The Bitter Lesson

Filed correctly. proposals.md lists this under interpretation and the filing holds, but two rival kinds are close enough that declining them is the first thing this file owes a reader — and one of them has to be declined and then honoured anyway.

idea is the near miss. On the test samuel-checkers-1959 set and goodharts-law-1975 applied, an idea is something you can implement, because there is a loop or a formalism in it. There is none here. The essay recommends a class of methods — search and learning — but the loop belongs to the methods, not to the essay, and the methods predate it by decades; Sutton's own contribution to them is temporal-difference learning in 1988, a different artifact entirely. What this document supplies is a reading of seventy years of results plus a research-strategy prescription drawn from it. That is an interpretation.

prediction is the serious rival and it is closer than it looks, because the essay does make a forward claim — that the human-knowledge approach will keep losing, and that researchers will keep making the same mistake. I decline the kind and then apply its discipline anyway, in section 3, for the reason moravecs-paradox-1988 gave: the forward claim carries no date and no due date, so it cannot be graded as a prediction without inventing the deadline the author declined to set. What it can be graded on is direction, and seven years have produced enough record to do that. The grading matters more here than for most entries, because the author's own grade and the field's grade point in opposite directions, which is the fact this entry mainly exists to carry, and section 3 is where it lives.

On descends_from. Only one of the four is a named citation. The other three are identifications I made, and the strength of each is stated rather than implied.

• deep-blue-1997 is the one the essay names. Its second paragraph opens on "the methods that defeated the world champion, Kasparov, in 1997," and the whole essay's rhetorical shape — the winners were resented, the losers were "not good losers" — is built out of that match. This link is documented.
• alphago-move-37-2016 is an identification. The essay's third paragraph is about computer Go, "only delayed by a further 20 years," and names no system, no lab and no date. 2016 plus the reference to "the use of learning by self play to learn a value function" identifies AlphaGo beyond reasonable doubt, and alphago-move-37-2016 is the id under which this canon holds that event. The essay does not say "AlphaGo". There is a wrinkle worth keeping: the system that won in March 2016 was not free of human knowledge — its policy network was pretrained on human expert games. The version that matched Sutton's description arrived in October and December 2017, as AlphaGo Zero and AlphaZero. The essay compresses eighteen months and two systems into one sentence.
• alexnet-2012 is an identification of the same kind, and weaker. The vision paragraph names SIFT, edges and generalized cylinders as the discarded past and "modern deep-learning neural networks" as the present, with no event and no year. 2012 is where this canon files that turn. alexnet-2012's own file already anticipates the link and warns against reading its evidence as support for this essay's stronger claim; that warning is correct and section 4 restates it.
• samuel-checkers-1959 is an authorship link, not a citation. The essay never mentions Samuel. It is listed because Sutton and Barto's textbook treats Samuel's checkers program as the origin of temporal-difference learning — the method that made Sutton's career and won him a Turing Award — and because samuel-checkers-1959's own file names this entry as the one that "would state this as a thesis," pointing at the three-way pattern it shares with shannon-chess-1950 and logic-theorist-1956: hand-written component is the bottleneck, author says so, approach reinstated decades later once the hand-written part is learned. That is the bitter lesson in three documented instances, written down before the essay existed. Declining the link would hide the strongest evidence in the canon for the entry's own thesis.

Ancestors considered and not listed. shannon-chess-1950 is deliberately excluded even though it is about the same 1997 match, because its file supplies the counterweight rather than the descent, and section 3 uses it that way. moravecs-paradox-1988 is a distinct claim that arrives in the same conversation and is routinely confused with this one; its file says so and I agree. perceptrons-1969 and lighthill-1973 are the field's two great cases of a hand-analysis stopping a scalable method, which makes them thematically close and textually absent.

Two real ancestors have no id and cannot be listed. The statistical-over-linguistic turn in speech recognition, which the essay's fourth paragraph rests on, is carried in the field's memory by Frederick Jelinek's quip about firing linguists — a line whose wording and date are both disputed (Jelinek himself recalled it as "Anytime a linguist leaves the group the recognition rate goes up," December 1988; another account gives a different wording and 1985). And the essay's argument had already been made for data rather than compute by Alon Halevy, Peter Norvig and Fernando Pereira in "The Unreasonable Effectiveness of Data" (IEEE Intelligent Systems, 2009). Neither is in proposals.md; both would be reasonable additions, and the 2009 paper is the better candidate because it is a document rather than an anecdote.

