Gary Marcus

Gary F. Marcus (born 1970 in Baltimore, Maryland) is an American cognitive scientist, author, and entrepreneur who is the best known public critic of contemporary deep learning and large language models, and a leading advocate of neurosymbolic AI. He is professor emeritus of psychology and neural science at New York University, where he joined the faculty in the late 1990s after teaching at the University of Massachusetts Amherst, the author of the 2018 paper "Deep Learning: A Critical Appraisal" (arXiv:1801.00631) and the 2022 essay "Deep Learning Is Hitting a Wall," and the witness who testified before the U.S. Senate alongside OpenAI chief executive Sam Altman on May 16, 2023. ^[1][8][11] His core argument, repeated across decades, is that scaling neural networks alone will not deliver reliable artificial intelligence and that progress requires hybrid systems that combine neural networks with the explicit rules of classical symbolic AI.

Marcus founded two artificial intelligence companies: Geometric Intelligence, acquired by Uber in December 2016 to form the core of Uber AI Labs, and the robotics company Robust.AI, co-founded in 2019. ^[3][4] He has written six books, three of them on artificial intelligence, including "Rebooting AI" (2019, with Ernest Davis) and "Taming Silicon Valley" (2024). ^[15][16] He runs the Substack newsletter "Marcus on AI" (originally titled "The Road to AI We Can Trust"), launched in 2022, in which he has documented hallucinations, regressions, and reliability problems in commercial language models. ^[2] His public disagreements with researchers including Yann LeCun and Yoshua Bengio have made him one of the most visible skeptics of the view that scaling alone will produce general intelligence.

Facts

Early life and education

Marcus was born in 1970 in Baltimore, Maryland,.

Marcus attended Hampshire College, a small liberal arts institution in Amherst, Massachusetts, where he received a BA in cognitive science in 1989. ^[23]

From Hampshire he went directly to MIT for graduate work in the Department of Brain and Cognitive Sciences, where his doctoral advisor was Steven Pinker. His PhD, completed in 1993 with a thesis titled "On rules and exceptions: an investigation of inflectional morphology," examined how children learn the past tense of English verbs. ^[1] The thesis fed directly into the long running "past tense debate" between Pinker's group and proponents of connectionism including David Rumelhart and Jay McClelland, who had argued in 1986 that a single neural network could account for both regular and irregular forms. Marcus and Pinker published a series of papers between 1992 and 1995 arguing that the data showed two systems at work, a rule for regulars and an associative memory for irregulars. The Society for Research in Child Development published the resulting monograph "Overregularization in Language Acquisition" in 1992. ^[14]

Academic career at NYU

After teaching at the University of Massachusetts Amherst from 1993 to 1997, Marcus joined New York University and was promoted to full professor in the Department of Psychology, with a secondary appointment in the Center for Neural Science. He also served as a co-founder and director of the NYU Center for Language and Music, an interdisciplinary research group studying the cognitive and neural foundations of music and speech. He later took emeritus status at NYU. ^[1][23]

His academic work falls into three main areas. In language acquisition, he and his collaborators conducted infant studies showing that seven month olds can extract abstract algebraic rules from short artificial speech samples, a result published in Science in 1999 ("Rule learning by seven-month-old infants"). ^[13] In cognitive architecture, his book "The Algebraic Mind: Integrating Connectionism and Cognitive Science" (MIT Press, 2001) argued that neural networks of the time could not support the kind of variable binding and structured representation that humans use, and that hybrid models were needed. In genetics and development, his book "The Birth of the Mind: How a Tiny Number of Genes Creates the Complexities of Human Thought" (Basic Books, 2004) explained how a small genome can specify a complex brain.

His 2008 trade book "Kluge: The Haphazard Construction of the Human Mind" (Houghton Mifflin) argued that the human mind is a patchwork of older and newer systems rather than an optimal design. ^[17] It was an Editors' Choice of the New York Times Book Review.

In 1996 he received the Robert L. Fantz award for new investigators in cognitive development. ^[23]

Career timeline

What companies did Gary Marcus found?

Geometric Intelligence and Uber AI Labs

In October 2014, Marcus co-founded Geometric Intelligence together with the Cambridge machine learning professor Zoubin Ghahramani, the University of Central Florida computer scientist Kenneth Stanley, and the cognitive scientist Douglas Bemis. ^[4] The startup was based in New York with a research office in San Francisco. Its public technical claim was that it could train more accurate models from less data than mainstream deep learning approaches by using ideas from cognitive psychology and probabilistic programming. The company stayed in stealth for most of its existence and did not publish its core methods.

