II · THE IDEA · ARTIFICIAL INTELLIGENCE
Why Hand-Built Knowledge Lost
▶ Listen · narrated
The systems that won encoded almost no domain knowledge at all. They learned from examples instead, and the lesson has shaped two decades of research priorities.
At a glance
- Central claim
- General methods that scale with computation beat approaches that rely on human knowledge
- Named by
- Richard Sutton, researcher in reinforcement learning
- Empirical support
- Scaling laws show smooth power-law improvements with model size and data
- Time horizon
- Pattern observed across speech, vision, game-playing and language over decades
Imagine you are teaching someone to recognise birds. You could write down rules: if it has a red breast and is small, it is a robin; if it has a curved beak and is large, it is a hawk. Or you could show them thousands of labelled photographs and let them learn the patterns. The first method is faster to start and you can explain every decision. The second method takes more examples and more time, but it keeps improving as long as you give it more pictures, and it will notice details you forgot to mention. The bitter lesson is that the second method, given enough data and computation, has beaten the first method in domain after domain — speech, vision, games, language — even though it feels less intelligent because it cannot tell you the rules it is using.
The claim is that methods which scale with computation — larger models, more data, longer training — have consistently outperformed methods that rely on encoding human knowledge, even when the knowledge is correct. The mechanism is that hand-coded features or rules impose a structure that may fit the data well initially but cannot adapt as the data distribution shifts or the task becomes more complex. A learned representation, by contrast, can allocate capacity wherever the data demands it. The scaling laws documented by Kaplan et al. formalise this: loss decreases as a power law in parameters N, dataset size D, and compute C, with exponents that are empirically stable across orders of magnitude. The relationship is L(N) ∝ N^(-α), with α around 0.076 for models trained to convergence, meaning you need to increase model size by roughly an order of magnitude to halve the loss. The laws also show that there is an optimal allocation of compute between model size and training duration, and that undertraining a large model is better than fully training a small one. The objections are practical and philosophical: learned models are expensive to train, opaque in their decision-making, and dependent on the quality and bias of the training data. They also require infrastructure that not every organisation has. Whether the trade-off favours learning or encoding depends on whether you have access to large datasets, whether you need interpretability, and whether the task structure is stable enough that hand-coded knowledge remains valid.
Look closer
The argument is about leverage, not capability
Sutton's essay does not claim that human knowledge is worthless. It claims that methods which incorporate human knowledge tend to stop scaling earlier than methods which learn from data, because the knowledge becomes a ceiling rather than a foundation. When compute was scarce, hand-coded expertise gave better results per operation. As compute became abundant, the systems that could use more of it — by learning from larger corpora or exploring more game positions — pulled ahead.
The scaling laws are empirical fits, not theorems
Kaplan and colleagues found that loss decreases as a power law in model parameters, dataset size and compute budget, with exponents that can be estimated from smaller runs. The fits are smooth across several orders of magnitude, which is what makes them useful for planning. But they are observations about a particular architecture family trained in a particular way on particular data, not proofs that the relationship holds indefinitely. The paper itself notes that the trends must eventually break.
The objection is that general methods waste resources and miss structure
Critics point out that learning everything from scratch requires enormous datasets and energy expenditure to rediscover regularities that humans already know. A model trained on language will eventually learn something like syntax, but it does so by seeing millions of examples where a linguist could have written the constraints in an afternoon. The counterargument is that the hand-written constraints are always incomplete, often wrong in edge cases, and expensive to maintain as the domain shifts. Which cost matters more depends on what resources you have and what problem you are solving.
The story
In the 1980s, expert systems were the dominant paradigm in artificial intelligence. Teams of knowledge engineers would interview domain experts — doctors, geologists, loan officers — and encode their decision rules in formal logic. The resulting systems could explain their reasoning step by step, and they worked well enough in constrained settings that companies built businesses around them. The assumption was that intelligence meant having the right knowledge, and that progress meant accumulating more of it.
Then the pattern started to break. In speech recognition, systems built on hidden Markov models and large corpora of recorded speech outperformed systems built on phonetic rules. In computer vision, statistical methods trained on labelled images outperformed systems built on hand-designed feature detectors. In game-playing, search algorithms that evaluated millions of positions beat systems encoded with chess strategy. The shift was not immediate and it was not uniform, but by the early 2000s it was clear that the approaches winning competitions were the ones that learned from data rather than the ones that encoded expertise.
