# Seven years of debate over AI's most cited blog post Rich Sutton's "The Bitter Lesson" has become the unofficial scripture of the scaling era — required reading at OpenAI, cited in DeepMind papers, and invoked in hundreds of arguments about AI's future. Yet the essay's apparent simplicity masks a remarkably contested intellectual object. In the seven years since publication, a rich ecosystem of criticism, qualification, extension, and reappropriation has emerged — and Sutton himself has significantly evolved his position. The discourse reveals not just disagreements about AI strategy, but deep unresolved questions about the nature of knowledge, the limits of scaling, and whether the lesson's greatest vindication (large language models) may actually be its most important counterexample. ## The critics drew blood early and their arguments still stand Within six days of publication, **Rodney Brooks** (MIT emeritus, co-founder of iRobot) posted "A Better Lesson" (March 19, 2019, rodneybrooks.com/a-better-lesson/), the most widely cited rebuttal. Brooks mounted six numbered objections, but the core thrust was that Sutton had drawn a false boundary around "human knowledge." CNNs encode translational invariance through human-designed convolution. Network architectures themselves are products of human ingenuity. Even specialized AI hardware represents "simply shifting where human intelligence needs to be applied." Brooks' summary: "The lesson we should learn from the last seventy years of AI research is not at all that we should just use more computation and that always wins. Rather I think 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." **Thomas Dietterich** (former AAAI president, Oregon State) offered a diplomat's synthesis in a comment on Brooks' post: "Rod and Rich are both right... the top-performing systems will always combine knowledge engineering with brute force." **Andy Kitchen** wrote a thoughtful "Meta Lesson" reconciling both positions, arguing that the real trajectory is one of continuously pushing human ingenuity to higher levels of abstraction — "in the past we hand-crafted features, now we hand-craft architectures... we should be always seeking to do more with learning and search." The architecture-as-knowledge objection proved the most durable line of attack. **Beren Millidge** (AI researcher, beren.io, July 2020) made the case precisely: "CNNs, LSTMs, transformers, ResNets, attention, initialization schedules, SGD, and a whole bunch of empirical tweaking has been the primary driver of continued success, not simply more compute." He noted that basic RNNs "fail miserably compared to transformers at language modelling even with substantially greater compute and data." **Shimon Whiteson** (Oxford, now Waymo) argued in a March 2019 Twitter thread that complexity "necessitates, rather than obviates, human ingenuity." **Jason McEwen** (geometric deep learning researcher) sharpened this further in his Substack "Inductive Bias," arguing that encoding symmetries is not injecting "human knowledge" in the pejorative sense — "they are the stage upon which the play of reality is performed. Ignoring them isn't avoiding human bias — it's ignoring the structure of the universe." **Michael Nielsen** (physicist, author of *Neural Networks and Deep Learning*) published "Reflections on 'The Bitter Lesson'" in July 2021 (cognitivemedium.com/bitter-lesson) with two particularly sharp observations. First, he documented that **Deep Blue was actually a hybrid system** using approximately 8,000 human-engineered evaluation features, making Sutton's use of it as evidence for pure computation misleading. Second, he noted that Sutton's qualifier "ultimately" renders the thesis unfalsifiable: "'Ultimately' isn't a very informative stance. Does it mean in 5 years? In 20 years? In 100 years?" **Gary Marcus** (NYU emeritus, neuroscientist) has been the essay's most persistent critic across dozens of Substack posts, tweets, and media appearances from 2019 through 2026. His core position is that the Bitter Lesson is a "weak inductive argument" that is "true of the past, in particular domains" but doesn't generalize. Marcus frames his alternative as "the bittersweet lesson" — scaling delivers **System 1**-type heuristics (fast pattern matching), but System 2 understanding (compositional, causal reasoning) requires fundamentally different methodologies. In November 2025, Marcus wrote in Newsweek: "The bitter lesson about the bitter lesson is that it only works for some problems." **François Chollet** (creator of Keras, founder of ARC-AGI benchmark) provided perhaps the sharpest empirical challenge. He observed that LLMs were scaled