The One Sentence Behind the AI Boom
On March 13, 2019, Richard Sutton posted a short essay to his personal website. Sutton is not a hype man. He is one of the founders of reinforcement learning, co-author of the field's standard textbook, and later a Turing Award winner. The essay is barely over a page. It has arguably shaped more billion-dollar decisions than any business book published that decade.
Its opening line 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."
Read that twice, because it's less obvious than it looks. Sutton is not saying compute helps. He's saying that in the long run it's the only thing that reliably wins, and that the human knowledge we lovingly build into our systems tends to get in the way. The methods that scale with raw computation beat the methods that encode what smart people already figured out. Not by a little. By a large margin.
That claim sounds almost anti-intellectual. It says the domain expert, the person who understands the special structure of the problem, is usually the one who loses. And Sutton didn't argue it from theory. He argued it from six decades of AI researchers making the same mistake, watching it fail, and making it again.
Sixty Years, Four Times the Same Mistake
The power of the essay is that it's a pattern, not an opinion. Sutton walks through four fields where the same movie played out.
Chess. For decades, computer-chess researchers tried to encode human chess understanding: opening theory, positional heuristics, the wisdom of grandmasters. Then in 1997 IBM's Deep Blue beat world champion Garry Kasparov using massive, deep search, evaluating hundreds of millions of positions per second. It didn't "understand" chess the way a grandmaster does. It searched. The knowledge-based approaches lost to raw look-ahead running on special-purpose hardware.
Go. Go was supposed to be different. The board is far too large for brute-force search, so surely this was the game where human intuition would stay essential. Then, about twenty years after Deep Blue, DeepMind's AlphaGo beat Lee Sedol in 2016, and its successor AlphaZero learned superhuman Go, chess, and shogi from scratch through self-play, with no human games at all. Same story, same ending: search plus learning beat encoded human knowledge.
Speech recognition. Back in the 1970s, DARPA ran a speech-recognition competition. One camp used human knowledge of words, phonemes, and the human vocal tract. The other used statistical methods and hidden Markov models that did more computation and knew less about language. The statistics won. Decades later, deep learning took the same lesson further, and modern speech systems contain almost no hand-coded linguistics.
Computer vision. Early vision research searched for edges, corners, and clever hand-designed features like SIFT. Then in 2012 a deep neural network known as AlexNet cut the error rate on the ImageNet benchmark from roughly 26% to about 16% in a single leap, using convolution, GPUs, and data rather than human-designed features. The hand-crafted pipeline that a generation of researchers had refined was gone within a few years.
| Field | The human-knowledge approach | What actually won | Turning point |
|---|---|---|---|
| Chess | Grandmaster heuristics and opening books | Massive deep search | Deep Blue beats Kasparov, 1997 |
| Go | Human intuition and shape knowledge | Search plus self-play learning | AlphaGo beats Lee Sedol, 2016 |
| Speech | Phonetics and vocal-tract models | Statistics and hidden Markov models | DARPA competitions, 1970s onward |
| Vision | Hand-designed features like SIFT | Deep neural networks on GPUs | AlexNet on ImageNet, 2012 |
Four fields, four sets of experts, one outcome. Each time, the people who understood the problem best backed the approach that lost.
Why They Called It "Bitter"
Here's the part people skip. The lesson isn't just "compute wins." It's that the win feels bad, and that the bad feeling causes smart people to keep resisting it.
When Deep Blue won with search, Sutton notes, the reaction from human-knowledge chess researchers was not curiosity. It was something closer to sour grapes. They had built careers on the idea that chess mastery required deep positional understanding, and a machine that mostly searched had just embarrassed that idea. In his words, they "were not good losers."
The pattern has an emotional logic. Building your own knowledge into a system is satisfying. It makes the system work better today, and it makes you feel like the improvement came from your insight. That short-term payoff is real, which is exactly why it's a trap. You get rewarded now for the very move that plateaus later.
Sutton's summary of the cycle is worth holding in your head:
- Researchers build human knowledge into their systems.
- It helps, and it feels good, in the short term.
- Over time it plateaus and can actively block further progress.
- The breakthrough eventually comes from the opposite direction: scaling computation through search and learning.
The lesson is bitter because it never stops being tempting to ignore. Every new field arrives convinced that this time the human insight is essential, that this problem is special. So far, that conviction has a losing record.
Search and Learning: The Only Two Things That Scale
If human knowledge is the thing that plateaus, what's the thing that doesn't? Sutton names exactly two: search and learning.
Both share a rare property. They get better when you throw more computation at them, and they keep getting better. Search explores more possibilities per second as hardware improves. Learning extracts more structure from more data as models and compute grow. Neither one depends on a human having already figured out the answer, which is why neither one hits the same ceiling.
