In March 2019, a 63-year-old professor at the University of Alberta published a brief essay on his personal website. It had no peer review, no institutional backing, no promotion. Within months, it had become one of the most debated texts in artificial intelligence, a kind of manifesto that made some researchers defensive and others triumphant. Rich Sutton called it "The Bitter Lesson," and its argument was simple enough to sting: the history of AI proves that researchers who try to encode human knowledge into systems are wasting their time. What works is scale. What works is compute. What works is letting machines learn for themselves, given enough data and processing power.

Sutton's essay ran to roughly 800 words. It surveyed decades of AI development—chess, speech recognition, computer vision—and found the same pattern repeating. In each domain, researchers initially made progress by building in human expertise: handcrafted features, linguistic rules, domain-specific heuristics. And in each domain, those approaches were eventually demolished by methods that simply threw more computation at the problem and let learning algorithms figure out the rest.

The chess precedent

Sutton's canonical example was computer chess. For years, the field was dominated by systems that tried to encode grandmaster knowledge—opening books, positional evaluation rules, endgame databases compiled by human experts. These programs were impressive engineering achievements. They were also dead ends. When IBM's Deep Blue defeated Garry Kasparov in 1997, it relied heavily on specialized hardware that could evaluate hundreds of millions of positions per second. The human knowledge baked into its evaluation function mattered less than the sheer brutality of its search. Two decades later, DeepMind's AlphaZero learned to play chess at a superhuman level by playing against itself millions of times, starting from nothing but the rules. It discovered strategies that human grandmasters had never conceived. The bitter lesson had repeated itself.

Why it stings

The essay's title is precise. The lesson is bitter because it offends the self-image of researchers. Most people who enter AI do so because they find intelligence fascinating and want to understand it. They bring theories about how cognition works, how language is structured, how perception operates. Sutton's argument suggests that all this intellectual apparatus is, at best, a temporary scaffold—useful until compute catches up, then discarded. The researchers who spent careers hand-engineering features for image recognition watched as convolutional neural networks, trained on vast datasets, rendered their work obsolete. The linguists who built elaborate syntactic parsers watched as transformer models learned grammar implicitly, without being told what a noun was.

This is not an argument that human intelligence is unimportant. It is an argument that human intelligence is poorly suited to the task of encoding itself into machines. The things we know explicitly—the rules we can articulate—turn out to be the least valuable parts of cognition. What matters is what we cannot easily express: the statistical regularities of the world, learned through exposure.

The lesson's limits

Sutton's thesis is not without critics. Some argue that he overstates the case, that algorithmic innovations—attention mechanisms, residual connections, reinforcement learning from human feedback—have been essential to progress, not merely incidental. Others point out that scale has costs: the environmental burden of training enormous models, the concentration of power among organizations that can afford the compute, the opacity of systems that learn their own representations. The bitter lesson may describe what has worked, but it does not prescribe what should work. There are domains where interpretability matters, where human oversight is non-negotiable, where the brute-force approach is socially unacceptable even if technically superior.

And yet, the pattern keeps asserting itself. Large language models have achieved capabilities that seemed decades away, and they have done so not through careful knowledge engineering but through scale: more parameters, more data, more compute. Researchers who bet against Sutton's thesis have repeatedly lost.

Our take

The bitter lesson is uncomfortable because it suggests that AI progress may be less about insight than about resources—that the field's future belongs not to the cleverest theorists but to whoever can marshal the most silicon. Sutton himself seems unbothered by this implication; he is, after all, a pioneer of reinforcement learning, a field premised on letting agents learn from experience rather than instruction. But for those who entered AI hoping to crack the mystery of mind, his essay reads like a rebuke. The mystery, it turns out, may not need cracking. It may just need computing.