# Peter Norvig

> Source: https://aiwiki.ai/wiki/peter_norvig
> Updated: 2026-07-24
> Fact-checked: 2026-07-24
> Categories: AI History, Artificial Intelligence, Google, People
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Peter Norvig." aiwiki.ai, 24 Jul 2026. https://aiwiki.ai/wiki/peter_norvig
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Peter Norvig is an American computer scientist known above all as co-author, with [Stuart Russell](https://aiwiki.ai/wiki/stuart_russell), of *Artificial Intelligence: A Modern Approach*, the textbook that has shaped how [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) is taught since its first edition in 1995 [1][6]. He directed Google's core search algorithms group from 2002 to 2005 and then its research organization, and before that headed the Computational Sciences Division at NASA Ames Research Center [2][3].

Outside the textbook, Norvig is associated with three things: an argument that large quantities of ordinary data beat clever models, made in the 2009 essay "The Unreasonable Effectiveness of Data" with Alon Halevy and Fernando Pereira [4]; a long public disagreement with Noam Chomsky over whether statistical models count as science [5]; and the free online AI class he taught with [Sebastian Thrun](https://aiwiki.ai/wiki/sebastian_thrun) at [Stanford University](https://aiwiki.ai/wiki/stanford_university) in the fall of 2011, which registered roughly 160,000 students and helped set off the MOOC boom [2][11][13].

He is a Distinguished Education Fellow at the Stanford Institute for Human-Centered AI [10]; his own site lists a 2025 Stanford course, CS 139: Human-Centered Artificial Intelligence [1]. That site also still lists a research director role at [Google](https://aiwiki.ai/wiki/google), while a June 2025 interview called him a former director of research there [1][16].

## Early life and education

Norvig was born on December 14, 1956 [3]. He took a bachelor of science in applied mathematics at Brown University and a Ph.D. in computer science from [UC Berkeley](https://aiwiki.ai/wiki/uc_berkeley) in 1986, supervised by Robert Wilensky, with a dissertation titled "A Unified Theory of Inference for Text Understanding" [2][3].

His undergraduate years surface in the argument he became known for. The 2009 data essay opens by recalling that "One of us, as an undergraduate at Brown University, remembers the excitement of having access to the Brown Corpus, containing one million English words" [4]; Norvig took his own bachelor's degree at Brown [3].

## Career before Google

Norvig taught at the University of Southern California as an assistant professor and held research faculty positions at Berkeley [2][3]. In industry he was a senior scientist at Sun Microsystems Laboratories, chief designer at Harlequin, and chief scientist at Junglee, an early internet comparison-shopping service where he says he was employee number 8 [2][3].

Two books came out of this period. *Verbmobil: A Translation System for Face-to-Face Dialog*, written with Martin Kay and Jean Mark Gawron and published by CSLI Publications in 1994, surveyed speech-based [machine translation](https://aiwiki.ai/wiki/machine_translation) and the difficulty of joining [speech recognition](https://aiwiki.ai/wiki/speech_recognition) to language processing [23]. *Paradigms of Artificial Intelligence Programming: Case Studies in Common Lisp* runs to roughly 950 pages and rebuilds classic [symbolic AI](https://aiwiki.ai/wiki/symbolic_ai) systems in [Common Lisp](https://aiwiki.ai/wiki/common_lisp) while teaching advanced language technique; Morgan Kaufmann published it in 1992, although the publisher's current catalogue entry dates the first edition to October 1991 [3][22][27].

## NASA Ames

Norvig headed the Computational Sciences Division at NASA Ames Research Center, later renamed the Intelligent Systems Division, overseeing about 200 scientists working on autonomy, robotics, automated software engineering, and data analysis [3].

By his own account his division developed the Remote Agent experiment, built with the Jet Propulsion Laboratory and flown on the Deep Space 1 spacecraft [2]. It was the first artificial-intelligence control system to run a spacecraft without human supervision, and during the mission it handled three simulated failures: reactivating a failed electronics unit, recognizing and ignoring bad sensor data, and compensating for a stuck attitude-control thruster [24]. The system won the 1999 NASA Software of the Year award, and Norvig received NASA's Exceptional Achievement Award in 2001 [2].

