to those exact scenarios before the real test arrives. Source: Tatsunori Hashimoto, The Cognitive Revolution (or similar Stanford AI podcast), Small Language Models and AI Democratization"