by approximately 10,000× from 2019 to 2024, yet their scores on his ARC benchmark — designed to test genuine novel reasoning — "stayed near 0 (e.g. GPT-4o at ~5%). Meanwhile a very crude program search approach could score >20% with hardly any compute." Chollet's distinction between **skill** and **intelligence** cuts at the heart of the Bitter Lesson: "You are increasing the skill of the system... but not its intelligence because skill is not intelligence." ## Sutton's own position has evolved dramatically Sutton has not retracted the Bitter Lesson — he has extended it into a far more radical thesis. His evolution proceeded through several stages. The **Alberta Plan** (arXiv:2208.11173, August 2022), co-authored with Michael Bowling and Patrick Pilarski, doubled down on the original thesis while redirecting it. The plan explicitly states: "Although there are many ways human input and domain knowledge can be used to improve the performance of an AI, such methods typically do not scale with computational resources and as such are not a research priority for us." But it also introduced a crucial emphasis on **continual learning from ordinary experience** rather than static training on curated datasets. This thread culminated in "**Welcome to the Era of Experience**" (April 2025), a vision paper co-authored with **David Silver** (DeepMind VP of Reinforcement Learning, Sutton's former PhD student). The paper argues that "the knowledge extracted from human data is rapidly approaching a limit" and proposes that the next era of AI will be defined by agents learning from their own experience through reinforcement learning, developing "non-human reasoning methods" and building their own world models. On the **Dwarkesh Podcast** (September 26, 2025), Sutton went further, calling LLMs a "dead end" for achieving true intelligence. He stated: "I consider reinforcement learning to be basic AI. What is intelligence? The problem is to understand your world. Whereas large language models are about mimicking people, doing what people say you should do. They're not about figuring out what to do." He dismissed next-token prediction as a goal: "That's not a goal. It doesn't change the world. Tokens come at you and if you predict them, you don't influence them." In September 2023, Sutton joined **John Carmack's Keen Technologies** as Research Scientist. Carmack framed their shared position against the mainstream: "The AI space is awash in capital, compute, and data, but it is still dominated by fashions that may yet hinder important breakthroughs." At the **Upper Bound 2025** conference and **RLC 2025**, Sutton presented his OaK (Options and Knowledge) architecture — a manifesto for building agents that learn through lifetime interaction rather than static pretraining. The net effect is paradoxical. **The Bitter Lesson's most famous advocate now argues that the Bitter Lesson's most famous vindication (LLMs) is a dead end.** Sutton sees the industry as having learned only the *first* bitter lesson (scale beats hand-coding), while missing the *second*: that static data training, even at enormous scale, cannot produce genuine intelligence. Only experiential, goal-directed learning can do that. ## How the scaling community adopted — and arguably distorted — the thesis The Bitter Lesson became institutional doctrine at the major AI labs. OpenAI reportedly **required engineers to memorize it**. The essay's influence can be traced through a clear intellectual lineage. **Jared Kaplan** (Anthropic co-founder, CSO) and collaborators at OpenAI published "Scaling Laws for Neural Language Models" (January 2020), establishing power-law relationships between model size, dataset size, compute, and loss. This formalized Sutton's qualitative observation into quantitative law. Kaplan initially found that performance scaled more steeply with model parameters than with data, which was read as vindication of the "just scale up" interpretation. At his 2025 Y Combinator talk, Kaplan stated: "So many times in my experience of the last five years when it seemed like scaling was broken, it was because we were doing it wrong." **DeepMind's Chinchilla paper** (Hoffmann et al., March 2022) corrected Kaplan's exponents, showing that model size and training data should be scaled roughly equally — approximately **20 tokens per parameter**. As Kushal Chakrabarti argued in his influential "The Bitter Lesson Is Misunderstood" (Obviously Wrong Substack, August 2025), Chinchilla revealed that compute scales quadratically with dataset size (C ≈ 6N·D, and N ≈ D, so C ≈ D²). "If you double your GPUs, you also have to increase your data by ~40% or you're just lighting cash on fire." Chakrabarti's key reframing: "**The Bitter Lesson lays out the spirit of the laws, but the Scaling Laws specify their letter. And the letter revealed that the god of Compute we worship is actually yoked to an even greater one — the god of Data.