The bitter lesson

Richard Sutton, who had formalised temporal-difference learning in 1988, published a thousand-word essay on his website on 13 March 2019 that became the most quoted document of the scaling era. Its argument is historical. In chess, the researchers who encoded human knowledge lost to Deep Blue's search. In Go, they lost to self-play. In speech and vision, hand-built features lost to learning from data. Each time, the field's instinct was to build in what it knew, and each time a method that instead used more computation, general search and general learning, won, because computation gets cheaper by Moore's law and human knowledge does not.

The lesson is bitter, he wrote, because it is a loss for the researchers' own contributions: "the only thing that matters in the long run is the leveraging of computation." What should be built in is not knowledge but the capacity to acquire it.

The essay was published a year before the scaling laws quantified it and three years before ChatGPT made it obvious. It is cited by those who build ever-larger models as a justification and by their critics as the confession of a field that has stopped thinking. Either way, the decade since has not produced a counterexample.