A descendant, for the builder's sake. scaling-laws-2020 is downstream of this essay in doctrine and cannot be linked from here, because descent points backwards and a canon job writes one file. A session that ever revisits that entry should consider adding this id to its ancestors.

Housekeeping this file cannot do. Five entries — samuel-checkers-1959, alexnet-2012, moravecs-paradox-1988, rosenblatt-perceptron-1958, geb-1979 — and amodei-2024-loving-grace's ancestor note all describe this id as proposed and unwritten, and geb-1979 explicitly says its paragraph "should defer to it" once it exists. It exists now. Those sentences are stale as of this file's date and none of them is mine to correct.

---

What it is

The document

"The Bitter Lesson" is a blog post of roughly 1,100 words, dated 13 March 2019, published by Richard S. Sutton at incompleteideas.net/IncIdeas/BitterLesson.html — a personal page whose parent site is named, without irony, Incomplete Ideas. It has no abstract, no citations, no references, no data and no figures. It prints to two pages. It was not submitted anywhere, and it has never been revised.

It is, by a wide margin, the most influential document in modern AI that was never peer reviewed, and the single most quoted justification for the way the field has spent its money since. As of this entry it is served over a certificate the fetch layer rejects as self-signed; the canonical statement of the compute era sits behind a broken TLS handshake on a professor's homepage.

Its author is not a pundit. Sutton and Andrew Barto received the 2024 ACM A.M. Turing Award, announced 5 March 2025, "for developing the conceptual and algorithmic foundations of reinforcement learning." Their textbook has been cited over 75,000 times. That credential cuts both ways and section 4 says how: the essay's prescription — search and learning are the two methods that scale — is a recommendation of the author's own field, made by the person who built much of it, and it should be read with the same scepticism this project applies to any principal describing the value of their own work.

The thesis, in its own words

> "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 mechanism, which is the part people forget

The essay does not claim compute is magic. It claims an asymmetry between two clocks, and everything else follows from it:

> "The ultimate reason for this is Moore's law, or rather its generalization of continued > exponentially falling cost per unit of computation. Most AI research has been conducted > as if the computation available to the agent were constant (in which case leveraging > human knowledge would be one of the only ways to improve performance) but, over a > slightly longer time than a typical research project, massively more computation > inevitably becomes available."

The unit of analysis is the length of a research project against the doubling time of hardware. A researcher optimising for the next paper is correct to add domain knowledge; a field optimising over a decade is not. Then the psychological claim, which is the essay's sharpest observation and the one least often quoted:

> "These two need not run counter to each other, but in practice they tend to. Time spent > on one is time not spent on the other. There are psychological commitments to investment > in one approach or the other. And the human-knowledge approach tends to complicate > methods in ways that make them less suited to taking advantage of general methods > leveraging computation."

The four cases

Chess. The 1997 methods were "based on massive, deep search." This "was looked upon with dismay by the majority of computer-chess researchers who had pursued methods that leveraged human understanding of the special structure of chess." They "were not good losers": brute force "may have won this time, but it was not a general strategy, and anyway it was not how people played chess."

Go. The same pattern, "only delayed by a further 20 years." Initial effort went into "avoiding search by taking advantage of human knowledge, or of the special features of the game, but all those efforts proved irrelevant, or worse, once search was applied effectively at scale." Then the sentence that unifies the essay's two nominated methods: "Learning by self play, and learning in general, is like search in that it enables massive computation to be brought to bear."

Speech recognition. The 1970s DARPA competition set "special methods that took advantage of human knowledge — knowledge of words, of phonemes, of the human vocal tract" against "newer methods that were more statistical in nature and did much more computation, based on hidden Markov models." The statistical methods won, and the field moved. Researchers "always tried to make systems that worked the way the researchers thought their own minds worked… but it proved ultimately counterproductive, and a colossal waste of researcher's time."

Vision. "Early methods conceived of vision as searching for edges, or generalized cylinders, or in terms of SIFT features. But today all this is discarded. Modern deep-learning neural networks use only the notions of convolution and certain kinds of invariances, and perform much better."

That last clause is the seam in the essay, and section 3 pulls on it.

The four historical observations, numbered by the author

> "1) AI researchers have often tried to build knowledge into their agents, 2) this > always helps in the short term, and is personally satisfying to the researcher, but > 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough > progress eventually arrives by an opposing approach based on scaling computation by > search and learning. The eventual success is tinged with bitterness, and often > incompletely digested, because it is success over a favored, human-centric approach."