**" This data constraint is where the thesis encounters its most interesting complications. **Microsoft's Phi model series** (beginning with "Textbooks Are All You Need," Gunasekar et al., June 2023) demonstrated that a **1.3 billion parameter model** trained on "textbook quality" synthetic data could outperform models 10× its size on coding benchmarks. The Phi-1.5 paper explicitly states this "challenges the prevailing notion that the capabilities of LLMs are solely determined by their scale." Yet the challenge is subtler than it appears: the "textbook quality" data was generated by GPT-3.5 and GPT-4, meaning the quality curation was itself a product of computation at scale. As Chakrabarti notes, this is "the Alchemist's path" — still consistent with the Bitter Lesson if you view the entire pipeline. **Ari Morcos** (CEO of Datology, formerly Meta FAIR) describes his own "bitter lesson" moment around 2020 as realizing "all that really matters is the data." Datology achieved baseline performance **12× faster** through data curation alone, "bending the naive scaling laws." This aligns with **Andrew Ng's** data-centric AI movement (launched 2021): "In the last decade, the biggest shift in AI was a shift to deep learning. I think it's quite possible that in this decade the biggest shift will be to data-centric AI." Ng advocates for domain experts improving data quality — seemingly the kind of human knowledge injection Sutton warned against, but applied to the training signal rather than the model architecture. ## The unanswered question about pre-filtered training data One question from the user's prompt deserves special attention because it identifies what appears to be a genuinely underexplored argument. The idea that the Bitter Lesson works partly because training data is **already pre-filtered for quality by domain-internal processes** — peer review in science, competitive play in games, communicative success in language, market forces in economic text — has not been explicitly articulated as a coherent thesis in the published literature. Several authors gesture at adjacent ideas. **Andrew Trask** (OpenMined, 2025, openmined.org/blog/the-bitter-lessons-bitter-lesson/) argues that human-generated text represents "the compressed output" of billions of years of evolution, and that training on it inherits the quality of that evolutionary process — but he frames this as a computational inheritance argument, not a domain-internal filtering argument. The **CACM article** "Data Quality May Be All You Need" (2024) notes that "filtering data using some basic metrics of quality and training across two epochs produced better results than training the full dataset over just one epoch." **Assaf Pinhasi** (Medium, July 2025) observes that the LLM lifecycle uses progressively more explicit domain knowledge, with even the "self-supervised" phase depending on data produced by institutional quality processes. But **no one has made the specific philosophical argument that the Bitter Lesson's success is partly parasitic on quality guarantees already embedded in the training distribution by domain-internal selection pressures**. This appears to be a genuinely novel observation. ## RLHF sits at an awkward boundary the essay cannot easily resolve The relationship between RLHF and the Bitter Lesson has been explicitly debated. **Christopher Rytting** (Scale Venture Partners, April 2023) provided the most direct analysis: "Maybe the more refined takeaway is that models must learn for themselves as **gently steered by rather than entirely built by** human designers." He also warned that "the truth is that humans were driving all along, despite ideas you might get from The Bitter Lesson that we backed ourselves out of the process." Two interpretations coexist. RLHF can be read as hand-encoding human preferences (violating the Bitter Lesson) or as providing a learning signal that computation then amplifies (consistent with the Bitter Lesson). **Jianing Qi** (Medium, 2025) frames RL as "search" — one of Sutton's two scalable methods — "optimizing policies computationally, searching beyond what data explicitly provides." The trend toward **RLAIF** (AI feedback replacing human feedback) and **RLVR** (reinforcement learning from verifiable rewards) pushes the field progressively toward the Bitter Lesson's preference for computation over human knowledge. But the initial human signal remains, and its