## Artificial Intelligence: A Modern Approach

The textbook Norvig wrote with Stuart Russell of Berkeley has appeared in four editions over 25 years.

| Edition | Year | Notes |
| --- | --- | --- |
| 1st | 1995 | Prentice Hall [8][9] |
| 2nd | 2003 | Prentice Hall [8][9] |
| 3rd | 2009 | Prentice Hall [8][9] |
| 4th | 2020 | Released April 28, 2020; 1,136 pages [9] |

The authors' site describes it as the most-used AI textbook, adopted by more than 1,500 schools [6]. Norvig's biography page puts sales at over 500,000 copies [2]. The book's editions page lists translations into more than a dozen languages, among them Chinese, French, German, Greek, Hungarian, Italian, Japanese, Kazakh, Korean, Persian, Portuguese, Russian, Serbian, and Spanish [8].

The fourth edition is organized into 28 chapters across seven parts. Chapter 2 introduces intelligent agents; later chapters cover search, adversarial games, constraint satisfaction, logic and [knowledge representation](https://aiwiki.ai/wiki/knowledge_representation), automated planning, probabilistic reasoning and [Bayesian networks](https://aiwiki.ai/wiki/bayesian_network), probabilistic programming, multiagent decision making, [machine learning](https://aiwiki.ai/wiki/machine_learning), [deep learning](https://aiwiki.ai/wiki/deep_learning), [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning), [natural language processing](https://aiwiki.ai/wiki/natural_language_processing) both classical and neural, [computer vision](https://aiwiki.ai/wiki/computer_vision), and robotics. The final part contains a chapter titled "Philosophy, Ethics, and Safety of AI" and one on the future of the field [7].

The fourth edition remains the current one [6]. Norvig has said publicly that he doubts the format still fits the subject: "I don't think the 'one textbook' model works anymore. What we need is something interactive and personalized" [16].

## Google

Norvig ran Google's core search algorithms group, work in [information retrieval](https://aiwiki.ai/wiki/information_retrieval), from 2002 to 2005, and the research group from 2005 [2][10]. In 2012 he described the company's research model with Alfred Spector and Slav Petrov in "Google's Hybrid Approach to Research" (*Communications of the ACM* 55(7), pages 34-37), which set out how Google integrates research with product development and weighed the benefits and risks of that arrangement [18]. The ACM elected him a fellow in 2006 [3].

## The unreasonable effectiveness of data

"The Unreasonable Effectiveness of Data" appeared as an Expert Opinion column in *IEEE Intelligent Systems* for March/April 2009, pages 8 to 12, credited to Alon Halevy, Peter Norvig, and Fernando Pereira of Google [4]. The title borrows from Eugene Wigner's 1960 paper "The Unreasonable Effectiveness of Mathematics in the Natural Sciences," and the essay opens by noting that human behavior has resisted the elegant equations physics enjoys [4].

The argument is that the biggest wins in language technology, statistical speech recognition and statistical machine translation, came from tasks where enormous input-output data already exists in the wild, not from tasks that are intrinsically easier [4]. The authors contrast the one-million-word Brown Corpus with the trillion-word corpus Google released in 2006, which carried frequency counts for every sequence up to five words long, and observe that the newer corpus is full of spelling errors and ungrammatical fragments but a million times larger [4]. Their summary line is blunt: "But invariably, simple models and a lot of data trump more elaborate models based on less data" [4].

