The bitter lesson and Moravec’s paradox

Working Definition

The bitter lesson and Moravec’s paradox are two complementary principles that organize the history and limits of artificial intelligence. The bitter lesson, formulated by reinforcement-learning pioneer Richard S. Sutton, states that over the long run, general statistical methods that scale with exponentially growing computation—specifically search and learning—invariably defeat systems into which researchers try to hand-code human domain knowledge. Moravec’s paradox, articulated by roboticist Hans Moravec, identifies an inverted hierarchy of computational difficulty: tasks that require years of conscious study from adult humans (proving theorems, playing chess, drafting legal briefs) are relatively easy for computers, whereas skills that humans and animals perform effortlessly without conscious thought (navigating a cluttered room, folding a towel, catching a ball) prove computationally hardest. Taken together, the two concepts explain the strange landscape of contemporary technology: algorithms pass bar exams and predict protein folding, while household robots still struggle to unload a delicate wine glass from a dishwasher.

Genesis and Global Lineage

Moravec’s paradox grew out of the frustration of 1980s robotics pioneers (Hans Moravec, Rodney Brooks, Marvin Minsky) with the assumptions of symbolic “Good Old-Fashioned AI” (GOFAI). Early AI researchers in the 1950s and 1960s assumed that once computers mastered logical deduction and checkers, computer vision and physical grasping would be a summer student project. In Mind Children (1988), Moravec offered an evolutionary explanation for why the opposite held true: abstract logical reasoning is at most a hundred thousand years old and relies on thin, recently evolved cortical machinery that we experience as effortful precisely because it is inefficient; visual perception and motor control, by contrast, were honed by natural selection over half a billion years and run on massively parallel, unconscious neural hardware that reverse-engineering struggles to match. Three decades later, in March 2019, Canadian computer scientist Richard S. Sutton (2024 Turing Award laureate) published his two-page essay The Bitter Lesson. Surveying seventy years of AI—from chess (Deep Blue, 1997) and Go (AlphaGo and AlphaZero, 2016–2017) to speech recognition and computer vision—Sutton showed a recurring pattern: researchers initially win short-term gains by building human heuristics into their programs, but as Moore’s law cheapens compute by orders of magnitude, brute-force search and tabula rasa learning sweep the expert-crafted rules aside.

Central European Lineage

The mathematical foundation of one of the two pillars of Sutton’s bitter lesson—large-scale probabilistic search across state spaces—traces directly to the Lwów and Budapest schools of mathematics. Between 1946 and 1949 at Los Alamos, Stanisław Ulam—educated at Lwów Polytechnic and formed intellectually at the tables of the Scottish Café (Kawiarnia Szkocka) in interwar Lwów (now Lviv, Ukraine)—joined forces with Budapest-educated John von Neumann to invent the Monte Carlo method. Ulam’s starting intuition was a striking precursor to Sutton’s: while convalescing from encephalitis and playing solitaire, Ulam realized that instead of laboriously deriving closed-form combinatorial equations for the probability that a deal is winnable, it was far more effective to have the new ENIAC electronic computer play out thousands of random sample games and count the statistical outcomes. Seventy years later, a direct descendant of Ulam and von Neumann’s paradigm—Monte Carlo Tree Search (MCTS) coupled with deep neural networks—enabled DeepMind’s AlphaZero to master Go and chess from scratch without any human opening books. A methodological note on lineage also belongs here: although Hans Moravec was born in 1948 in Kautzen, Austria, he emigrated to Canada at age five and completed all his schooling in North America; his case illustrates why intellectual lineage is defined by formative academic environment rather than birthplace alone.

Why Now

Between 2019 and 2026, the collision between the bitter lesson and Moravec’s paradox defined the central strategic debate in artificial intelligence. Sutton’s March 2019 essay coincided with the release of GPT-2 (2019), GPT-3 (2020), and the Kaplan neural scaling laws (2020): frontier AI labs treated the bitter lesson as a multi-billion-dollar mandate to build giant GPU clusters and discard hand-engineered features. Between 2024 and 2026, with the arrival of reasoning models (OpenAI o1/o3, DeepSeek-R1), Sutton’s second pillar—search—returned to center stage via test-time compute, where models spend thousands of inference tokens searching chains of thought before answering. Simultaneously, Moravec’s paradox exacted revenge on labor-market forecasts: whereas 2010s economists predicted that automation would first displace truck drivers, plumbers, and warehouse pickers while sparing white-collar analysts, poets, and programmers, the post-2022 generative AI wave inverted that sequence—automating cognitive desk work first while dexterous physical robotics remained bottlenecked by sensorimotor data.

Explanatory Power

Pairing the two concepts illuminates three major technological trajectories. First, natural language processing: for decades computational linguists hand-coded phrase-structure grammars and morphological exceptions (prompting speech researcher Fred Jelinek’s famous quip: “Every time I fire a linguist, the performance of the speech recognizer goes up”); when those rules were replaced by Transformers trained on trillions of raw tokens, the network learned syntax and pragmatics far more robustly than any manual grammar. Second, structural biology: for half a century biophysicists tried to compute protein folding from explicit physical force fields, until AlphaFold 2 (2020) solved the problem by learning evolutionary co-variation patterns across massive sequence databases. Third, illustrating Moravec’s paradox, autonomous driving: a program that defeats the world chess champion requires a tiny fraction of the compute needed for a self-driving car on a rainy night to tell a wind-blown plastic bag from a dog on the curb—because chess is a closed, discrete world, whereas the street demands half a billion years of evolved sensorimotor grounding.

Axes of Debate

Underappreciated Angle

Why is Sutton’s lesson actually “bitter”? Most commentators read the bitterness as a wound to researchers’ professional vanity: it hurts to watch an elegant algorithm that took a decade of human ingenuity get crushed by matrix multiplication on ten thousand GPUs. Yet there is a deeper epistemological bitterness: the decoupling of predictive power from human scientific understanding. For three centuries after Newton, technological mastery marched hand in hand with comprehension—to build a better bridge or steam engine, scientists discovered compact causal laws that fit inside a human mind. The bitter lesson reveals that in domains of high messy complexity (natural language, cellular biology, atmospheric physics), reality refuses to compress into a handful of human-legible axioms; a model with a trillion opaque floating-point weights outperforms human theories precisely because it abandons the simplifications required for human intelligibility. The price of superhuman predictive engineering is that the builders themselves no longer understand step by step how their creation solves the problem.