importance may be irreducible. ## The essay has migrated far beyond engineering into philosophy and epistemology The most intellectually interesting developments involve the Bitter Lesson's implications for what knowledge *is*. **Steven Normore** (snormore.dev, February 2025) makes the connection explicit: "Understanding is compression — finding the shorter description of a complex phenomenon... A handcrafted heuristic is a human-compressed version of the problem. Newton's laws compress thousands of observations into F = ma. The chess grandmaster compresses thousands of games into 'control the center.' These compressions are genuinely powerful. But they're **human-scale compressions**, limited by what we can hold in our heads." He connects this to Herbert Simon's **bounded rationality**: "We build heuristics because we can't search exhaustively. That's bounded rationality... Our shortcuts are good. Evolution and experience gave us useful ones. But they're workarounds for a compute limitation." A blog post from Engineers of AI (Medium, March 2026) sharpens this further: "Human knowledge, encoded as inductive bias, is a **ceiling rather than a floor**. We gain performance short-term by baking in what we know. We pay for it long-term by preventing the system from discovering what we do not know." **Elan Moritz** (PhilArchive, July 2025) provides the most explicitly philosophical treatment, arguing that the Bitter Lesson instantiates what he calls **Operational Constructivism** — a synthesis of Peircean pragmaticism, Bridgman's operationalism, and constructivism. Moritz argues that "knowledge is validated not by descriptive accuracy but by constructive success" and that the debate between hand-crafted and outcome-trained AI "is a choice between two different epistemologies." His paper was written in response to **Ethan Mollick's** (Wharton) "The Bitter Lesson versus The Garbage Can" (OneUsefulThing, July 2025), which asked whether AI agents trained on outputs alone could bypass the need to understand organizational processes. **Michael Stoyanovich** published the most sustained philosophical engagement: "The Human Lesson: A Response to Sutton through Wittgenstein, Lewis, Dennett, and Nagel" (Version 1.6.2, November 2025). He deploys Dennett's distinction between **competence and comprehension**: "A system without comprehension cannot furnish reasons — only results. This makes it dangerously prone to amplify existing biases or enact morally arbitrary decisions, all while performing flawlessly." He invokes Nagel's phenomenological argument: "For our machines, there is nothing it is like. Even if we grant a system the ability to report internal states, that is not equivalent to experiencing them." **Drew Breunig** (dbreunig.com, August 2025) connects the Bitter Lesson to **Donna Haraway's** framework of successive "blows to the human ego" — the Copernican revolution, Darwinism, Freud's unconscious — positioning AI's displacement of expertise as the latest such blow. **Eva Dyer et al.** ("Accepting 'the bitter lesson' and embracing the brain's complexity," The Transmitter, 2025) apply the lesson to neuroscience itself, while noting a crucial caveat: "We have no optimization metric for scientific insight. The path to understanding runs through human creativity and rigorous analysis." Notably, **formal academic philosophy** has been slow to engage. The Moritz paper sits on PhilArchive; the Stoyanovich paper is self-published. No major peer-reviewed philosophy journal has published a dedicated analysis of the Bitter Lesson's epistemological implications. The connection to **Hubert Dreyfus's** phenomenological critique of AI — which argued that skilled expertise is grounded in embodied, tacit "knowledge-how" that resists formalization — is natural but has not been drawn explicitly in the published literature. ## The AI safety community has forked the lesson into three incompatible readings The Bitter Lesson has become a contested terrain in alignment debates, generating at least three distinct positions. **Adam Karvonen, Richard Ren, Dan Hendrycks** and others published "The Bitter Lesson for AI Safety Research" (LessWrong/arXiv:2407.21792, August 2024), a meta-analysis finding that roughly half of AI safety benchmarks are highly correlated with general capabilities. Their practical recommendation: focus safety resources on problems that scale does *not* solve — bias, dynamic adversarial robustness, sycophancy, and crucially, **weapons of mass destruction knowledge** (which is *negatively* correlated with capabilities). An Alignment Forum post titled "The Sweet Lesson: AI Safety Should Scale with Compute" (2025) takes the