The bitter lesson is the observation in artificial intelligence

Richard Sutton's 2019 essay 'The Bitter Lesson' argues that in artificial intelligence, general approaches that scale with computational power outperform domain-specific methods in the long run, as exemplified by Deep Blue, AlphaGo Zero, and deep learning in speech recognition and computer vision. Sutton concludes that investing in scalable techniques like search and learning is more effective than incorporating human insights.

Bitter lesson | Part of | Artificial intelligence AI https://en.wikipedia.org/wiki/Artificial intelligence Glossary https://en.wikipedia.org/wiki/Glossary of artificial intelligence The bitter lesson is the observation in artificial intelligence https://en.wikipedia.org/wiki/Artificial intelligence that, in the long run, general approaches that scale https://en.wikipedia.org/wiki/Scalability with available computational power tend to outperform ones based on domain-specific understanding https://en.wikipedia.org/wiki/Domain knowledge because they are better at taking advantage of the falling cost of computation over time. The principle was proposed and named in a 2019 essay by Richard Sutton https://en.wikipedia.org/wiki/Richard S. Sutton 1 and is now widely accepted. 2 cite note-generalist-agent-2 3 cite note-switch-transformers-3 4 cite note-learning-bitter-lesson-4 5 cite note-eng-less-artificial-5 6 cite note-brain-lesson-6 7 cite note-beyond-imitation-game-7 8 cite note-overestimation-overfitting-actor-critic-8 The essay edit /w/index.php?title=Bitter lesson&action=edit§ion=1 Sutton gives several examples that illustrate the lesson: Game playing https://en.wikipedia.org/wiki/General game playing . In chess https://en.wikipedia.org/wiki/Chess , the Deep Blue https://en.wikipedia.org/wiki/Deep Blue chess computer system that became the first computer opponent https://en.wikipedia.org/wiki/Computer chess to defeat a world champion https://en.wikipedia.org/wiki/World Chess Championship relied on a relatively simple alpha–beta search algorithm https://en.wikipedia.org/wiki/Alpha–beta pruning that scaled up by applying large amounts of specialized hardware to search for the best move. This defeated previous attempts to exploit the unique structure of chess or to include grandmaster https://en.wikipedia.org/wiki/Grandmaster chess knowledge directly. Likewise in the game of Go https://en.wikipedia.org/wiki/Go game , the AlphaGo https://en.wikipedia.org/wiki/AlphaGo algorithm that surpassed human performance relied much less on expert skill at the game itself than previous generations of AI, and was further surpassed by AlphaGo Zero https://en.wikipedia.org/wiki/AlphaGo Zero , which removed human expertise completely and trained only by self-play https://en.wikipedia.org/wiki/Self-play . Speech recognition https://en.wikipedia.org/wiki/Speech recognition . Approaches based on training a general-purpose hidden Markov model https://en.wikipedia.org/wiki/Hidden Markov model with large numbers of speech samples consistently outperformed the hand-crafted approaches of the 1970s, and deep learning https://en.wikipedia.org/wiki/Deep learning has continued this trend. Computer vision https://en.wikipedia.org/wiki/Computer vision . Algorithms that were assumed to approximate the human visual system https://en.wikipedia.org/wiki/Visual system such as explicitly encoded edge detection https://en.wikipedia.org/wiki/Edge detection or detecting high-level features with SIFT https://en.wikipedia.org/wiki/Scale-invariant feature transform were outperformed by convolutional neural networks https://en.wikipedia.org/wiki/Convolutional neural network that make far fewer assumptions about the nature of visual perception https://en.wikipedia.org/wiki/Visual perception . Sutton concludes that time is better invested in finding simple scalable solutions that can take advantage of Moore's law, rather than introducing ever-more-complex human insights, and calls this the "bitter lesson". He also cites two general-purpose techniques that have been shown to scale effectively: search https://en.wikipedia.org/wiki/Search algorithm and learning https://en.wikipedia.org/wiki/Machine learning . The lesson is considered "bitter" because it is less anthropocentric https://en.wikipedia.org/wiki/Anthropocentrism than many researchers expected and so they have been slow to accept it. Impact edit /w/index.php?title=Bitter lesson&action=edit§ion=2 The essay was published on Sutton's website incompleteideas.net in 2019, and has received hundreds of formal citations according to Google Scholar https://en.wikipedia.org/wiki/Google Scholar . Some of these provide alternative statements of the principle; for example, the 2022 paper "A Generalist Agent" from Google DeepMind https://en.wikipedia.org/wiki/Google DeepMind summarized the lesson as: 2 cite note-generalist-agent-2 Historically, generic models that are better at leveraging computation have also tended to overtake more specialized domain-specific approaches, eventually. Another phrasing of the principle is seen in a Google paper on switch transformers https://en.wikipedia.org/wiki/Transformer deep learning architecture coauthored by Noam Shazeer https://en.wikipedia.org/wiki/Noam Shazeer : 3 cite note-switch-transformers-3 Simple architectures—backed by a generous computational budget, data set size and parameter count—surpass more complicated algorithms. The principle is further referenced in many other works on artificial intelligence. For example, From Deep Learning to Rational Machines draws a connection to long-standing debates in the field, such as Moravec's paradox https://en.wikipedia.org/wiki/Moravec's paradox and the contrast between neats and scruffies https://en.wikipedia.org/wiki/Neats and scruffies . 