Bitter lesson - Wikipedia

Jump to content

Search

Search

Donate

Create account

Log in

Personal tools

Donate

Create account

Log in

Bitter lesson

1 language

Svenska

Edit links

From Wikipedia, the free encyclopedia

Principle in artificial intelligence

Part of a series onArtificial intelligence (AI)<br>Major goals

Artificial general intelligence<br>superintelligence

Intelligent agent

Recursive self-improvement

Planning

Computer vision

General game playing

Knowledge representation

Natural language processing

Robotics

AI safety

Approaches

Machine learning

Data mining

Symbolic

Deep learning

Bayesian networks

Evolutionary algorithms

Neuro-symbolic AI

Systems integration

Open-source

Open weights

AI data centers

Applications

Art<br>Music

Bioinformatics

Deepfake

Earth sciences

Finance

Generative AI<br>Audio

Images

Government

Healthcare

Industry

Software development

Translation

Military

Physics

Projects

Philosophy

AI alignment

Artificial consciousness

The bitter lesson

Chinese room

Friendly AI

Ethics

Existential risk

Turing test

Uncanny valley

Human–AI interaction

History

Timeline

Progress

AI winter

AI boom

AI bubble

Controversies

Deepfake pornography<br>Taylor Swift deepfake pornography controversy

Grok sexual deepfake scandal

Google Gemini image generation controversy

It's the Most Terrible Time of the Year

Pause Giant AI Experiments

Removal of Sam Altman from OpenAI

Statement on AI Risk

Tay (chatbot)

Théâtre D'opéra Spatial

Voiceverse NFT plagiarism scandal

Glossaryv

The bitter lesson is the observation in artificial intelligence that, in the long run, general approaches that scale with available computational power tend to outperform ones based on domain-specific understanding because they are better at taking advantage of the falling cost of computation over time. The principle was proposed and named in a 2019 essay by Richard Sutton[1] and is now widely accepted.[2][3][4][5][6][7][8]

The essay<br>[edit]

Sutton gives several examples that illustrate the lesson:

Game playing. In chess, the Deep Blue system that became the first computer opponent to defeat a world champion relied on a relatively simple alpha–beta search algorithm that scaled up by applying large amounts of specialized hardware to search for the best move. This defeated previous attempts to exploit the unique structure of chess or to include grandmaster knowledge directly. Likewise in the game of Go, the AlphaGo algorithm that surpassed human performance relied much less on expert skill at the game itself than previous generations of AI, and was further surpassed by AlphaGo Zero, which removed human expertise completely and trained only by self-play.

Speech recognition. Approaches based on training a general-purpose hidden Markov model with large numbers of speech samples consistently outperformed the hand-crafted approaches of the 1970s, and deep learning has continued this trend.

Computer vision. Algorithms that were assumed to approximate the human visual system (such as explicitly encoded edge detection or detecting high-level features with SIFT) were outperformed by convolutional neural networks that make far fewer assumptions about the nature of visual perception.

Sutton concludes that time is better invested in finding simple scalable solutions that can take advantage of Moore's law, rather than introducing ever-more-complex human insights, and calls this the "bitter lesson". He also cites two general-purpose techniques that have been shown to scale effectively: search and learning. The lesson is considered "bitter" because it is less anthropocentric than many researchers expected and so they have been slow to accept it.

Impact<br>[edit]

The essay was published on Sutton's website incompleteideas.net in 2019, and has received hundreds of formal citations according to Google Scholar. Some of these provide alternative statements of the principle; for example, the 2022 paper "A Generalist Agent" from Google DeepMind summarized the lesson as:[2]

Historically, generic models that are better at<br>leveraging computation have also tended to overtake more specialized domain-specific approaches, eventually.

Another phrasing of the principle is seen in a Google paper on switch transformers coauthored by Noam Shazeer:[3]

Simple architectures—backed by a generous computational budget, data set size and parameter count—surpass more complicated algorithms.

The principle is further referenced in many other works on artificial intelligence. For example, From Deep Learning to Rational Machines draws a connection to long-standing debates in the field, such as Moravec's paradox and the contrast between neats and scruffies.[9] In "Engineering a Less Artificial Intelligence", the authors concur that "flexible methods so far have always outperformed handcrafted domain knowledge in the long run" although note that "[w]ithout the right (implicit) assumptions, generalization is impossible".[5]...