Much of the essay is about [n-gram](https://aiwiki.ai/wiki/n-gram) models and memorization. The authors note that statistical translation systems consist mostly of large memorized phrase tables and add general rules only where those rules improve on memorizing particular phrases, and that simple n-gram models or linear classifiers built on millions of specific features perform better than elaborate models that try to discover general rules [4]. The closing advice reads: "So, follow the data. Choose a representation that can use [unsupervised learning](https://aiwiki.ai/wiki/unsupervised_learning) on unlabeled data, which is so much more plentiful than labeled data" [4]. In a 2025 interview Norvig extended the same reasoning to video, calling it a comparatively unbiased record of how the world actually behaves and the data source he was most interested in next [16].

## The exchange with Chomsky

At the Brains, Minds, and Machines symposium held during MIT's 150th anniversary in 2011, Noam Chomsky criticized machine learning researchers who use statistical methods to mimic behavior without explaining it. Norvig's essay "On Chomsky and the Two Cultures of Statistical Learning" opens from *Technology Review*'s report of the event, then turns to the symposium transcript to quote Chomsky directly: "There is a notion of success ... which I think is novel in the history of science. It interprets success as approximating unanalyzed data" [5].

Norvig's reply works through five disagreements: that engineering success is evidence that something is working and not scientifically irrelevant; that fact-gathering is ordinary science rather than mere butterfly collecting; that a model with billions of parameters can still yield insight through study of where and how it fails; that probabilistic models are needed because language is a noisy, ambiguous, contingent biological and cultural process; and that Gold's 1967 theorem does not establish innate grammar, since probabilistic learners escape the setting the theorem assumes [5]. He also quotes Chomsky's 1969 line that "it must be recognized that the notion of 'probability of a sentence' is an entirely useless one, under any known interpretation of this term" [5].

The essay is built from a run of blunt section headings, among them "What did Chomsky mean, and is he right?", "What is a statistical model?", and "How successful are statistical language models?" [5]. Its closing section, "The two cultures," borrows [Leo Breiman](https://aiwiki.ai/wiki/leo_breiman)'s distinction between data modeling, which assumes a simple generative model underneath, and algorithmic modeling, which accepts a complex and uninterpretable function so long as it predicts well [5].

## The 2011 online AI class

In the fall of 2011 Norvig and Sebastian Thrun taught Stanford's Introduction to Artificial Intelligence to 175 students on campus and more than 100,000 people over an interactive webcast [11]. Registration reached roughly 160,000 students from about 190 countries [2][14]. Norvig described the experiment in a TED2012 talk called "The 100,000-student classroom" [11].

The consequences arrived quickly. Thrun co-founded Udacity with David Stavens and Mike Sokolsky; the company was announced at the 2012 Digital Life Design conference and launched its first two courses in February 2012 [12]. Andrew Ng and Daphne Koller started Coursera, and 2012 became the year massive open online courses entered the mainstream [13]. Norvig went on to teach Design of Computer Programs (CS212) on Udacity, and the Stanford AI class was relaunched there as CS271 in November 2012 [1][12]. At Google he founded the Course Builder project, an open source package for building online classes [2].

## Programming writing and open source

Norvig's short technical essays are widely read outside academia. "Teach Yourself Programming in Ten Years," written in 2001 as a rebuttal to books promising mastery in 24 hours, argues that expertise takes about a decade of deliberate practice, citing psychology research going back to an 1899 study by Bryan and Harter; it has been translated into more than 20 languages and, by his own count, read by over two million people [2][20].

"How to Write a Spelling Corrector" (2007) came about when two accomplished engineers, both of whom used Google's spelling correction daily, turned out to have no idea how it worked. Norvig wrote a probabilistic corrector in about half a page of [Python](https://aiwiki.ai/wiki/python), then measured it honestly: 75 percent correct on a 270-word development set at 41 words per second, and 68 percent on a 400-word final test set at 35 words per second [19].

Other pieces on his site include a constraint-propagation Sudoku solver, Lispy (a Scheme interpreter in Python), JScheme, the letter-frequency study "English Letter Frequency Counts: Mayzner Revisited," and "The Gettysburg PowerPoint Presentation," a satire of slide-deck culture [1]. His pytudes repository, described as short Python programs of considerable difficulty meant to perfect particular skills, has drawn more than 24,000 stars on GitHub and was still receiving new notebooks in 2026 [26].

