The Bitter Lesson (2019) [pdf] (www.cs.utexas.edu)

In the influential paper "The Bitter Lesson," AI researcher Rich Sutton emphasizes a pivotal realization in AI and machine learning: leveraging large-scale computation has consistently outperformed approaches that incorporate human knowledge. Sutton argues that many researchers have historically prioritized methods based on human expertise, such as domain-specific heuristics, leading to temporary successes but ultimately stalling progress. Key examples illustrate this trend, including the triumph of computational strategies in chess and Go over human-centric approaches, and the shift from knowledge-based methods to data-driven deep learning in speech recognition and computer vision. Sutton's insights are vital for the AI/ML community, as they highlight the need to embrace general-purpose techniques like search and learning that scale with available computation. He cautions that reliance on human knowledge may inhibit innovation and that AI should focus on developing methods that can independently discover complex patterns instead of embedding prior human understanding. This perspective serves as a reminder to the field that future breakthroughs are likely to come from scalable computational approaches rather than attempts to replicate our cognitive processes.