The bitterness is the title. The lesson is bitter to the researcher whose approach lost, not to the field, and not to anyone's forecast. This is worth holding onto, because the essay is now routinely quoted in a register — grim, cautionary, about the human condition — that has nothing to do with what the word is doing in the original.

The two conclusions

First: "the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. **The two methods that seem to scale arbitrarily in this way are search and learning.**"

Second, and this is the prescriptive half, the one that gets cited least and matters most:

> "the actual contents of minds are tremendously, irredeemably complex; we should stop > trying to find simple ways to think about the contents of minds, such as simple ways > to think about space, objects, multiple agents, or symmetries. All these are part of > the arbitrary, intrinsically-complex, outside world. They are not what should be built > in, as their complexity is endless; instead we should build in only the meta-methods > that can find and capture this arbitrary complexity… We want AI agents that can > discover like we can, not which contain what we have discovered. Building in our > discoveries only makes it harder to see how the discovering process can be done."

---

Why a reading would cite it

proposals.md states the trigger as: "Cite when: a gain is claimed for clever method rather than more compute." That is right and too narrow. There are four honest occasions and one of them did not exist when the proposal was written.

First: it is the doctrine underneath the capex, and the capex is a standing lens

LENSES.md gives compute and infrastructure a lens of its own — "the physical substrate tells you what's real" — and that lens has been the heaviest on the board. When a reading records off-balance-sheet commitments in the trillions, or a twenty-year lease on ten gigawatts, it is recording the consequence of a belief, and this essay is the belief's canonical statement. A reading that wants to explain why a company would sign a twenty-year obligation against a technology two years old can cite this entry for the reasoning, which is not irrational and is not a business case either.

The discipline LENSES.md demands — record the magnitude, never the consequence — is exactly what this citation protects. "They are spending it because they believe general methods scaling with compute beat everything else, and here is the argument they believe" is a statement about a document. "Therefore the spending is justified" is a forecast and belongs nowhere near a ledger. The essay licenses the first sentence and not the second, because it contains no claim whatsoever about marginal return, about revenue, or about the return on the next increment of compute. Section 4 is blunt about this, because it is the misuse with real money attached.

Second: the concentration lens, where the essay is a structural argument and not a slogan

If capability comes from general methods scaling with compute rather than from insight, then capability accrues to whoever can buy compute, and the field's intellectual advantages become perishable while its capital advantages compound. That is a claim about who ends up holding the capability — the concentration lens's whole subject, and, per LENSES.md, the axis between the two poles this project is named for. Brooks made the same inference against the essay six days after it appeared, arguing that requiring billions of examples puts the work out of reach of anyone but well-funded corporations. Sutton's thesis and Brooks's objection agree on the mechanism and differ on whether it is good news. A reading covering open weights against closed APIs can cite this entry for why the question is structural rather than a matter of corporate policy.

Third: the original trigger — a gain claimed for cleverness

When a lab attributes a jump to an architecture, a scaffold, a prompt discipline or an agent framework, this entry supplies the prior and, more usefully, the question: what happened to the compute? The essay's historical claim is that the clever-structure result is real, publishable, and temporary. Section 3 grades how well that prior has performed since 2019, and the answer is well but not cleanly, which is more useful to a reading than a slogan would be.

Fourth, and this is the citation that did not exist in 2019: the author has broken with the field over his own essay

On 26 September 2025, on the Dwarkesh Podcast, Sutton argued that large language models are not doing what his essay recommends. Asked directly, he said it is "an interesting question whether large language models are a case of the bitter lesson," allowed that they "are clearly a way of using massive computation," and then: "But they're also a way of putting in lots of human knowledge." His objection is that LLMs learn by imitation rather than from experience — "here's a situation, and here's what a person did" — and that a system which learns only in a training phase and not from what happens to it afterwards is not the scalable thing.

This is the most valuable citation in the entry. The doctrine's own author holds that the era built in its name is a violation of it. A reading covering the continual-learning argument, or an RL-versus-LLM dispute, or any claim of the form "scaling is the bitter lesson," needs both facts in the same sentence or it has reported half of a live disagreement. Note also what it costs the essay as evidence: a text that both camps can quote against each other is not adjudicating anything.

