Rich Sutton put a one-page essay on his personal website in March 2019. It has since been cited hundreds of times, summarized in DeepMind papers, quoted in Google research, and given its own Wikipedia entry. The essay itself is about 1,000 words and never left that page.
Its thesis: across 70 years of AI research, systems built from handcrafted human knowledge keep losing to simpler systems built to consume as much computation as possible. That idea, known as the Bitter Lesson, is the closest thing the scaling era has to a founding document. It also explains why Sutton, its author, thinks today's large language models have not finished applying it.
The claim
Sutton's opening sentence is the whole argument: "The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin."
Two premises support it. First, computation keeps getting cheaper. Sutton frames this as Moore's law generalized: any method evaluated at a fixed compute budget will look, in hindsight, like it was designed for the wrong era, because massively more computation always arrives. Second, only a couple of methods can absorb that arriving computation without limit. In his words: "The two methods that seem to scale arbitrarily in this way are search and learning."
The lesson is called bitter for a specific reason. Building in human expertise works in the short term, satisfies the researchers doing it, and then plateaus. The simpler scalable approach eventually wins anyway, and it wins over methods the field liked and had invested in. Progress arrives as a loss.

Four case studies
Chess. The program that beat Garry Kasparov in 1997, IBM's Deep Blue, won with massive, deep search at a time when most chess researchers pursued methods built on human understanding of the game's structure. Sutton notes that the losing camp responded that brute force "may have won this time, but it was not a general strategy."
Go. The same pattern, delayed about 20 years. Early Go work tried to avoid search using human knowledge of the game. Once search and self-play learning were applied at scale, those efforts proved irrelevant or worse. The 2016 AlphaGo paper in Nature already leaned on self-play, and its successors dropped human data entirely: AlphaZero defeated world-champion-level programs in chess, shogi and Go "starting from random play and given no domain knowledge except the game rules."
Speech recognition. In the 1970s, a DARPA-sponsored competition pitted methods encoding human knowledge of words, phonemes and the vocal tract against statistical methods that did far more computation, based on hidden Markov models. The statistical side won, and the shift reshaped all of natural language processing over the following decades. Deep learning's arrival in speech in the 2010s was, in Sutton's telling, just the latest step in the same direction.
Computer vision. Early vision research modeled the world the way people thought they saw it: edges, generalized cylinders, SIFT features. Modern deep networks discarded all of it in favor of convolution, a few invariances, and very large training runs.
Translation supplies a post-essay example Sutton did not need. In 2016, Google replaced its decade-old phrase-based machine translation pipeline with a single neural model, and said the switch improved quality in one leap more than the previous ten years combined.
From blog post to doctrine
The essay's formal citations cluster in the 2020s, which read as its confirmation. GPT-3 (2020) took the no-task-specific-engineering stance to its limit: a 175-billion-parameter model applied to dozens of tasks with no gradient updates, where scaling itself produced few-shot ability. DeepMind's Gato paper (2022) restated the lesson almost verbatim: "Historically, generic models that are better at leveraging computation have also tended to overtake more specialized domain-specific approaches, eventually." Google's Switch Transformers paper (2022) argued the design implication: "Simple architectures - backed by a generous computational budget, data set size and parameter count - surpass more complicated algorithms." A 2024 study of 20 years of CVPR proceedings found computer vision's publication record shows "a strong adherence to the core principles" of the lesson.
Sutton's second point is less quoted than the first, and probably the deeper one. The contents of minds, he argues, are irredeemably complex, so researchers should stop encoding their theories of space, objects or symmetry into systems and instead build only the meta-methods that discover those structures themselves: "We want AI agents that can discover like we can, not which contain what we have discovered."
The author's objection to the scaling era
Sutton shared the 2024 Turing Award with Andrew Barto for the development of reinforcement learning, and he does not think the current LLM wave has closed the loop he described. In a September 2025 interview on the Dwarkesh Patel podcast, he argued that language models mimic what people said rather than learn what actually works: "The large language models are learning from training data. It's not learning from experience." Because their weights stop changing at release, they never collect the feedback stream that, in his framing, is what intelligence is for.
In April 2025, with DeepMind's David Silver, he published Welcome to the Era of Experience, a chapter for the MIT Press volume Designing an Intelligence. It predicts that progress from supervised learning on human data slows, and that the next gains come from agents learning "predominantly from experience," in continuous streams, grounded in real environments, with rewards measured from consequences instead of human ratings. The paper's proof case is AlphaProof, which trained on self-generated proofs in a simulated interaction environment to reach International Mathematical Olympiad medal-level results. That is the Bitter Lesson pointed at the next decade: static weights are the new handcrafted knowledge, and the field will eventually route around them the same way it routed around handcrafted chess strategy.
One qualification, from researchers who accept the general direction: a 2019 Neuron paper30740-8) agrees that "flexible methods so far have always outperformed handcrafted domain knowledge in the long run," while noting that without the right implicit assumptions, generalization is impossible. The lesson is a claim about the long race, not a guarantee that structure never helps at any given budget.

What to take from it
- Treat handcrafted domain engineering as rented, not owned. It buys short-term wins that cheapening computation erodes.
- When two approaches compete, favor the one whose bottleneck is compute rather than researcher time. Search and learning are the two that have never stopped scaling.
- Read capability announcements through this lens: jumps that come from scale are predictable; claims that a clever new structure will permanently beat scaling are the pattern that keeps losing.
- Take the 2025 twist seriously. The same argument that beat handcrafted knowledge now points at frozen weights, which makes continual learning from experience the open problem.
Sources
- Rich Sutton, The Bitter Lesson, incompleteideas.net, March 13, 2019
- Wikipedia, Bitter lesson (impact and citations)
- Silver et al., Mastering the game of Go with deep neural networks and tree search, Nature, 2016
- Silver et al., A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play, Science, 2018
- Quoc Le and Mike Schuster, A Neural Network for Machine Translation, at Production Scale, Google Research blog, September 27, 2016
- Found in translation: More accurate, fluent sentences in Google Translate, Google blog, November 2016
- Brown et al., Language Models are Few-Shot Learners, 2020
- Reed et al., A Generalist Agent, 2022
- Fedus, Zoph and Shazeer, Switch Transformers, JMLR, 2022
- Yousefi and Collins, Learning the Bitter Lesson: Empirical Evidence from 20 Years of CVPR Proceedings, 2024
- Sinz et al., Engineering a Less Artificial Intelligence30740-8), Neuron, 2019
- David Silver and Richard Sutton, Welcome to the Era of Experience, 2025
- Dwarkesh Patel, Richard Sutton: Father of RL thinks LLMs are a dead end, September 2025
- ACM, 2024 Turing Award