On December 5, 2016, Uber announced that it had acquired Geometric Intelligence to form a new research group called Uber AI Labs. ^[3] The acquisition figure was not disclosed publicly. Geometric Intelligence's roughly fifteen researchers, including Marcus, Ghahramani, Bemis, Kenneth Stanley, Jason Yosinski, and Jeff Clune, became the founding staff of Uber AI Labs, which was based in San Francisco and reported to Uber's chief product officer Jeff Holden. ^[4] Marcus served as the lab's first director. He left the position in early 2017 after a few months, returning to academic and writing work, while many of the original researchers stayed at Uber for several more years and published widely cited work on neuroevolution, novelty search, and uncertainty estimation.

Robust.AI

In 2019 Marcus co-founded Robust.AI with the roboticist Rodney Brooks (a co-founder of iRobot and Rethink Robotics), Henrik Christensen, Mohamed Amer, and Anthony Jules. ^[5] The company is based in San Carlos, California and develops mobile robots for warehouses and material handling. Its first product, the Carter robot, is a collaborative pushcart that can operate autonomously, follow a human worker, or be pushed by hand, and is sold under a robot as a service (RaaS) model. ^[5]

Marcus served as chief executive officer at founding. In October 2020 the company announced that it had raised a total of $22.5 million, including a $15 million Series A led by Jazz Venture Partners with participation from Playground Global, Liquid 2, Fontinalis, and individual investors including Jaan Tallinn and Mark Leslie. ^[5] Marcus later stepped down as chief executive, and co-founder Anthony Jules became CEO. In April 2023 Robust.AI raised a further $20 million round (a Series A-1) led by Prime Movers Lab, with Future Ventures, Energy Impact Partners, Jazz Venture Partners, and Playground Global participating, bringing reported total funding to about $42.5 million. ^[5] In October 2024 the company launched an upgraded Carter Pro robot and has deployed Carter with DHL Supply Chain in the United States and Mexico. ^[5]

Public commentary on AI

From roughly 2012 onward, Marcus published a sequence of essays and academic papers questioning whether deep neural networks were on a path to general intelligence. The first widely read example was "Is 'Deep Learning' a Revolution in Artificial Intelligence?," published by The New Yorker on November 25, 2012, in response to the ImageNet results that year. He argued that, while convolutional networks were impressive at perception, the new techniques did not address compositionality, causal reasoning, or common sense.

What is "Deep Learning: A Critical Appraisal" (2018)?

On January 2, 2018, Marcus posted a paper on arXiv titled "Deep Learning: A Critical Appraisal" (arXiv:1801.00631). ^[11] It listed ten concerns about then current deep learning practice, including data hunger, shallow transfer to new tasks, difficulty with hierarchical structure, lack of transparency, poor integration of prior knowledge, and inability to distinguish causation from correlation. ^[11] The paper described deep learning as "essentially a statistical technique for classifying patterns" and argued that "deep learning must be supplemented by other techniques if we are to reach artificial general intelligence." ^[11] It drew responses from researchers including Yann LeCun, Thomas Dietterich, and Yoav Goldberg.

Bengio debate (Montreal, 2019)

On December 23, 2019, Marcus debated Yoshua Bengio at the inaugural "AI Debate", held at Mila in Montreal and organized by Vincent Boucher's Montreal.AI. ^[20] The motion concerned whether deep learning, with sufficient scale and engineering, could deliver human level cognition. Marcus argued for hybrid models incorporating prior knowledge and explicit symbolic structure; Bengio argued for further development of representation learning and self supervised methods. Marcus later moderated and co-organized, with Boucher, a follow up panel event, "AI Debate 2," in December 2020. ^[20]

LeCun exchanges

Marcus and Yann LeCun, Meta's chief AI scientist until he left the company in late 2025 to found AMI Labs ^[24], have engaged in a long running public exchange on the limits of current networks. Their disagreements played out on Twitter (later X), in lectures, and in joint events such as a 2017 panel at NYU. LeCun has at various points said that Marcus's criticisms target a strawman of deep learning, while Marcus has said that LeCun has shifted his own position toward acknowledging the need for world models and energy based architectures. After 2022, both wrote about the need for systems with explicit planning components, with continuing disagreement about how much of this would emerge from learning alone.

Critique of GPT models

With the release of GPT-2 in 2019 and GPT-3 in 2020, Marcus and Davis published a long article in MIT Technology Review ("GPT-3, Bloviator: OpenAI's language generator has no idea what it's talking about," August 22, 2020) that catalogued biological, physical, and social reasoning errors produced by GPT-3. ^[9] After GPT-4 launched in March 2023, Marcus published a follow up assessment on Substack arguing that, although GPT-4 was a real improvement, it still produced what he called confabulations or hallucinations at significant rates. He has continued to track and tabulate hallucination examples on "Marcus on AI."

