What Is This?
On March 13, 2019, Rich Sutton — one of the founding figures of reinforcement learning and co-author of the standard textbook on the subject — published a short essay on his website. It was 1,189 words. It may be the most important document about AI written in the 21st century.
He called it "The Bitter Lesson." The lesson: the biggest thing we've learned from 70 years of AI research is that general methods that scale with computation consistently beat methods that encode human knowledge.
Every time. Without exception. And AI researchers keep having to relearn it because it's deeply unsatisfying, because it seems to diminish the value of intellectual effort, because it runs against every instinct that experts have about what should work.
The historical record he assembled is striking:
Chess: When computers first tackled chess, researchers built evaluation functions encoding grandmaster knowledge — piece values, positional heuristics, strategic principles accumulated over centuries. These approaches worked reasonably well. Then came programs that used less human knowledge and more search, looking ahead more moves rather than evaluating positions more cleverly. The search-based approach won. Deep Blue, which beat Kasparov in 1997, was primarily a massive search engine with specialised hardware, not a chess-smart system.
Go: Go was considered impervious to the brute search approach because the branching factor is too high. Researchers spent decades building systems that encoded expert Go knowledge — territory evaluation, influence patterns, joseki sequences. Then DeepMind trained AlphaGo on raw game positions and outcomes, using general self-play reinforcement learning and Monte Carlo tree search at scale. It beat the world champion in 2016. AlphaZero, which learned Go from zero human knowledge (only the rules), became even stronger.
Speech recognition: The dominant approach for decades was phoneme-based models built on linguistic knowledge of how language works. Hidden Markov Models with expert-crafted features. Then deep learning on raw audio with enough data and compute began outperforming all of it. Every expert linguistic feature turned out to be something the network could learn itself, better than the feature engineers had built.
Computer vision: Image recognition systems used engineered features — edge detectors, SIFT features, histogram of oriented gradients — until convolutional neural networks trained on raw pixels at scale swept them away entirely. AlexNet (2012) didn't win by being clever about vision. It won by being large and trained on a lot of data.
The pattern is identical each time: experts build something clever that works, then something general with more compute surpasses it, and the experts are always surprised and resistant.
"We should stop trying to find ways to build in knowledge," Sutton concluded. "The world is vastly more complex than we can understand or model. The only scalable approach is general methods that find and capture that complexity through massive computation."^1
Why Does It Matter?
- It's the principle that explains the entire current AI moment. GPT-3, GPT-4, Claude, Gemini — none of these were built by encoding linguistic knowledge, reasoning rules, or world models. They were built by taking a general architecture (transformers) and scaling it — more parameters, more data, more compute. The Bitter Lesson predicted exactly this in 2019. LLMs are the Bitter Lesson's most extreme vindication to date. The question of whether this continues to hold as we hit new capability ceilings (reasoning, planning, physical world grounding) is the central empirical question in AI.^2
- DeepSeek's story is a Bitter Lesson story. When DeepSeek released R1 in January 2025, the narrative was "Chinese lab beats OpenAI on a budget." The deeper story is more Bitter Lesson: DeepSeek didn't win by being cleverer about AI design. They won by being more rigorous about applying general learning methods efficiently — MoE architecture, efficient RLHF, focused training on reasoning tasks. They proved the lesson works even with constrained compute: general methods at scale beat clever specialisation.
- It's a direct challenge to the instinct that expertise adds value in AI. The Bitter Lesson is psychologically bitter for two groups: AI researchers who built their careers on hand-crafted representations, and builders who believe that clever prompt engineering, custom fine-tuning, or specialised architectures are the path to better AI systems. Sutton's evidence suggests that in the long run, the general learner with more scale will outcompete the expert who bakes in structure. This doesn't mean clever engineering is worthless — it means it has a shelf life measured in months, not years.
- It tells you where to invest effort. If the Bitter Lesson holds, the most durable competitive advantage in AI is not clever technique but access to compute, data, and the ability to scale general methods. This is why the companies with the most compute (Google, Microsoft, Amazon) have structural advantages. It's why data moats (proprietary training data) matter more than algorithm moats. And it's why a solo builder's advantage isn't in out-clevering frontier labs — it's in moving faster in a specific domain where the general models haven't been applied yet.
- The counterargument matters: human structure may be necessary at higher complexity levels. Sutton's own reinforcement learning research acknowledges that learning from scratch in complex environments can be sample-inefficient to the point of infeasibility. AlphaZero needed millions of games of self-play. In domains with expensive real-world feedback, some human structure may be necessary to bootstrap. The debate between "scale everything" and "scale informed by structure" is live and unresolved — the Bitter Lesson is a strong prior, not an absolute law.
Key People & Players
Rich Sutton (University of Alberta / DeepMind) — Author of the essay and co-author (with Andrew Barto) of Reinforcement Learning: An Introduction, the standard textbook in the field. His career has been spent on general learning methods — particularly reinforcement learning — rather than hand-crafted AI. The Bitter Lesson is partly a vindication of his own research programme.^3
Geoffrey Hinton — The primary empirical demonstrator of the Bitter Lesson in deep learning. His lab's work on training deep neural networks on raw data (rather than engineered features) produced the results that demonstrated the principle across vision, speech, and language. His departure from Google in 2023 and subsequent warnings about AI risk are the coda to a career spent proving the Bitter Lesson true.
Ilya Sutskever — Co-founder of OpenAI and the person who most aggressively pursued the scaling hypothesis (the quantitative version of the Bitter Lesson): more parameters + more data + more compute = better models, in a predictable log-linear relationship. GPT-3 and GPT-4 are the direct products of this conviction.
Yann LeCun (Meta AI) — The most prominent critic of the pure Bitter Lesson view. LeCun argues that brute scaling will hit a wall — that general intelligence requires more structured world models, causal reasoning, and physical grounding than pure scale-and-learn can provide. His JEPA (Joint Embedding Predictive Architecture) is the bet against the Bitter Lesson continuing to hold indefinitely.
The Current State
The Bitter Lesson continues to be validated in 2025-2026. Test-time compute scaling (chain-of-thought reasoning, extended inference) is the newest chapter: giving models more compute at inference time — to think longer — produces better outcomes than the same model with less thinking time. It's another version of the same principle: general methods (more compute) beat clever prompting or architectural tricks.
The essay is freely available at Sutton's website. It's 1,189 words and takes 8 minutes to read. It's had more influence on how AI labs think than most 100-page papers.
The uncomfortable implication for anyone building with AI: the tool that beats your clever solution may not arrive with fanfare. It may arrive as a general model that just had more compute thrown at it, trained on more data, released by a lab with more resources. The Bitter Lesson's bitterness isn't just historical — it's a standing prediction about the future.
Best Resources to Learn More
- "The Bitter Lesson" by Rich Sutton (2019) — The essay itself. Eight minutes. Read it.^4
- Wikipedia: Bitter Lesson — Good summary and context, including the empirical follow-up research.^5
- Reinforcement Learning: An Introduction by Sutton & Barto — The textbook that contextualises where the Bitter Lesson comes from as a principle.^6
- Gwern: "The Scaling Hypothesis" — The most thorough analysis of the scaling laws that make the Bitter Lesson quantitatively precise.^7
- Kaplan et al.: "Scaling Laws for Neural Language Models" (2020) — The OpenAI paper that formalised the log-linear scaling relationship that is the Bitter Lesson's mathematical expression.^8