Gary Marcus: Deep Learning Critic Who Backs AI Rules

Gary Marcus is a cognitive scientist and professor emeritus at New York University (NYU). He has argued for years that deep learning alone will not produce trustworthy artificial intelligence (AI). He has also criticized AI hype, called for AI regulation, and testified before the U.S. Senate. His dated predictions are a test of that argument, and the public record is still incomplete.

Who Gary Marcus Is and What He Argues

Gary Marcus is a cognitive scientist and professor emeritus of psychology and neural science at NYU. His core claim is that deep learning alone is insufficient, and that AI needs a hybrid approach. He says that approach combines neural networks with symbol manipulation.

He criticizes AI hype. He has called for AI regulation, greater public AI literacy, and "well-funded public thinktanks." For the wider debate over hype, see our explainer on whether AI is overhyped.

His Academic Background

Marcus was born in 1970 in Baltimore, Maryland. He earned a bachelor of science (BS) at Hampshire College, majoring in cognitive science. He holds a master's (MS) and a PhD from the Massachusetts Institute of Technology (MIT).

His 1993 thesis was titled "On rules and exceptions: an investigation of inflectional morphology." His doctoral advisor was the psychologist Steven Pinker. He also attended the Center for Talented Youth.

His Current Affiliations

The official record of his Senate testimony lists him as "Professor Emeritus, New York University, Vancouver, British Columbia, Canada." That record is dated May 16, 2023. Wikipedia doesn't list another current role, and it doesn't give a start date for the emeritus title.

He writes a paid-tier newsletter on Substack called "Marcus on AI." Its about page doesn't list titles, affiliations, books or a subscriber count. Its only description of him is a blurb from novelist Kim Stanley Robinson, who called him "one of our few indispensable public intellectuals."

His Companies and Uber

Marcus founded Geometric Intelligence, a machine-learning startup. Uber acquired it in December 2016, and Marcus became director of Uber's AI efforts. He left that role in March 2017.

Wikipedia gives two founding dates for Geometric Intelligence, 2014 in the lead and 2015 in the career section. The sources disagree on that date.

In 2019 he launched the robotics company Robust.AI with Rodney Brooks, an iRobot co-founder. Wikipedia cites a Forbes article dated October 28, 2020, reporting a $15 million raise. The sources checked do not say whether Robust.AI still operates, or whether Marcus still has a role or stake in it.

His Core Argument

In his March 10, 2022 essay "Deep Learning Is Hitting a Wall," published in Nautilus, Marcus wrote that "deep learning was only a tiny part of what we need to build if we're ever going to get trustworthy AI." The essay makes several specific claims:

• Scaling laws "aren't universal laws like gravity but rather mere observations that might not hold forever."
• "To think that we can simply abandon symbol-manipulation is to suspend disbelief."
• Large language models are "approximations to ... language use rather than language understanding."
• GPT-3 "can neither integrate information from basic web searches nor reason about the most basic everyday phenomena."
• Deep-learning systems are "particularly problematic" on outliers that differ from their training data.

The same essay points to AlphaGo and AlphaFold2 as early hybrid, or neuro-symbolic, successes. Marcus says he is "finally" optimistic about a neurosymbolic path forward. The essay's forecast is a turn toward hybrid systems. It does not forecast a halt in progress.

His Track Record on Predictions

Marcus has made several dated predictions and wagers. The sources checked do not contain an independent scorecard of them, so the record is only partly settled. Each claim below has a date and a recorded outcome, or a note that none is recorded.

On September 3, 2025, Marcus published a New York Times op-ed titled "The Fever Dream of Imminent Superintelligence Is Finally Breaking." Wikipedia says that after GPT-5's release in 2025, he wrote that adding more data "helps them improve only to a degree."

The Musk Wager

On April 9, 2024, Marcus published an open letter to Elon Musk. Musk had said, "My guess is that we'll have AI that is smarter than any one human probably around the end of next year." Investor Damion Hankejh, CEO of ingk.com, offered to raise the bet to $10 million.