This is where Sutton's argument turns philosophical, and where his best line lives. The reason to stop hand-coding knowledge, he writes, is that "the actual contents of minds are tremendously, irredeemably complex." The way we think about space, objects, other agents, and symmetry is far messier than any tidy rule we could write down. So when we try to compress our understanding into a system, we're shipping a lossy, oversimplified cartoon of real intelligence.
His prescription follows directly. Don't try to build in the contents of thought. Build in the meta-methods that can find that complexity on their own. As he puts it, we want AI agents "that can discover like we can, not which contain what we have discovered." Give the system the ability to search and to learn, then get out of the way and let scale do the work you were never going to do by hand.
For anyone who cares about knowledge, that reframe is bigger than AI. It suggests the highest-leverage move isn't packing more finished answers into your head. It's building a system that can keep discovering. That idea connects directly to why knowledge compounds like interest when you set it up to keep growing.
The Modern Proof: Scaling Laws and Self-Play
The reason the 2019 essay reads like prophecy is that the years right after it turned into one long confirmation.
In 2020, a team led by Jared Kaplan published "Scaling Laws for Neural Language Models," showing that a model's performance improves as a smooth power law as you increase model size, data, and compute. Bigger and more consistently meant better, across orders of magnitude. That paper became the intellectual license for the entire race to scale up large language models. In 2022, DeepMind's Chinchilla work refined the recipe, showing that many models had been undertrained on data relative to their size, and rebalancing the ratio.
The through-line from Deep Blue to GPT is the same lesson wearing new clothes. AlphaZero learned world-class play from self-play alone, no human game records required, which is search and learning at full volume. Language models learned staggering competence not from hand-written rules of grammar and reasoning but from predicting the next token across an ocean of text on enormous compute budgets. Every lab that bet on scaling over cleverness, from OpenAI to DeepMind to Anthropic, was, whether they cited him or not, running Sutton's playbook.
The compounding engine underneath all of it is exponential growth in available computation. Sutton's whole argument leans on the assumption that compute keeps getting cheaper, so any method that can absorb more of it eventually wins. That's a close cousin of the superlinear returns that Paul Graham describes in startups: pick the thing that compounds, and time does the heavy lifting.
Where the Bitter Lesson Bites Back
A good explainer has to say where the idea is contested, because it is. Treating "The Bitter Lesson" as gospel is its own kind of mistake.
The most cited rebuttal came fast. In March 2019, roboticist Rodney Brooks published "A Better Lesson." His point: the human-knowledge camp isn't actually dead, it just moved. The convolutional neural networks that "won" computer vision have translational invariance built in by human designers. Model architectures, tokenizers, and training setups are full of human priors. We didn't remove human cleverness, Brooks argued. We relocated it into the structure of the learning machine and stopped noticing. Others added that Moore's law is slowing, so "just add compute" is not free forever.
Then the strangest twist of all: Sutton himself. In 2025, in the essay "Welcome to the Era of Experience" with David Silver and in a widely discussed interview on Dwarkesh Patel's podcast that September, Sutton argued that today's large language models don't actually satisfy the bitter lesson. LLMs learn by imitating a fixed pile of human text, which he called learning by mimicry rather than learning from experience. Real intelligence, in his newer view, comes from agents that learn continually from their own interaction with the world, not from predicting the next human word. The father of the bitter lesson looked at the LLM era and said, in effect, this still has too much human knowledge in it.
So the honest reading is layered. The core claim, that scalable search and learning beat hand-coded knowledge, has an extraordinary track record. But "which method scales best" is still a live fight, and the person who wrote the essay thinks the current champion is a way station, not the destination. Holding both ideas at once is the mark of actually understanding it, the kind of thinking-in-tradeoffs that good range across fields tends to produce.
The Bitter Lesson of Note-Taking
Now the part that isn't about AI at all. The bitter lesson has a personal-knowledge version, and once you see it you can't unsee it.
Look at how most people try to get smarter. They build elaborate systems. Nested folder hierarchies. Intricate tagging taxonomies. Color codes, templates, a perfect ontology for where every idea should live. This is the human-knowledge approach applied to your own brain. It feels productive. Organizing is satisfying in exactly the way Sutton warns about, and it helps a little at first. Then it plateaus, because the system takes more effort to maintain than it ever gives back, and most people quietly abandon it.