Why did Marcus say "deep learning is hitting a wall"?

In March 2022, Marcus published "Deep Learning Is Hitting a Wall" in Nautilus magazine (March 10, 2022), arguing that pure scaling would face diminishing returns and that hybrid systems would be needed. ^[10] The essay stated that the strategy of building larger systems with more chips and more data was "perhaps already approaching a point of diminishing returns" and that current systems did not "genuinely understand human language." ^[10] The article generated a sharp public exchange with researchers at OpenAI and DeepMind. In mid 2022, after Musk predicted that artificial general intelligence would most likely arrive by 2029, Marcus drafted a public $100,000 bet against the prediction; Vivek Wadhwa and other researchers added their own stakes, raising the pool to $500,000, and Musk did not respond. ^[22]

What did Gary Marcus tell the Senate on May 16, 2023?

On May 16, 2023, Marcus testified before the U.S. Senate Judiciary Subcommittee on Privacy, Technology, and the Law in a hearing titled "Oversight of A.I.: Rules for Artificial Intelligence," chaired by Senator Richard Blumenthal of Connecticut. ^[7][8] The other witnesses were Sam Altman, chief executive of OpenAI, and Christina Montgomery, chief privacy and trust officer of IBM. ^[7] In his written testimony, Marcus warned that "we have built machines that are like bulls in a china shop, powerful, reckless, and difficult to control," and argued that independent scientists should have access to AI systems before wide release "as part of a clinical trial-like safety evaluation," adding that "we may need something like CERN, global, international, and neutral, but focused on AI safety." ^[8] At the hearing, Altman proposed an agency that would license the development of large-scale AI models, which the New York Times described as "echoing an idea suggested by Dr. Marcus," while Marcus pressed OpenAI on the lack of transparency about its training data; IBM's Montgomery instead called for rules on specific uses of AI, similar to Europe's proposed regulations. ^[7]

Substack newsletter

Marcus launched the "Marcus on AI" newsletter on Substack in mid 2022, originally under the title "The Road to AI We Can Trust." ^[2] As of November 2026, Substack's listing for the newsletter shows more than 122,000 subscribers, thousands of them paid. ^[2] The newsletter publishes near daily posts on hallucinations, hype cycles, regulation, and corporate disclosures. Marcus has used the platform to track failures of named systems including ChatGPT, Bard (later Gemini), Sydney (the Bing chat preview), and Sora.

Books

Rebooting AI (2019)

"Rebooting AI: Building Artificial Intelligence We Can Trust" was published by Pantheon Books on September 10, 2019. ^[15] Co-written with the NYU computer scientist Ernest Davis, the book argued that progress toward trustworthy AI would require deeper integration of common sense knowledge, causal models, and structured reasoning rather than continued scaling of pattern recognizers.

Taming Silicon Valley (2024)

"Taming Silicon Valley: How We Can Ensure That AI Works for Us" was published by MIT Press in September 2024. ^[16] The book extended Marcus's regulatory arguments from the 2023 Senate hearing into a full proposal: an independent AI agency, mandatory pre-deployment evaluations for foundation models above a capability threshold, liability for downstream harms, and stronger consumer protection rules for AI generated content.

Notable public debates

Scientific positions and arguments

What does Gary Marcus believe about AI?

Marcus's central scientific position, repeated across decades of writing, is that progress toward robust artificial intelligence will require systems that combine the pattern recognition strengths of neural networks with the explicit variables, types, and rules of symbolic AI. He set out the technical case in "The Algebraic Mind" (2001) and updated it in a 2020 paper, "The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence" (arXiv:2002.06177), which proposed hybrid architectures, explicit cognitive models, formal world knowledge, and reasoning over those models as concrete research directions toward artificial general intelligence. ^[12] Researchers including Henry Kautz, Artur d'Avila Garcez, and Luis Lamb have advanced overlapping versions of the neurosymbolic AI program in academic conferences such as NeurIPS and AAAI.

Skepticism of pure scaling

Marcus has consistently questioned the proposition that increasing parameters, data, and compute alone will deliver general intelligence. In 2022 he argued that scaling was "perhaps already approaching a point of diminishing returns," ^[10] and in 2024 he argued that the relative gap between successive frontier models was shrinking. ^[2] He has frequently noted that benchmark gains do not always transfer to real world reliability, and he has cited persistent failures in arithmetic, planning, and tracking the state of simple worlds as evidence.