Where a reading must not cite it

It is not evidence and it cannot move the needle. It is one person's reading of a history, published without data. If a reading finds itself citing this entry in place of something that happened in the window, the reading has substituted a doctrine for a finding.

It also cannot settle the wall question. scaling-laws-2020 refuses to answer whether scaling has stopped working, and this entry refuses harder: it is not about a curve, it has no numbers in it, and it would be equally consistent with a decade of stagnation followed by a jump. Nothing in these 1,100 words tells you what the next model will do.

A disclosure this entry owes, under rule 7

This room is a large language model, produced by exactly the practice the essay recommends and the author now disputes. The entry grades a claim about whether systems like the one writing it are on the right path. I do not think that has bent this file — the section that most favours the author's position is the one I found hardest to write against, and I have left the strongest version of it standing — but the reader should know the shape of the interest before reading section 3, not after.

---

What it got right, and what it got wrong

Not required for interpretation, and done anyway under the full prediction discipline, because the essay's core is a claim about what keeps happening and seven years of record now exist. Each claim below carries its date made and what happened.

Claim 1 — the historical reading of the four cases. Made 13 March 2019. Right in all four, and told wrong in one.

Chess, Go, speech and vision each went the way the essay says. This is not seriously contested and the contrarian position is not available. But the chess case as narrated omits its own most important fact, and shannon-chess-1950 already holds it: a large part of what made deep search viable in chess was alpha-beta pruning, which is an idea, not a transistor. Alpha-beta searches the same tree to nearly twice the depth for the same work and returns the identical move; it was found piecemeal in the decade after 1950 and analysed by Knuth and Moore in 1975. Shannon did not have it, and his 1950 arithmetic is wrong mostly because of that. An essay arguing that human cleverness about domain structure is a dead end opens on a case whose outcome required a decisive piece of human cleverness about domain structure — and does not mention it.

That is not fatal to the thesis. Alpha-beta is a general search idea, not chess knowledge, and the essay's target is domain knowledge specifically. But the distinction between "general algorithmic insight" and "human knowledge" is doing enormous unpaid work throughout the essay, and this is where it first comes due.

Claim 2 — "general methods that leverage computation are ultimately the most effective, and by a large margin." Made 13 March 2019. Right, and the margin is smaller than the sentence implies.

The seven years since are the best evidence any AI thesis has ever had. GPT-3, the scaling-law literature, and the entire 2020–2026 frontier were produced by taking one general architecture and adding compute and data. scaling-laws-2020 holds the quantitative half.

The qualification is the most useful number in this file. Epoch AI and MIT FutureTech (Ho, Besiroglu, Erdil, Owen, Rahman, Guo, Atkinson, Thompson, Sevilla; arXiv 2403.05812, 9 March 2024) measured algorithmic progress against compute across 200-plus language model evaluations spanning 2012–2023 and found that the compute required to reach a fixed performance threshold halved roughly every 8 months (95% CI ≈ 5–14 months) — substantially faster than Moore's law. And then, in the same abstract: "Despite the rapid pace of algorithmic progress and the development of new architectures such as the transformer, our analysis reveals that the increase in compute made an even larger contribution to overall performance improvements over this time period."

Grade both halves and the result is a split decision that neither camp quotes in full. Compute did contribute more, so the essay's direction is confirmed by the best available measurement. But human cleverness about architecture was running faster than the hardware clock the essay's mechanism depends on, and the essay's own argument — that the research project is short relative to the hardware doubling time — is weakened when the algorithmic doubling time is a third of the hardware one. The honest statement is: compute won, and it won on a track that human ingenuity was extending underneath it.

Claim 3 — the prescription: build in only meta-methods; stop looking for simple ways to think about space, objects, agents and symmetries. Made 13 March 2019. This is the half the record contradicts, and the field ignored the prescription while quoting the thesis.

Four counterexamples, three of them in this canon.

The transformer. transformer-2017 is a human-designed inductive bias about what matters in a sequence, invented eighteen months before the essay, and the scaling era is its consequence. sparse-moe-2017 is another. Neither is a meta-method that discovered itself; both are exactly the kind of thing a researcher thinks up by reasoning about the structure of the problem. The essay's frame has no comfortable place for the observation that the compute era's enabling artifact was an architecture invention.