## Views on AGI and AI safety

In October 2023 Norvig and Blaise Agüera y Arcas published "Artificial General Intelligence Is Already Here" in *Noema Magazine*, arguing that frontier [large language models](https://aiwiki.ai/wiki/large_language_model) including [ChatGPT](https://aiwiki.ai/wiki/chatgpt) and its competitors already qualify as [artificial general intelligence](https://aiwiki.ai/wiki/artificial_general_intelligence) [15]. Their test is generality along five axes: topics, tasks, modalities, languages, and instructability, the last meaning the ability to pick up a new task from a prompt alone [15]. The central claim is that "frontier language models can perform competently at pretty much any information task that can be done by humans, can be posed and answered using natural language, and has quantifiable performance" [15]. They compare the situation to the 1945 ENIAC, recognized as the first true general-purpose electronic computer only in hindsight, and they concede that current systems fabricate scholarly citations and court cases and carry biases forward; flawless performance, they write, would require superintelligence rather than general intelligence [15].

On [AI safety](https://aiwiki.ai/wiki/ai_safety) Norvig has argued for layered defenses rather than a single guarantee, comparing the situation to ordinary locks that keep casual intruders out while determined attackers find a way around them [17]. He has also reframed [hallucination](https://aiwiki.ai/wiki/hallucination) as a control problem rather than a defect: "a synonym for a hallucination is creativity," with the real task being to teach a system when to invent and when to report [17]. A 2025 podcast appearance carried the title "Singularity Is in the Eye of the Beholder" [1][28].

## Recent work

*Data Science in Context: Foundations, Challenges, Opportunities*, written with Alfred Spector, Chris Wiggins, and Jeannette M. Wing, was published by Cambridge University Press in 2022; its subtitle sets out the three-part treatment of [data science](https://aiwiki.ai/wiki/data_science) the book gives [21].

Since taking the Stanford HAI fellowship, Norvig has concentrated on AI education. In the June 2025 interview he defined the field in terms of decision-making under uncertainty ("AI deals with uncertainty. You're trying to make the best decision based on incomplete, noisy, or ambiguous data") and argued that tools like code assistants only accelerate people who already have judgment: "The goal is not to produce code. It's to produce understanding" [16]. He noted that when he entered the field AI papers came out slowly, while a dozen papers a week now claim a breakthrough, which is part of why he thinks AI education has to be continuous rather than something a degree completes [16].

## Honors and recognition

| Year | Honor |
| --- | --- |
| 1999 | NASA Software of the Year, awarded to Remote Agent, built by the division he headed [2] |
| 2001 | Elected AAAI Fellow [3]; NASA Exceptional Achievement Award [2] |
| 2006 | Elected ACM Fellow [3]; UC Berkeley distinguished alumni award [2] |
| 2013 | Elected to the American Academy of Arts and Sciences, class of Mathematical and Physical Sciences [25] |
| Undated | Fellow of the California Academy of Sciences [2] |

## See also

- [Stuart Russell](https://aiwiki.ai/wiki/stuart_russell)
- [Sebastian Thrun](https://aiwiki.ai/wiki/sebastian_thrun)
- [Google Research](https://aiwiki.ai/wiki/google_research)
- [Artificial general intelligence](https://aiwiki.ai/wiki/artificial_general_intelligence)
- [Machine learning](https://aiwiki.ai/wiki/machine_learning)
- [Judea Pearl](https://aiwiki.ai/wiki/judea_pearl)

