"The Bitter Lesson" is a short essay that Richard Sutton, a pioneer of reinforcement learning, posted on his website on 13 March 2019. Its thesis: the biggest lesson of seventy years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin (Sutton, 2019). The story repeats itself in chess, Go, speech recognition and vision: researchers build in their domain knowledge, which helps in the short term and is personally satisfying, but in the long run it plateaus, and the breakthrough comes from the opposite direction, with search and learning scaled up. Hence the bitterness: success arrives against the human-centric approach one preferred.
The essay's most philosophical conclusion is that the contents of minds are "tremendously,
irredeemably complex" and should not be hand-coded; we should build only the meta-methods that
discover them: "We want AI agents that can discover like we can, not which contain what we have
discovered" (Sutton, 2019). In San Francisco the essay became scripture for the
scaling hypothesis and an epitaph for GOFAI. The song celebrates a model that works
without a single line of Lisp: "Without a single cdr" (osmarks, n.d.).
Buckner treats it as the pole of "radical empiricism" in the empiricist–nativist debate, and offers a better reading: good structural tweaks are those that turn more computation into better performance (Buckner, 2024, pp. 25–26) (see empiricism and nativism). Rodney Brooks replied in 2019 that learning's successes depend on built-in human ingenuity, such as the translation invariance of convolutional networks (inductive bias). And Sutton himself argued in a 2025 interview that language models are not fully faithful to his lesson, because they rely on the human knowledge contained in texts instead of learning from experience.