CONCEPT
The Bitter Lesson
Sutton’s 2019 thesis that seventy years of AI history teach a single humbling truth: general methods that leverage computation beat human cleverness in the long run, by a large margin, every time.
In March 2019,
Richard Sutton posted a short essay that became one of the most discussed pieces of writing in artificial intelligence. He called it “The Bitter Lesson,” and its argument was as blunt as its title: the biggest lesson to be drawn from seventy years of AI research is that general methods which leverage computation are ultimately the most effective, and by a large margin. The lesson is bitter because it humbles. It says that researcher cleverness about the structure of a problem—the knowledge encoded, the symmetry exploited, the heuristic installed—matters less than the willingness to build systems that search and learn at scale. Chess, Go, speech recognition, computer vision: in each domain, the systems that finally succeeded did so through massive search and data-driven learning, displacing the carefully engineered approaches that had dominated the field and that researchers had been proud of. The positive content of the essay—that the methods worth building are those that scale, specifically search and learning—is as