## References

1. Peter Norvig, personal homepage. https://norvig.com/
2. Peter Norvig, "Peter Norvig" (biography page). https://norvig.com/bio.html
3. Wikipedia, "Peter Norvig." https://en.wikipedia.org/wiki/Peter_Norvig
4. Alon Halevy, Peter Norvig, and Fernando Pereira, "The Unreasonable Effectiveness of Data," IEEE Intelligent Systems, March/April 2009, pp. 8-12. https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/35179.pdf
5. Peter Norvig, "On Chomsky and the Two Cultures of Statistical Learning," 2011. https://norvig.com/chomsky.html
6. Stuart Russell and Peter Norvig, "Artificial Intelligence: A Modern Approach" official site. https://aima.cs.berkeley.edu/
7. "Artificial Intelligence: A Modern Approach," 4th edition table of contents. https://aima.cs.berkeley.edu/contents.html
8. "Artificial Intelligence: A Modern Approach," editions and translations. https://aima.cs.berkeley.edu/translations.html
9. Wikipedia, "Artificial Intelligence: A Modern Approach." https://en.wikipedia.org/wiki/Artificial_Intelligence:_A_Modern_Approach
10. Stanford Institute for Human-Centered Artificial Intelligence, "Peter Norvig." https://hai.stanford.edu/people/peter-norvig
11. TED, "Peter Norvig: The 100,000-student classroom," TED2012. https://www.ted.com/talks/peter_norvig_the_100_000_student_classroom
12. Wikipedia, "Udacity." https://en.wikipedia.org/wiki/Udacity
13. Wikipedia, "Massive open online course." https://en.wikipedia.org/wiki/Massive_open_online_course
14. Wikipedia, "Sebastian Thrun." https://en.wikipedia.org/wiki/Sebastian_Thrun
15. Blaise Agüera y Arcas and Peter Norvig, "Artificial General Intelligence Is Already Here," Noema Magazine, October 10, 2023. https://www.noemamag.com/artificial-general-intelligence-is-already-here/
16. Cloudera, "Beyond the Textbook: Peter Norvig on the Future of AI Literacy," June 4, 2025. https://www.cloudera.com/blog/business/beyond-the-textbook-peter-norvig-on-the-future-of-ai-literacy.html
17. Fiddler AI, "AI Safety in Generative AI with Peter Norvig" (podcast). https://www.fiddler.ai/podcasts/ai-safety-in-generative-ai-with-peter-norvig
18. Alfred Spector, Peter Norvig, and Slav Petrov, "Google's Hybrid Approach to Research," Communications of the ACM 55(7), 2012, pp. 34-37. https://research.google/pubs/googles-hybrid-approach-to-research/
19. Peter Norvig, "How to Write a Spelling Corrector," 2007. https://norvig.com/spell-correct.html
20. Peter Norvig, "Teach Yourself Programming in Ten Years," 2001. https://norvig.com/21-days.html
21. Alfred Spector, Peter Norvig, Chris Wiggins, and Jeannette M. Wing, "Data Science in Context: Foundations, Challenges, Opportunities," Cambridge University Press, 2022. https://datascienceincontext.com/
22. Elsevier, "Paradigms of Artificial Intelligence Programming: Case Studies in Common Lisp" product page. https://shop.elsevier.com/books/paradigms-of-artificial-intelligence-programming/norvig/978-1-55860-191-8
23. CSLI Publications, "Verbmobil: A Translation System for Face-to-Face Dialog," 1994. https://web.stanford.edu/group/cslipublications/cslipublications/site/0937073954.shtml
24. Wikipedia, "Deep Space 1." https://en.wikipedia.org/wiki/Deep_Space_1
25. American Academy of Arts and Sciences, "Peter Norvig." https://www.amacad.org/person/peter-norvig
26. GitHub, "norvig/pytudes." https://github.com/norvig/pytudes
27. Open Library, catalogue records for "Paradigms of Artificial Intelligence Programming." https://openlibrary.org/search?q=Paradigms+of+Artificial+Intelligence+Programming+Norvig
28. Weights and Biases, Gradient Dissent, "Peter Norvig: Singularity Is in the Eye of the Beholder." https://wandb.ai/site/resources/podcast/episodes/peter-norvig-singularity-is-in-the-eye-of-the-beholder