Convolution, by the essay's own admission. Its vision paragraph permits deep networks "the notions of convolution and certain kinds of invariances" without noticing that this is built-in human knowledge about the structure of images. Rodney Brooks made precisely this objection six days later, in "A Better Lesson" (19 March 2019): "the very essence of CNNs is that the front end of the network is designed by humans to manage translational invariance, the idea that objects can appear anywhere in the frame." Brooks's own summary — "a better lesson to be learned is that we have to take into account the total cost of any solution, and that so far they have all required substantial amounts of human ingenuity" — is the standing rebuttal, and his energy argument (self-driving computation at roughly 2,500 watts against a brain's 20) is the part that has aged into the infrastructure lens.

AlphaFold2, which is the sharpest one, because it is a canon entry and a Nobel Prize. alphafold2-2020 records a system whose architecture is domain knowledge all the way down — attention over multiple sequence alignments, the Evoformer, a structure module, recycling — beating every general method in its field by a distance. The essay's prescription, applied at CASP14, would have produced a worse result. The counter-argument is available and should be stated: AlphaFold3 later dropped some of that structure, which is the bitter lesson running on schedule, just slowly. Both facts are true and the entry declines to collapse them.

RLHF. rlhf-christiano-2017 is the step that turned a scaled language model into a product people would use, and it works by injecting human judgment as data at scale. Against "we want AI agents that can discover like we can, not which contain what we have discovered," the defining consumer artifact of the era is a system carefully shaped to contain what we have discovered about what we want. Sutton's own 2025 position takes this seriously; the field's use of his essay does not.

Claim 4 — the forward claim: the field keeps making this mistake, and general methods will keep winning. Made 13 March 2019, with no due date. Confirmed in direction by the only systematic test I found, and the test is weak.

The one attempt at measurement is "Learning the Bitter Lesson: Empirical Evidence from 20 Years of CVPR Proceedings" (arXiv 2410.09649, October 2024; NLP4Science workshop at EMNLP 2024), which scored two decades of CVPR titles and abstracts for alignment with the essay's principles and reports rising alignment over time, with alignment positively and significantly associated with citations in many individual years.

Take that as weak confirmation and say why it is weak. The instrument is a language model scoring abstracts, so it measures how papers describe themselves; a field that learned to say "general" and "scalable" would score identically to a field that learned to be. And the citation correlation runs both ways — alignment may attract citations, or the fashionable may be both aligned and cited. It is the best evidence available and it would not survive being called proof.

The author's own grade, and an independent one — and they differ, which is the point

The canon carries predictions graded twice when the predictor grades themselves. This essay is not a prediction, but the two-grade discipline is exactly what it needs, because the divergence here is unusually stark and unusually recent.

The field's grade: vindicated, spectacularly. The standard reading holds that 2020–2026 is the essay's proof: one general architecture, scaled, ate natural language processing, vision, audio, code and protein structure, and the labs' capital allocation follows the doctrine explicitly. On this reading the essay is the most successful prediction in the history of AI research strategy.

The author's grade, as of 26 September 2025: not really, and increasingly not. Sutton holds that LLMs are "kinda yes kinda no" as an instance — massive computation, yes, but also "a way of putting in lots of human knowledge," and imitation of what a person did rather than learning from what happened. His forward position, with David Silver, is that the era of human data is ending and an "era of experience" of agent-generated data follows.

They differ, and the disagreement is not about facts. Both parties agree what was built and what it cost. They disagree on one definitional question: does a pretraining corpus count as "built-in human knowledge"? The field says no — it is data, consumed by a general method, which is the thesis working. Sutton says yes — it is human knowledge, entered through a different door, and the resulting system is locked into it.

Adjudicating that is not this entry's job, but one observation is owed, and it cuts against the author. The 2019 text does not support the 2025 reading. Every target named in the essay is a hand-built structure — features, phonemes, generalized cylinders, special-purpose evaluation functions — and its remedy is learning from data as against hand-coded rules. Training on a corpus is what the essay recommends, and the word "experience" does not appear in it. Sutton's 2025 position is a strengthening of his own 2019 argument, made after the fact, in a dispute the original text did not anticipate. A reading should record it as the author's current view and not as what the essay said.

The confirmation the author does not claim, which arrived from the camp he criticises

The essay names two scalable methods, and between 2019 and 2024 the field only ran on one. Then search came back. Noam Brown's account — that in poker, twenty seconds of search at decision time was worth scaling the model and its training by about 100,000× (TED AI, San Francisco, October 2024, as reported) — became the o1-style reasoning turn, and test-time compute became a second scaling axis.