9 In "Engineering a Less Artificial Intelligence", the authors concur that "flexible methods so far have always outperformed handcrafted domain knowledge in the long run" although note that " w ithout the right implicit assumptions, generalization https://en.wikipedia.org/wiki/Generalization learning is impossible". More recently, "The Brain's Bitter Lesson: Scaling Speech Decoding With Self-Supervised Learning" continues Sutton's argument, contending that as of 2025 the lesson has not been fully learned in the fields of speech recognition and 5 cite note-eng-less-artificial-5 brain data https://en.wikipedia.org/wiki/Brain–computer interface . 6 cite note-brain-lesson-6 Other work has looked to apply the principle and validate it in new domains. For example, the 2022 paper "Beyond the Imitation Game" applies the principle to large language models https://en.wikipedia.org/wiki/Large language model to conclude that "it is vitally important that we understand their capabilities and limitations" to "avoid devoting research resources to problems that are likely to be solved by scale alone". 7 In 2024, "Learning the Bitter Lesson: Empirical Evidence from 20 Years of CVPR Proceedings" looked at further evidence from the field of computer vision and pattern recognition https://en.wikipedia.org/wiki/Pattern recognition , and concludes that the previous twenty years of experience in the field shows "a strong adherence to the core principles of the 'bitter lesson'". In "Overestimation, Overfitting, and Plasticity in Actor-Critic: the Bitter Lesson of Reinforcement Learning", the authors look at generalization of 4 cite note-learning-bitter-lesson-4 actor-critic algorithms https://en.wikipedia.org/wiki/Actor-critic algorithm and find that "general methods that are motivated by stabilization of gradient-based learning https://en.wikipedia.org/wiki/Gradient descent significantly outperform RL https://en.wikipedia.org/wiki/Reinforcement learning -specific algorithmic improvements across a variety of environments" and note that this is consistent with the bitter lesson. 8 cite note-overestimation-overfitting-actor-critic-8 References edit /w/index.php?title=Bitter lesson&action=edit§ion=3 ↑ cite ref-1 Sutton, Rich March 13, 2019 . "The Bitter Lesson" http://www.incompleteideas.net/IncIdeas/BitterLesson.html . www.incompleteideas.net . Retrieved September 7, 2025. 1 cite ref-generalist-agent 2-0 2 cite ref-generalist-agent 2-1 Reed, Scott; Zolna, Konrad; Parisotto, Emilio; et al. 2022 . "A Generalist Agent" https://openreview.net/forum?id=1ikK0kHjvj . Transactions on Machine Learning Research 2834–8856 . arXiv https://en.wikipedia.org/wiki/ArXiv identifier : 2205.06175 https://arxiv.org/abs/2205.06175 . Retrieved September 7, 2025. 1 cite ref-switch-transformers 3-0 2 cite ref-switch-transformers 3-1 Fedus, William; Zoph, Barret; Shazeer, Noam 2022 . "Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity" http://jmlr.org/papers/v23/21-0998.html . Journal of Machine Learning Research . 23 120 : 1–39. Retrieved September 14, 2025. 1 cite ref-learning-bitter-lesson 4-0 2 cite ref-learning-bitter-lesson 4-1 Yousefi, Mojtaba; Collins, Jack. "Learning the Bitter Lesson: Empirical Evidence from 20 Years of CVPR Proceedings" https://aclanthology.org/2024.nlp4science-1.15/ . Proceedings of the 1st Workshop on NLP for Science NLP4Science . Association for Computational Linguistics. pp. 175–187. Retrieved September 7, 2025. 1 cite ref-eng-less-artificial 5-0 2 cite ref-eng-less-artificial 5-1 Sinz, Fabian H.; Pitkow, Xaq; Reimer, Jacob; et al. 2019 . "Engineering a Less Artificial Intelligence" https://www.cell.com/neuron/fulltext/S0896-6273 19 30740-8 . Neuron . 103 6 . Elsevier: 967–979. doi https://en.wikipedia.org/wiki/Doi identifier : 10.1016/j.neuron.2019.08.034 https://doi.org/10.1016%2Fj.neuron.2019.08.034 . Retrieved September 13, 2025. 1 cite ref-brain-lesson 6-0 2 cite ref-brain-lesson 6-1 Jayalath, Dulhan; Landau, Gilad; Shillingford, Brendan; Woolrich, Mark; Parker Jones, ʻŌiwi 2025 . "The Brain's Bitter Lesson: Scaling Speech Decoding With Self-Supervised Learning" https://icml.cc/virtual/2025/poster/44019 . Forty-second International Conference on Machine Learning . Proceedings of Machine Learning Research. Retrieved September 13, 2025. 1 cite ref-beyond-imitation-game 7-0 2 cite ref-beyond-imitation-game 7-1 Srivastava, Aarohi; Rastogi, Abhinav; Rao, Abhishek; Awal, Abu; Abid, Abubakar; et al. "Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models" https://arxiv.org/abs/2206.04615 . The Fourteenth International Conference on Learning Representations . 1 cite ref-overestimation-overfitting-actor-critic 8-0 2 cite ref-overestimation-overfitting-actor-critic 8-1 Nauman, Michal; Bortkiewicz, Michał; Miłoś, Piotr; Trzciński, Tomasz; Ostaszewski, Mateusz; et al. 2024 . "Overestimation, Overfitting, and Plasticity in Actor-Critic: the Bitter Lesson of Reinforcement Learning" https://arxiv.org/pdf/2403.00514 . Proceedings of the 41st International Conference on Machine Learning . Proceedings of Machine Learning Research. Retrieved September 13, 2025. ↑ cite ref-9 Buckner, Cameron J. December 11, 2023 . From Deep Learning to Rational Machines: What the History of Philosophy Can Teach Us about the Future of Artificial Intelligence . Oxford University Press. doi https://en.wikipedia.org/wiki/Doi identifier : 10.1093/oso/9780197653302.001.0001 https://doi.org/10.1093%2Foso%2F9780197653302.001.0001 . ISBN https://en.wikipedia.org/wiki/ISBN identifier 9780197653302 https://en.wikipedia.org/wiki/Special:BookSources/9780197653302 .