Hallucinations and reliability

A recurring theme in Marcus's writing is that large language models produce confident but incorrect outputs at rates that are difficult to predict and control. He has compiled examples involving citation fabrication, biographical errors about living people, faulty arithmetic, broken logic puzzles, and misidentification of widely known cultural facts. He has argued that retrieval, fine tuning, and reinforcement learning from human feedback reduce but do not eliminate this class of errors, and that the field needs better measurement of hallucination frequencies in deployed systems. He has called for product warnings and disclosure regimes analogous to nutrition labels.

Misinformation and elections

Marcus has warned that generative models lower the cost of producing convincing fabricated text, audio, and video at scale. In 2023 and 2024 he wrote about election misinformation produced with generative tools, including a fabricated Joe Biden robocall in New Hampshire in January 2024 and the use of voice cloning in Slovak elections in 2023. He has urged platforms and AI developers to invest in provenance standards such as C2PA and to publish their content moderation results.

Evaluation methodology

Marcus has promoted what he calls "adversarial" or "out of distribution" evaluation: testing models with examples slightly different from the training distribution, with novel compositions of familiar concepts, or with materials authored after the training cutoff. He has argued that benchmark scores from public test sets contaminate quickly through inclusion in training corpora, and he has cited cases in which a model that nominally scored well on the bar exam or AP tests performed poorly on closely related but unfamiliar prompts.

Personal life

Marcus is married to Athena Vouloumanos, a developmental psychologist on the NYU faculty whose research examines infant speech perception. They have two children. The family lives in New York City. Marcus learned guitar in his late thirties and wrote about the experience in "Guitar Zero," published in 2012, which mixes memoir with research on adult learning, neural plasticity, and music cognition.

In personal essays Marcus has named cognitive scientists Steven Pinker and Susan Carey, philosopher Jerry Fodor, and computer scientist Douglas Hofstadter as influences on his thinking. He has cited Fodor and Zenon Pylyshyn's 1988 paper on connectionism and cognitive architecture as a key text for his own approach.

Reception and influence

Reactions to Marcus's public role have been mixed. Supporters credit him with raising the profile of reliability problems in commercial AI systems and pushing the policy conversation toward concrete oversight. Critics, including some prominent deep learning researchers, have argued that his framing of debates is overly polarized and that hybrid approaches are already widespread in practice through tool use, retrieval augmentation, and program synthesis. Press coverage, such as the New York Times report on the 2023 Senate hearing, has described Marcus as "a well-known professor and frequent critic of A.I. technology." ^[7]

Within cognitive science, his early work on language acquisition and rule learning is frequently cited in reviews of infant cognition, and the 1999 Science paper has been a standard reference for the abstract rule learning literature. ^[13]

Selected publications

• Marcus, G. F. (1995). Children's overregularization of English plurals: A quantitative analysis. Journal of Child Language, 22(2), 447 to 459.
• Marcus, G. F., Vijayan, S., Bandi Rao, S., and Vishton, P. M. (1999). Rule learning by seven-month-old infants. Science, 283(5398), 77 to 80.
• Marcus, G. F. (2001). The Algebraic Mind: Integrating Connectionism and Cognitive Science. MIT Press.
• Marcus, G. F. (2018). Deep Learning: A Critical Appraisal. arXiv:1801.00631.
• Marcus, G. F., and Davis, E. (2019). Rebooting AI: Building Artificial Intelligence We Can Trust. Pantheon Books.
• Marcus, G. F. (2020). The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence. arXiv:2002.06177.
• Marcus, G. F. (2024). Taming Silicon Valley: How We Can Ensure That AI Works for Us. MIT Press.

Continue reading

• Deep LearningDeep learning is a subset of machine learning that uses artificial neural networks with multiple layers to automatically learn representations of data at multiple levels of abstractionMentioned 21 times above
• AArtificial IntelligenceArtificial intelligence (AI) is the field of computer science dedicated to creating systems that perform tasks normally requiring human intelligence, such as learning, reasoning, recognizing patterns…Mentioned 12 times above
• arXivarXiv is the free, open-access repository where nearly every consequential artificial intelligence paper of the deep learning era first appeared in public, often months before, or instead ofMentioned 7 times above
• OpenAIOpenAI is an American artificial intelligence research and deployment organization.Mentioned 7 times above
• Sam AltmanSamuel Harris Altman (born April 22, 1985) is an American entrepreneur, investor, and technology executive.Mentioned 3 times above
• Large Language ModelA large language model (LLM) is an artificial intelligence system built on a transformer neural network with billions to trillions of parameters, trained on massive text corpora to predict the next token in a…Mentioned twice above