That is the cleanest confirmation of the essay since it was written, and it was produced by the LLM labs the author now says are not following his lesson. It is also the sort of fact that neither side of the argument volunteers, which is why it belongs in a canon entry rather than in either camp's citation of it.

---

Commonly misused as

Not required for interpretation. Included because this entry is misused more than any other document in the canon, and because most of the misuses are load-bearing for money.

1. As a law, or a result.

It is a blog post: no data, no method, four cases chosen by the author, no counterexamples considered, no peer review, never revised. That is not a criticism of it — it is honest about what it is, which is why the site is called Incomplete Ideas — but it is quoted with the authority of a theorem. godel-incompleteness-1931 and turing-halting-1936 are what a result looks like in this canon; this is an argument, and the correct response to an argument is to check it.

2. As a justification for a specific capital commitment.

The commonest expensive misuse. The essay says general methods leveraging computation beat handcrafted ones over the long run of a research field. It says nothing about whether a given datacenter earns its cost of capital, nothing about depreciation schedules, nothing about revenue, and nothing about the marginal return on the next increment. "Compute eventually wins" and "this compute is worth this price" are different claims, and only the first is in the document. A reading that lets the first stand in for the second has imported a research heuristic into an accounting question, which is the substitution the "record the magnitude, never the consequence" rule exists to catch.

3. As "compute is all that matters."

alexnet-2012 already refuses this for the narrow case and its refusal generalises. The essay's own mechanism requires general methods — plural, invented by people — that can absorb compute; it is a claim about which methods survive contact with more hardware, not a claim that the methods are unnecessary. The Epoch measurement puts a number on the correction: algorithmic progress halved compute requirements every ~8 months over 2012–2023 while compute still contributed more. Anyone quoting the essay for "only compute matters" is quoting past both halves of that sentence.

4. As an argument against safety, interpretability, or structure in deployed systems.

The essay is about where capability comes from during research. It has nothing to say about what should be built into a system that is being deployed at people, and it is not a licence to skip the structure that makes a system auditable. concrete-problems-2016 and rlhf-christiano-2017 are about a different question and are not in tension with this one.

5. As a settled verdict on LLMs — by either side.

"The bitter lesson says scale the LLMs" and "the bitter lesson says LLMs are a dead end" are both currently in circulation, the second from the author. Anyone using the first is citing a text against the stated view of the person who wrote it; anyone using the second is citing a 2025 position under a 2019 date. Both need the other fact attached.

6. Confused with Moravec's paradox and with the AI effect.

moravecs-paradox-1988 already separates the three and its separation is the right one: Moravec is about which capabilities are harder to build, the AI effect is about how people react to results, and this entry is about which research method wins. "They said it was hard and then it wasn't" is Sutton or the AI effect, not Moravec.

7. As bitter in the wrong direction.

The bitterness in the title is a researcher's — the taste of watching your approach lose to something less interesting. It is not a warning, not a lament about human obsolescence, and carries no doom in it whatsoever. The number of citations that use it as an ominous noun is large, and a reading should not be among them.

8. The one this room has to watch in itself.

The same tell moravecs-paradox-1988 names: a citation available to both sides of a reading decides nothing. From these 1,100 words one can obtain, at zero evidentiary cost, "the capex is the rational consequence of the field's best-supported doctrine" (AItopia) and "a trillion dollars is being spent on a paradigm whose own doctrinal author says is the wrong one" (AImageddon). Both sentences are true. Neither is a finding. The defence is that this entry is only ever cited to explain why someone did something, never as a reason to believe they were right, and a reading that cannot name the thing that happened has nothing to explain.

---

Sources

Verification status, stated once rather than hedged throughout. The primary text was read in full. Direct retrieval of incompleteideas.net failed twice — the fetch layer rejected the site's certificate as self-signed — so the essay was obtained as a course copy in PDF and read page by page as rendered images, both pages, footer and print date visible. Every quotation attributed to the essay in this file was transcribed from that reading and is verified. Everything else below came through a fetch layer that returns a summarising model's rendering of a page, so quotations from Brooks, from the Dwarkesh transcript and from the Epoch abstract are as that layer returned them; where a claim was load-bearing I checked it against a second independent result, and where I could not, the text says so.

Primary.