Richard Sutton, a researcher in reinforcement learning, named the pattern in a 2019 essay he called The Bitter Lesson. The lesson is that general methods that leverage computation are ultimately more effective than methods that leverage human knowledge. The bitterness is for researchers who spent careers building in domain expertise, only to watch it become a liability. Sutton argues that the history of AI is a history of researchers learning this lesson, resisting it, and learning it again in the next domain.
The scaling laws documented by Kaplan and colleagues in 2020 gave the pattern a quantitative form. They trained language models of different sizes on different amounts of data and found that performance improved as a smooth power law in each resource. Double the parameters and the loss drops by a predictable amount. Double the dataset and it drops again. The relationship held across three orders of magnitude, which suggested that the returns to scale were not about to stop. If you had more compute, the correct move was to build a bigger model and train it on more data, not to spend time encoding what you knew about language.
But the argument has serious objections. One is about efficiency: learning from scratch is wasteful when the structure is already known. A second is about safety: systems that learn everything from data inherit all the biases in the data, with no easy way to impose constraints. A third is about interpretability: a model that has learned millions of correlations is harder to audit than a system built from rules you wrote. And a fourth is about the future: the scaling laws are empirical observations, not physical limits, and no one knows whether they will hold for another three orders of magnitude or break next year.
Why it mattered then
The essay appeared in 2019, at the end of a decade in which deep learning had moved from a minority interest to the dominant research programme. Sutton was not describing a new phenomenon — he traces the pattern back to the 1990s — but naming one that had become undeniable. The timing mattered because the field was still absorbing the implications. Academic labs that had spent years building symbolic reasoning systems were retooling to train neural networks. Companies were hiring machine learning engineers instead of knowledge engineers. The essay gave researchers a frame for understanding why their expertise had stopped being the bottleneck, and why the skills that mattered now were the ones that let you use more data and more compute. It was also, as the title suggests, a lament. Sutton had watched colleagues invest in approaches that turned out not to scale, and he was arguing that the field should stop making the same mistake.
Why it matters now
The argument shapes resource allocation across the industry. If you believe the scaling laws will continue, you invest in data pipelines, compute infrastructure and model architectures that can use both. If you believe they are near a ceiling, you invest in ways to encode structure, priors and constraints. The current generation of large language models is a bet on the first view, and their capabilities have so far supported it. But the objections have not gone away. The cost of training frontier models is now high enough that efficiency matters again, and researchers are exploring ways to incorporate structure without giving up the ability to scale. The debate is also about what kind of understanding we want from these systems. A model that has learned to translate by reading millions of sentence pairs knows something about translation, but it may not know it in a way we can inspect, correct or trust in a high-stakes setting. Whether that matters depends on the application, which is why both approaches persist.
The surprising detail
Sutton's essay is often cited as a vindication of scale, but the essay itself is more conflicted than that. He describes the lesson as bitter precisely because he thinks researchers, including himself, keep wanting to build in their understanding of the problem. The desire is not irrational — it feels like progress, and in the short term it often is — but history suggests it does not compound. What compounds is the ability to use more computation. The essay is less a celebration than a resignation, and Sutton ends it by saying that the lesson has to be learned again and again because it goes against human instinct. We want our knowledge to matter.
Remember this
General methods that scale with compute have historically beaten methods that rely on human knowledge, but the pattern is empirical, not inevitable.
Test yourself
The scaling laws show smooth power-law improvement across several orders of magnitude. Name two distinct reasons why this does not guarantee that building a model ten times larger will be worth the cost.
First, the laws describe loss on a training objective, not performance on the task you actually care about. A model with lower perplexity may still fail at reasoning, instruction-following or factual accuracy in ways that matter more than the headline number. Second, cost scales faster than capability. If loss drops as a power law, then each additional unit of improvement costs more than the last, and at some point the marginal gain is not worth the marginal expense — especially if you are paying for inference at scale, where a larger model costs more per query for its entire deployed lifetime. A third reason: the laws are fits to data from a particular architecture and training recipe, and they may not survive changes to either.
Go deeper
- The Bitter Lesson · incompleteideas.net
- Scaling Laws for Neural Language Models · arXiv · Jared Kaplan et al. · 2020-01-23
Image: Original diagram, The Daily Triptych. Licence: Original work. Source.