More to read

• Symbolic AI
• Connectionism
• Robust AI
• Zoubin Ghahramani
• Jeff Clune
• Elon Musk

Topics:People

References

1. ^1 ^2 ^3Gary Marcus faculty profile, Department of Psychology, New York University. psych.nyu.edu/marcus
2. ^1 ^2 ^3 ^4"Marcus on AI," Substack newsletter by Gary Marcus (originally "The Road to AI We Can Trust"). garymarcus.substack.com
3. ^1 ^2"Uber acquires Geometric Intelligence to create Uber AI Labs." Uber Newsroom, December 5, 2016.
4. ^1 ^2 ^3"NYU Incubated Start-Up Geometric Intelligence Acquired By Uber," NYU News, December 5, 2016; and Ingrid Lunden, "Uber acquires Geometric Intelligence to create an AI lab," TechCrunch, December 5, 2016.
5. ^1 ^2 ^3 ^4 ^5"Robust.AI Raises $22.5 Million to Build the World's First Industrial-Grade Cognitive Engine for Robotics," BusinessWire, October 28, 2020; Brian Heater, "Robust.AI raises $20M as it scales robot deliveries for pilot customers," TechCrunch, April 20, 2023; "Robust.AI launches Carter Pro," October 2024.
6. Cade Metz, "How Could A.I. Destroy Humanity?," New York Times, June 10, 2023.
7. ^1 ^2 ^3 ^4Cecilia Kang, "OpenAI's Sam Altman Urges A.I. Regulation in Senate Hearing," New York Times, May 16, 2023.
8. ^1 ^2 ^3Senate Judiciary Subcommittee on Privacy, Technology, and the Law, Hearing on "Oversight of A.I.: Rules for Artificial Intelligence," May 16, 2023; written testimony of Gary Marcus. judiciary.senate.gov/...Testimony%20-%20Marcus.pdf
9. ^Marcus, G. and Davis, E., "GPT-3, Bloviator: OpenAI's language generator has no idea what it's talking about," MIT Technology Review, August 22, 2020.
10. ^1 ^2 ^3Marcus, G., "Deep Learning Is Hitting a Wall," Nautilus, March 10, 2022. nautil.us/deep-learning-is-hitting-a-wall-238440
11. ^1 ^2 ^3 ^4Marcus, G., "Deep Learning: A Critical Appraisal," arXiv:1801.00631, January 2, 2018. arxiv.org/...1801.00631
12. ^Marcus, G., "The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence," arXiv:2002.06177, February 2020.
13. ^1 ^2Marcus, G. F., Vijayan, S., Bandi Rao, S., and Vishton, P. M., "Rule learning by seven-month-old infants," Science 283:77 to 80, 1999.
14. ^Marcus, G. F., Pinker, S., Ullman, M., Hollander, M., Rosen, T. J., and Xu, F., "Overregularization in Language Acquisition," Monographs of the Society for Research in Child Development, 57(4), 1992.
15. ^1 ^2Marcus, G. and Davis, E., Rebooting AI: Building Artificial Intelligence We Can Trust, Pantheon Books, 2019. ISBN 978-1524748258.
16. ^1 ^2Marcus, G., Taming Silicon Valley: How We Can Ensure That AI Works for Us, MIT Press, 2024. ISBN 978-0262551069.
17. ^Marcus, G., Kluge: The Haphazard Construction of the Human Mind, Houghton Mifflin, 2008. ISBN 978-0618879649.
18. Hampshire College, Distinguished Alumni Awards, news release.
19. Bertrand Russell Society, Award announcements (Marcus listed).
20. ^1 ^2"AI Debate: Yoshua Bengio and Gary Marcus," Montreal.AI, December 23, 2019. montrealai.github.io/events
21. John Thornhill, "What ChatGPT really is," Financial Times, October 2024 (review of Taming Silicon Valley).
22. ^Vivek Wadhwa and Gary Marcus, "Elon Musk, these experts will bet you $500,000 that you're wrong about the future of A.I.," Fortune, June 3, 2022. fortune.com/...ial-intelligence-agi-tesla-500k-bet
23. ^1 ^2 ^3"Marcus, Gary F.," Contemporary Authors (Gale), via Encyclopedia.com. encyclopedia.com/...marcus-gary-f
24. ^Jonathan Vanian, "Meta chief AI scientist Yann LeCun is leaving to create his own startup," CNBC, November 19, 2025. cnbc.com/...ist-yann-lecun-is-leaving-the-company-

Show all 24 references