the bitter lesson
In plain wordsThe argument, made by AI pioneer Rich Sutton in 2019, that AI has advanced mostly through general methods that use ever more computing power, not by programming in human knowledge.
The thesis, set out by Richard SuttonRichard Sutton, prop. n.: A Canadian computer scientist who, with Andrew Barto, laid the groundwork for teaching machines by trial, error and reward, and whose essay 'The Bitter Lesson' argues that methods that scale with computing power always win. in a short 2019 essay, that seventy years of AI research teach one overriding lesson: 'general methods that leverage computation are ultimately the most effective', and decisively so. Sutton described a recurring cycle. Researchers encode what they know about a domain — hand-crafted chess knowledge, linguistic rules, engineered visual features — and gain ground for a while, then stall, and are overtaken as computing grows cheaper by methods that simply exploit more of it, chiefly search and learning; his examples were chess, Go, speech recognition and computer vision. The lesson is bitter because it devalues the researcher's insight into the problem. The essay became a touchstone for the scaling hypothesisscaling hypothesis, n.: The bet that today's AI methods, simply made bigger with more computing power and data, will keep getting smarter — possibly all the way to human-level intelligence and beyond — without needing fundamentally new ideas. and for laboratories' heavy investment in computecompute, n.: The raw computing power used to train and run AI, usually measured by how many specialized chips are used and for how long. With data and algorithms, it is a main driver of AI progress..
Instead of hand-writing parsing rules, we took the bitter lesson to heart and trained on ten times more data.
Origin
Posted by Rich Sutton on his personal website, Incomplete Ideas, on 13 March 2019. Sutton shared the 2024 Turing AwardTuring Award, prop. n.: Nicknamed the 'Nobel Prize of computing', the top honor in computer science, awarded each year since 1966 by the Association for Computing Machinery. Several AI pioneers have received it. with Andrew BartoAndrew Barto, prop. n.: An American computer scientist who, with his student Richard Sutton, laid the foundations of reinforcement learning, the study of how machines learn by trial and error from rewards. for foundational work in reinforcement learning.
Usage note
Critics argue that architectures such as convolution and attention are themselves built-in knowledge, and that scaling faces limits of data and energy. Sutton has said that language models trained mainly on human-generated text do not fully follow his lesson.
Related
- scaling hypothesisscaling hypothesis, n.: The bet that today's AI methods, simply made bigger with more computing power and data, will keep getting smarter — possibly all the way to human-level intelligence and beyond — without needing fundamentally new ideas.
- scaling lawsscaling laws, n.: Measured patterns showing that an AI model's error shrinks smoothly and predictably as the model is made bigger and given more data and computing power, so labs can forecast a giant model's performance from small, cheap experiments.
- computecompute, n.: The raw computing power used to train and run AI, usually measured by how many specialized chips are used and for how long. With data and algorithms, it is a main driver of AI progress.
- Richard SuttonRichard Sutton, prop. n.: A Canadian computer scientist who, with Andrew Barto, laid the groundwork for teaching machines by trial, error and reward, and whose essay 'The Bitter Lesson' argues that methods that scale with computing power always win.
- neats and scruffiesneats and scruffies, phr.: A joking split of AI researchers into 'neats', who want clean mathematical theories that can be proved correct, and 'scruffies', who build messy collections of tricks that simply work.
- deep learning revolutiondeep learning revolution, n.: The rapid takeover of AI by many-layered neural networks from about 2012, when they suddenly began beating older methods at recognizing images and speech, leading to today's AI systems.
Sources
- Sutton, R. S. (2019). The Bitter Lesson. Incomplete Ideas (personal website), 13 March 2019.
Mentioned in
algorithmic progressalgorithmic progress, n.: Improvements in AI methods — not faster chips — that let a model reach the same ability with less computing power. Researchers estimate this has cut the computing needed by half every eight to sixteen months., architecturearchitecture, n.: The overall design of an AI model, meaning what kinds of parts it has and how they are connected, as opposed to the specific numbers it learns during training., deep learning revolutiondeep learning revolution, n.: The rapid takeover of AI by many-layered neural networks from about 2012, when they suddenly began beating older methods at recognizing images and speech, leading to today's AI systems., end-to-end learningend-to-end learning, n.: Training one AI system to go straight from raw input, such as camera images, to the final output, such as steering commands, instead of chaining hand-designed steps., inductive biasinductive bias, n.: The assumptions built into a learning method that decide how it fills the gaps between the examples it was shown, such as that similar inputs get similar answers. Every learner needs some to cope with new cases., Richard SuttonRichard Sutton, prop. n.: A Canadian computer scientist who, with Andrew Barto, laid the groundwork for teaching machines by trial, error and reward, and whose essay 'The Bitter Lesson' argues that methods that scale with computing power always win., scaling hypothesisscaling hypothesis, n.: The bet that today's AI methods, simply made bigger with more computing power and data, will keep getting smarter — possibly all the way to human-level intelligence and beyond — without needing fundamentally new ideas., Software 2.0Software 2.0, n.: Andrej Karpathy's 2017 name for software whose behavior is learned from examples by a neural network instead of being written line by line by programmers., Turing AwardTuring Award, prop. n.: Nicknamed the 'Nobel Prize of computing', the top honor in computer science, awarded each year since 1966 by the Association for Computing Machinery. Several AI pioneers have received it., Deep ThoughtDeep Thought, prop. n.: The giant computer in The Hitchhiker's Guide to the Galaxy that thinks for 7.5 million years and answers the meaning of life with '42'; also an early chess computer named after it., era of experienceera of experience, n.: The prediction, made by David Silver and Richard Sutton in 2025, that AI will soon learn mainly from its own experience of acting in the world rather than from human-written data., hardware lotteryhardware lottery, n.: The idea that research ideas succeed partly by luck — because they happen to suit the computer chips available at the time — rather than purely because they are better., neats and scruffiesneats and scruffies, phr.: A joking split of AI researchers into 'neats', who want clean mathematical theories that can be proved correct, and 'scruffies', who build messy collections of tricks that simply work., -pilled-pilled, adj.: A slang word ending meaning someone has been won over to a belief, often suddenly and completely; 'AGI-pilled' means convinced that human-level AI is coming soon., reward is enoughreward is enough, phr.: The claim, made by DeepMind researchers in 2021, that maximizing a single reward in a rich enough world could by itself produce all the abilities we associate with intelligence., X is all you needX is all you need, phr.: A joking title formula copied from a famous 2017 AI paper, used to claim that one simple ingredient is enough to solve a problem; hundreds of research papers now use it.