Richard Sutton’s “bitter lesson” is that, across several histories of AI, general methods that can exploit more computation have tended to outlast methods built around researchers’ hand-coded understanding of a particular domain. In Sutton’s account, the methods with the clearest capacity to scale are search and learning—not compute by itself, and not a rule that specialist knowledge is always useless.
What is the bitter lesson?
Sutton’s essay, “The Bitter Lesson,” published March 13, 2019, looks back over what he frames as 70 years of AI research. He writes: “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.” The 70-year figure is Sutton’s retrospective framing, not a separately measured statistic.
As an Amazon Associate I earn from qualifying purchases.
The “26 Words” in some descriptions of the idea is an editorial framing; Sutton’s essay does not call its thesis a 26-word statement. The sentence above is an exact quotation. Any shorter summary should be treated as a paraphrase, not as Sutton’s own slogan.
Why computation matters in Sutton’s argument
Sutton’s contrast is between what a system relies on to improve. One approach encodes substantial human understanding of a domain—for example, knowledge about how people play a game or how speech works. The other relies more on general-purpose procedures that can use additional computation to search possibilities or learn from data and experience.
#1 Best Overall
| Question | Specialized approach | General, computation-intensive approach |
|---|---|---|
| What is built into the system? | Human understanding of the particular domain | General procedures such as search and learning |
| What drives improvement? | More or better domain-specific insight | More computation applied through the method |
| What is the scaling question? | Whether added specialization continues to help | Whether the method can keep benefiting as computation grows |
These are explanatory contrasts drawn from Sutton’s essay, not a universal scorecard for every AI system. He argues that specialist knowledge can help in the short term yet plateau or inhibit later progress when broader computation-intensive methods become practical. He points to falling computation costs as the backdrop to the pattern; he does not claim that compute always increases or that it alone explains every advance.
Sutton summarizes the methods he sees as especially scalable this way: “The two methods that seem to scale arbitrarily in this way are search and learning.” The qualification “seem” matters: this is his conclusion from the examples he reviews, not a guarantee about every task or future system.
Four examples Sutton uses
Chess
Sutton says the methods that defeated world champion Garry Kasparov in 1997 relied on massive, deep search. He contrasts that approach with systems that emphasized human understanding of chess. The example supports his argument that a general procedure capable of using substantial computation can surpass an approach built around expert knowledge of the domain.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Go
In Sutton’s account, a similar shift came later in Go. Search and learning from self-play played central roles. The comparison reinforces his emphasis on methods that improve through computation rather than relying chiefly on researchers’ explicit encoding of how the game should be played.
Speech recognition
Sutton contrasts early systems based on human knowledge of linguistics and articulation with statistical approaches, then describes deep learning as a later step that used more computation and large training sets. His point is about the direction of the methods he reviews, not that linguistic knowledge or earlier techniques vanished from all speech-recognition practice.
Computer vision
For vision, Sutton describes earlier approaches built around edges, generalized cylinders, and SIFT features, and contrasts them with deep-learning networks using convolution and certain invariances. The example illustrates his argument about general learning methods benefiting from computation; it does not establish that every older feature or technique disappeared from use.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the lesson does—and does not—claim
The essay is a historical argument based on selected areas, not proof that all domain knowledge, engineering, or specialist methods are futile. Sutton’s focus is on what to build into an AI agent and which methods can continue to benefit as computation grows. A useful reading is therefore: ask whether a proposed source of improvement can scale through search or learning as more computation becomes available, rather than assuming that adding more hand-coded expertise will keep producing gains.
Recommended Free Tools
That distinction leaves room for domain knowledge to matter in a system’s design, data, constraints, or practical deployment. The essay’s examples support Sutton’s claim about a recurring pattern in AI research; they do not settle every engineering trade-off or show that compute alone is sufficient.
Further reading on reinforcement learning
Readers who want a textbook introduction to reinforcement-learning ideas and algorithms can consult Reinforcement Learning: An Introduction, second edition, by Richard S. Sutton and Andrew G. Barto, published by MIT Press on November 13, 2018 (hardcover ISBN 9780262039246). It is a book about reinforcement learning, not a commentary on “The Bitter Lesson,” and it is not required to understand the essay.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
