Richard Sutton says that in the long run, general-purpose, compute-driven learning methods beat hand-crafted, human-designed rules. This pushes us to think of modern AI success as a systems engineering achievement rather than a clever human knowledge design achievement. Why is this a "bitter lesson"? because it hurts our ego. We love to inject expert knowledge, handcrafted features, tricks, etc. but historically (e.g. chess, Go, speech recognition, NLP, and more) whenever the field matured, the handcrafted approaches got crushed by large-scale compute-heavy general algorithms that learned directly from data.

reliability; scalability; robustness.