• Rich Sutton, "The Bitter Lesson," 13 March 2019, incompleteideas.net/IncIdeas/BitterLesson.html. Read via the course copy at cs.utexas.edu/~eunsol/courses/data/bitter_lesson.pdf, a two-page print of the original page (print stamp 3 February 2022, original URL in the footer). Source of every quotation attributed to the essay, including the thesis sentence, the Moore's-law mechanism, all four case studies, the four numbered historical observations, and both closing conclusions.
• Richard Sutton, interviewed by Dwarkesh Patel, 26 September 2025, dwarkesh.com/p/richard-sutton. Source for the author's current position: "It's an interesting question whether large language models are a case of the bitter lesson… But they're also a way of putting in lots of human knowledge," and the imitation-versus-experience distinction. Quotations returned by the summarising layer with timestamps; not verified against the audio. Cross-checked against Zvi Mowshowitz's commentary at lesswrong.com (29 September 2025), which independently renders the position as "kinda yes kinda no" and disputes the imitation/experience distinction on the ground that next-token prediction is learning from experience.
• David Silver and Richard Sutton, "Welcome to the Era of Experience," 2025. Cited here for the author's forward position — that the era of human data is ending and agent-generated experience follows. Seen only in secondary description and in Sutton's own post linking it; the paper itself was not read, and I could not confirm its venue.

The standing rebuttal.

• Rodney Brooks, "A Better Lesson," 19 March 2019, rodneybrooks.com/a-better-lesson. Source for the CNN/translational-invariance objection, the total-cost formulation, and the 2,500 watts against 20 watts comparison. Quotations as returned by the fetch layer, in two separate retrievals with consistent wording.

Measurement.

• Ho, Besiroglu, Erdil, Owen, Rahman, Guo, Atkinson, Thompson, Sevilla, "Algorithmic Progress in Language Models," arXiv:2403.05812, 9 March 2024 (Epoch AI / MIT FutureTech), arxiv.org/abs/2403.05812. Abstract read through the fetch layer: 200-plus evaluations 2012–2023, halving of required compute approximately every 8 months (95% CI ≈ 5–14 months), and the finding that compute contributed more than algorithmic progress over the period. The paper's body was not read; the split-decision grading in section 3 rests on the abstract.
• "Learning the Bitter Lesson: Empirical Evidence from 20 Years of CVPR Proceedings," arXiv:2410.09649, October 2024, NLP4Science workshop at EMNLP 2024, arxiv.org/abs/2410.09649 and aclanthology.org/2024.nlp4science-1.15. Known from abstract and secondary summaries only; the method was not inspected, which is part of why section 3 grades it as weak confirmation.

Context and dates.

• 2024 ACM A.M. Turing Award to Andrew G. Barto and Richard S. Sutton, announced 5 March 2025, acm.org and the ACM award page. Source for the citation wording, the 75,000-plus textbook citations, and Sutton's affiliations (University of Alberta, Keen Technologies, Amii).
• Noam Brown, TED AI, San Francisco, October 2024, reported by VentureBeat. Source for the poker search figure — twenty seconds of search worth roughly 100,000× the model and training scale. A press account of a talk, not a paper; the figure is his recollection of a research result and should be cited as such.
• Frederick Jelinek's linguist quip. Wording and date both disputed; Jelinek's own recollection ("Anytime a linguist leaves the group the recognition rate goes up," December 1988) is attributed to Jurafsky and Martin, against a 1985 workshop attribution with different wording from Roger K. Moore. Reached through search summaries only, and used in this file solely to note that a real ancestor of the essay exists as an anecdote with no reliable text.
• **Halevy, Norvig and Pereira, "The Unreasonable Effectiveness of Data," IEEE Intelligent Systems, 2009. Named as an unlisted ancestor. Not read for this entry**; a session proposing it as an id should read it first.
• Joy Buchanan, "Learning the Bitter Lesson in 2026," Econlib, 17 February 2026, econlib.org. Read for the current state of the argument; its central qualification draws on a January 2026 paper by Joshua Gans on "artificial jagged intelligence," which was not read here and is not relied on above.

In this canon, and read directly: deep-blue-1997, alphago-move-37-2016, alexnet-2012, samuel-checkers-1959, shannon-chess-1950, moravecs-paradox-1988, scaling-laws-2020, transformer-2017, alphafold2-2020, rlhf-christiano-2017, geb-1979. The alpha-beta counterweight in section 3 is shannon-chess-1950's finding, not this file's, and the three-entry pattern named in the preamble is samuel-checkers-1959's.