The Bitter Lesson

"these human-knowledge-based chess researchers were not good losers"

In 1997, IBM's Deep Blue beat the world chess champion, Garry Kasparov, mostly by examining hundreds of millions of positions per second — "massive, deep search," in Sutton's phrase. The detail worth savoring is the reaction: the researchers who had spent careers encoding human chess understanding dismissed the win as "brute force," as if that were a criticism. Sutton's point is that their objection was aesthetic, not scientific. The brute-force program won. What "should" have won is not a scientific category — and the essay's whole argument is that this aesthetic preference for human-like methods keeps misleading the field.

"Search and learning are the two most important classes of techniques for utilizing massive amounts of computation"

Go is a board game so vast that chess-style search alone cannot crack it, and for twenty years researchers again tried to build in human knowledge. What finally beat the world's best humans (AlphaGo, in 2016 — Sutton does not name it, writing for readers who know) combined search with learning by self play: the program played millions of games against itself, learning to judge positions with no human examples needed. This paragraph is where the essay names its two heroes. Search and learning are special because they are open-ended: give them ten times the computation and they produce better play, better judgment, better everything — no researcher required.

"researchers always tried to make systems that worked the way the researchers thought their own minds worked"

Two more fields, same plot. In speech recognition, systems built on knowledge of words, phonemes, and the vocal tract lost — starting in the 1970s — to statistical methods that just computed more (the "hidden Markov models" here are a statistical technique; the details do not matter for the argument). In computer vision, decades of hand-designed features for finding edges and shapes were, in Sutton's blunt summary, "discarded" once deep learning arrived. Note the phrase "through Moore's law": computation kept getting exponentially cheaper, so any method that could use more computation had an exponential wind at its back. Methods built from fixed human insight did not.

"building in how we think we think does not work in the long run"

Here is the essay's core, stated as a four-step historical pattern: build in human knowledge; enjoy short-term gains that feel scientifically satisfying; plateau; get overtaken by scaled search and learning. Read this paragraph twice, because it is the one everyone quotes and many overstate. Sutton is not saying knowledge never helps (he concedes it "always helps in the short term"), and he is not saying size alone is destiny. He is making a claim about the long run and about where breakthrough progress comes from. The bitterness is precise: the victories keep coming as defeats for the approach researchers loved, which is why the lesson stays "incompletely digested."

"We want AI agents that can discover like we can, not which contain what we have discovered."

The final paragraphs are the most philosophical and the easiest to misread. Sutton's claim is that the contents of minds — how we actually represent space, objects, other people — are "tremendously, irredeemably complex," so any simple theory of thinking we build into a machine will be wrong. What we should build in instead are meta-methods: machinery that can find complexity on its own. The last sentence is the essay in miniature — and it quietly explains the entire strategy of the modern AI industry, which builds systems that learn from data rather than systems programmed with what we know.