opposite approach, arguing that effective safety techniques must themselves be scalable: deliberative alignment through chain-of-thought, AI control through game-theoretic protocols, and debate protocols where longer exchanges yield more reliable honesty assessments. The darkest reading comes from **Adam Elwood's** "The Second Bitter Lesson" (LessWrong/EA Forum): "To solve alignment, we are currently assuming that we can develop some clever techniques which will allow us to both understand what an AI is planning and make sure its incentives go in the direction we want. But again we're making the mistake of thinking we will be able to understand what's going on in detail." Elwood argues that natural selection dynamics will override carefully engineered alignment in sufficiently autonomous AI systems — making the Bitter Lesson a prediction that alignment itself may be futile. The Bitter Lesson is included as **required reading in Week 1** of the AGI Safety Fundamentals curriculum, alongside Ngo's "AGI safety from first principles," indicating its foundational status in the safety discourse. ## Gwern, emergence, and the scaling hypothesis pipeline **Gwern Branwen's** "The Scaling Hypothesis" (gwern.net/scaling-hypothesis, first published May 2020, updated through 2025) explicitly extends the Bitter Lesson into a comprehensive theory: "The bitter lesson is the harder and bigger, the better." Gwern identifies what he calls "**blessings of scale**": stability → generalization → meta-learning. His scaling hypothesis — that "intelligence is 'just' simple neural units & learning algorithms applied to diverse experiences at a (currently) unreachable scale" — traces an intellectual lineage from Hans Moravec's compute forecasts through Sutton to the present. The **emergent abilities** literature (Jason Wei et al., "Emergent Abilities of Large Language Models," 2022) initially appeared to validate the strongest version of the thesis: scaling produces qualitatively new capabilities that "cannot be predicted simply by extrapolating the performance of smaller models." But **Rylan Schaeffer et al.** ("Are Emergent Abilities of Large Language Models a Mirage?", Stanford, 2023) argued that emergence is an artifact of metric choice, not fundamental model behavior — still validating smooth improvement with scale, but undermining the narrative of miraculous qualitative leaps. Perhaps the most telling development is the **emerging counter-narrative from the scaling movement's own champions**. **Ilya Sutskever** (co-founder of OpenAI, now CEO of Safe Superintelligence), in his 2025 Dwarkesh Podcast interview, stated: "We're moving from the age of scaling to the age of research" and "Is the belief that if you just 100x the scale, everything would be transformed? I don't think that's true." This represents a notable shift from the man who, more than perhaps anyone, operationalized the Bitter Lesson into the GPT scaling trajectory. ## Conclusion: the lesson's meaning keeps shifting under our feet Seven years on, the discourse around the Bitter Lesson reveals something more interesting than a simple right-or-wrong verdict. The essay functions as a **Rorschach test for AI philosophy**: scaling maximalists cite it to justify ever-larger training runs; critics use it to expose the hidden human knowledge in every "general" method; Sutton himself now wields it against the very systems most people consider its vindication. Three genuinely novel insights emerge from the full discourse. First, the strongest version of the Bitter Lesson — that *all* human knowledge injection is counterproductive — has been empirically refuted by the symmetry-encoding results in geometric deep learning and the Phi model series, but the **weaker, longitudinal version** (that over decades, scalable methods tend to displace fixed human heuristics) remains largely unscathed. Second, the pre-filtering argument — that the Bitter Lesson's success is partly parasitic on quality guarantees already embedded in training data by domain-internal selection pressures — represents a genuine gap in the literature that, if developed, could reconcile the data-centric and scale-centric camps. Third, the essay's deepest implication may be epistemological rather than engineering: if general methods consistently surpass human domain knowledge, then that knowledge may be best understood not as insight into the structure of problems, but as **lossy compression under the constraints of bounded rationality** — useful, but fundamentally limited by the hardware it runs on. The Bitter Lesson began as a blog post about AI engineering. It has become a live argument about the nature of knowledge itself.