{"slug": "the-bitter-lesson", "title": "The Bitter Lesson", "summary": "In a 2019 essay, computer scientist Rich Sutton argues that 70 years of AI research show general methods that leverage computation are ultimately the most effective, citing examples from computer chess, Go, and speech recognition. He contends that researchers' reliance on human knowledge often complicates methods and hinders progress, while search and learning enable massive computation to be brought to bear.", "body_md": "# The Bitter Lesson\n\n## Rich Sutton\n\n### March 13, 2019\n\nThe biggest lesson that can be read from 70 years of AI research is\nthat general methods that leverage computation are ultimately the most\neffective, and by a large margin. The ultimate reason for this is\nMoore's law, or rather its generalization of continued exponentially\nfalling cost per unit of computation. Most AI research has been\nconducted as if the computation available to the agent were constant\n(in which case leveraging human knowledge would be one of the only ways\nto improve performance) but, over a slightly longer time than a typical\nresearch project, massively more computation inevitably becomes\navailable. Seeking an improvement that makes a difference in the\nshorter term, researchers seek to leverage their human knowledge of the\ndomain, but the only thing that matters in the long run is the\nleveraging of computation. These two need not run counter to each\nother, but in practice they tend to. Time spent on one is time not\nspent on the other. There are psychological commitments to investment\nin one approach or the other. And the human-knowledge approach tends to\ncomplicate methods in ways that make them less suited to taking\nadvantage of general methods leveraging computation. There were\nmany examples of AI researchers' belated learning of this bitter\nlesson,\nand it is instructive to review some of the most prominent.\n\nIn computer chess, the methods that defeated the world champion,\nKasparov, in 1997, were based on massive, deep search. At the time,\nthis was looked upon with dismay by the majority of computer-chess\nresearchers who had pursued methods that leveraged human understanding\nof the special structure of chess. When a simpler, search-based\napproach with special hardware and software proved vastly more\neffective, these human-knowledge-based chess researchers were not good\nlosers. They said that ``brute force\" search may have won this time,\nbut it was not a general strategy, and anyway it was not how people\nplayed chess. These researchers wanted methods based on human input to\nwin and were disappointed when they did not.\n\nA similar pattern of research progress was seen in computer Go, only\ndelayed by a further 20 years. Enormous initial efforts went into\navoiding search by taking advantage of human knowledge, or of the\nspecial features of the game, but all those efforts proved irrelevant,\nor worse, once search was applied effectively at scale. Also important\nwas the use of learning by self play to learn a value function (as it\nwas in many other games and even in chess, although learning did not\nplay a big role in the 1997 program that first beat a world champion).\nLearning by self play, and learning in general, is like search in that\nit enables massive computation to be brought to bear. Search and\nlearning are the two most important classes of techniques for utilizing\nmassive amounts of computation in AI research. In computer Go, as in\ncomputer chess, researchers' initial effort was directed towards\nutilizing human understanding (so that less search was needed) and only\nmuch later was much greater success had by embracing search and\nlearning.\n\nIn speech recognition, there was an early competition, sponsored by\nDARPA, in the 1970s. Entrants included a host of special methods that\ntook\nadvantage of human knowledge---knowledge of words, of phonemes, of the\nhuman vocal tract, etc. On the other side were newer methods that were\nmore statistical in nature and did much more computation, based on\nhidden Markov models (HMMs). Again, the statistical methods won out\nover the human-knowledge-based methods. This led to a major change in\nall of natural language processing, gradually over decades, where\nstatistics and computation came to dominate the field. The recent rise\nof deep learning in speech recognition is the most recent step in this\nconsistent direction. Deep learning methods rely even less on human\nknowledge, and use even more computation, together with learning on\nhuge training sets, to produce dramatically better speech recognition\nsystems. As in the games, researchers always tried to make systems that\nworked the way the researchers thought their own minds worked---they\ntried to put that knowledge in their systems---but it proved ultimately\ncounterproductive, and a colossal waste of researcher's time, when,\nthrough Moore's law, massive computation became available and a means\nwas found to put it to good use.\n\nIn computer vision, there has been a similar pattern. Early methods\nconceived of vision as searching for edges, or generalized cylinders,\nor in terms of SIFT features. But today all this is discarded. Modern\ndeep-learning neural networks use only the notions of convolution and\ncertain kinds of invariances, and perform much better.\n\nThis is a big lesson. As a field, we still have not thoroughly learned\nit, as we are continuing to make the same kind of mistakes. To see\nthis, and to effectively resist it, we have to understand the appeal of\nthese mistakes. We have to learn the bitter lesson that building in how\nwe think we think does not work in the long run. The bitter lesson is\nbased on the historical observations that 1) AI researchers have often\ntried to build knowledge into their agents, 2) this always helps in the\nshort term, and is personally satisfying to the researcher, but 3) in\nthe long run it plateaus and even inhibits further progress, and 4)\nbreakthrough progress eventually arrives by an opposing approach based\non scaling computation by search and learning. The eventual success is\ntinged with bitterness, and often incompletely digested, because it is\nsuccess over a favored, human-centric approach.\n\nOne thing that should be learned from the bitter lesson is the great\npower of general purpose methods, of methods that continue to scale\nwith increased computation even as the available computation becomes\nvery great. The two methods that seem to scale arbitrarily in this way\nare search and learning.\n\nThe second general point to be learned from the bitter lesson is that\nthe actual contents of minds are tremendously, irredeemably complex; we\nshould stop trying to find simple ways to think about the contents of\nminds, such as simple ways to think about space, objects, multiple\nagents, or symmetries. All these are part of the arbitrary,\nintrinsically-complex, outside world. They are not what should be built\nin, as their complexity is endless; instead we should build in only the\nmeta-methods that can find and capture this arbitrary complexity.\nEssential to these methods is that they can find good approximations,\nbut the search for them should be by our methods, not by us. We want AI\nagents that can discover like we can, not which contain what we have\ndiscovered. Building in our discoveries only makes it harder to see how\nthe discovering process can be done.", "url": "https://wpnews.pro/news/the-bitter-lesson", "canonical_source": "http://www.incompleteideas.net/IncIdeas/BitterLesson.html", "published_at": "2026-08-27 06:59:42+00:00", "updated_at": "2026-08-27 07:19:00.000947+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "machine-learning"], "entities": ["Rich Sutton", "DARPA", "Kasparov"], "alternates": {"html": "https://wpnews.pro/news/the-bitter-lesson", "markdown": "https://wpnews.pro/news/the-bitter-lesson.md", "text": "https://wpnews.pro/news/the-bitter-lesson.txt", "jsonld": "https://wpnews.pro/news/the-bitter-lesson.jsonld"}}