The scalable alternative mirrors search and learning. Capture a lot, with almost no upfront structure, and let retrieval do the organizing later. Instead of pre-deciding where a highlight belongs, save everything and rely on full-text search and AI to surface the right thing when you need it. You bet on volume plus good retrieval instead of clever manual filing. That's the personal echo of "general methods that leverage computation."
| Approach | Human-knowledge style | Scalable style |
|---|---|---|
| Organizing | Elaborate folders and tag taxonomies decided up front | Capture everything, organize at retrieval time |
| Effort curve | High maintenance, helps early, plateaus | Low friction, compounds as the corpus grows |
| Finding things | Remember where you filed it | Search and AI surface it on demand |
| Failure mode | System collapses, you abandon it | Works better the more you save |
This is why a highlighting habit beats a filing habit. When you use Glasp's web highlighter to mark what matters as you read, or save YouTube video summaries and transcript highlights from talks and lectures, you're building the corpus. The volume is the point. Then Glasp's AI chat plays the role of search and learning, letting you ask questions across everything you've saved instead of manually remembering where each idea lives. You supply the meta-method, capture, and let scale handle the rest.
It also reframes a guilty feeling most readers know well. A giant backlog of saved-but-unread material isn't failure, it's raw compute for later. We made the full case for that in why saving now and reading never still makes you smarter. The bitter lesson agrees: a big messy corpus plus good retrieval beats a small, perfectly curated one.
How to Apply It to Your Own Learning
The essay is about machines, but the operating advice for humans is surprisingly concrete.
Stop over-engineering the system. The hours you spend perfecting your tagging scheme are the note-taking version of hand-coding chess heuristics. Satisfying, and a plateau. Keep the structure light enough that you'll actually keep using it.
Lower the cost of capture to almost zero. Scalable methods win partly because they're cheap to feed. Make saving a highlight, a quote, or a summary a one-click reflex, so the friction never talks you out of it. A smart-notes workflow works best when capture is effortless and connection happens later.
Invest in retrieval, not filing. Your leverage is in how well you can find and recombine what you saved, not in how neatly you stored it. Search, AI chat over your own notes, and good linking beat any folder tree.
Bet on compounding, not on finishing. You will never "finish" learning a field, and the bitter lesson says you shouldn't try to bake in a final answer anyway. Build the habit that keeps discovering. A corpus that grows a little every day quietly crosses thresholds you can't predict.
Stay a generalist about methods. The lesson's deepest move is intellectual humility: assume your current best insight will be outscaled, and design for that. Keep your inputs wide. The most durable edge is being the person who keeps learning, which ties back to the context you accumulate over time becoming your real advantage.
The meta-point is simple. Be the meta-method. Don't try to be the finished, hand-coded expert system. Be the thing that keeps searching and keeps learning, and let volume plus good retrieval compound in your favor.
Frequently Asked Questions
What is the bitter lesson in simple terms?
It's the observation, from Richard Sutton's 2019 essay, that in AI the approaches that win over time are the ones that use more computation through search and learning, not the ones that hand-code human expertise into the system. Cleverness helps in the short term and then plateaus; scale keeps improving. It's "bitter" because it means the expert's carefully built knowledge usually loses.
Who is Richard Sutton and why does he matter?
Richard Sutton is a computer scientist widely regarded as a founder of reinforcement learning, co-author of the field's standard textbook with Andrew Barto, and a Turing Award winner. When he writes one page about the direction of AI, the entire field reads it, because he helped build the methods that the modern AI boom relies on.
Is the bitter lesson actually true?
The historical pattern is strong: chess, Go, speech recognition, and computer vision all followed it, and scaling laws for language models extended it. But it's debated. Rodney Brooks argued human knowledge just moved into model architectures rather than disappearing, and in 2025 Sutton himself argued that today's large language models don't fully follow the lesson because they imitate human text instead of learning from real experience.
Does the bitter lesson mean human knowledge is useless?
No. It means human knowledge tends to lose as a way to hard-code the final answer into a system. Humans still choose the problem, design the meta-methods, build the training setups, and decide what to scale. The lesson is about where to put your effort: into scalable methods and good structure, not into freezing today's understanding into the machine.
How does the bitter lesson apply to note-taking and learning?
It suggests that elaborate manual organization systems behave like hand-coded expertise: satisfying at first, then a maintenance burden that plateaus. The scalable alternative is to capture a lot with minimal structure and rely on search and AI for retrieval. Build the corpus with tools like a web highlighter and let retrieval, not filing, do the work.
Conclusion: Bet on What Compounds
"The Bitter Lesson" is one page that quietly explains a decade of AI, and it lands as a warning about how any of us tries to get smarter. The seductive move is always to encode what you already know and call it progress. The durable move is to build something that keeps searching and keeps learning, then feed it scale.
You can't out-compute a data center, but you can copy the strategy. Capture more than feels tidy. Keep the structure light. Put your energy into retrieval, and let a growing body of highlights, notes, and summaries compound over years. That's the personal version of leveraging computation, and it beats a perfectly organized system you stop using by March.
Start building your corpus with Glasp's web highlighter and YouTube Summary, then let Glasp's AI chat search and learn across everything you've saved. Be the meta-method. Let scale do the rest.