The post ends by asking Musk, "Are you in?" The exact written terms were posted as an image on X and are not readable in the Substack post. Only the $10 million figure, the end-of-2025 deadline and the claim itself can be stated.

The Brundage Bet

On December 30, 2024, Marcus and Miles Brundage published a formal bet about AI at the end of 2027, at 10 to 1 odds. If AI systems can do 8 or more of 10 listed tasks by then, Marcus donates $2,000 to a charity of Brundage's choice. If they can do fewer than 8, Brundage donates $20,000 to a charity of Marcus's choice.

Both pledged to donate regardless of who wins. Marcus's own stake is a charitable donation, not a personal payout. A panel chosen by mutual agreement judges each task, with "broad discretion," on whether it is "technically feasible by the end of 2027 with modest costs on a per-task basis."

The bet defines modest cost as about $1 million of compute plus 10 person-weeks. The tasks include summarizing an unseen mainstream film without inventing events. They also include writing legal briefs without making up cases, and producing bug-free code over 10,000 lines from a spec written in plain language. Bespoke single-task systems do not count, and physical-action tasks were excluded.

The bet makes Marcus's view testable. A named expert took the other side, and a panel will judge the outcome. The timing debate behind bets like this one is covered in our guide to how close we are to artificial general intelligence (AGI).

What the Record Does Not Settle

Two wagers are still open or unresolved. The Brundage bet does not resolve until the end of 2027. The Musk wager has no recorded outcome in the sources checked.

No source checked gives a scored tally of his predictions. A claim that he is right most of the time cannot be checked against these sources.

His Books and Senate Testimony

Marcus has written or edited several books, and they span cognitive science and AI. Wikipedia gives the year for each one below, and the sources disagree on the Norton Reader date.

• "The Algebraic Mind: Integrating Connectionism and Cognitive Science" (MIT Press, 2001)
• "The Birth of the Mind" (Basic Books, 2004)
• "The Norton Psychology Reader," as editor (dated 2005 in the text and 2006 in the bibliography, so the sources disagree)
• "Kluge: The Haphazard Construction of the Human Mind" (Houghton Mifflin, 2008)
• "Guitar Zero: The New Musician and the Science of Learning" (Penguin Press, 2012), described as a New York Times bestseller
• "The Future of the Brain" (Princeton University Press, 2014), co-edited with Jeremy Freeman
• "Rebooting AI: Building Artificial Intelligence We Can Trust" (Pantheon/Random House, 2019), with Ernest Davis
• "Taming Silicon Valley: How We Can Ensure That AI Works for Us" (MIT Press, 2024)

His magazine writing includes "Am I Human?" in Scientific American in March 2017, on AI's difficulty with reliable disambiguation. It also includes "Artificial Confidence" in Scientific American in October 2022.

The Senate Hearing

On May 16, 2023, Marcus testified before the U.S. Senate Committee on the Judiciary, Subcommittee on Privacy, Technology, and the Law. The hearing was titled "Oversight of A.I.: Rules for Artificial Intelligence." It was held in the 118th Congress and is printed as S. Hrg. 118-037.

The other two witnesses were OpenAI chief executive Samuel Altman and IBM chief privacy and trust officer Christina Montgomery. The official hearing record is on govinfo.gov. The sources checked do not give what Marcus said at the hearing, so his testimony is not described here.

Is Gary Marcus Credible?

Gary Marcus is a credentialed scientist whose claims are specific enough to test. The sources checked do not score his predictions.

The Case For Him

His training is documented. He holds MIT degrees, his doctoral advisor was Steven Pinker, and he is professor emeritus at NYU. His books have appeared with MIT Press, Basic Books and Princeton University Press.

He puts his claims on the record with dates, odds and named counterparties. The Brundage bet names a co-signer and a judging panel. Critics respond to him by name, including Dietterich and LeCun.

He has testified before the Senate and signed a 2023 open letter calling for a pause in training larger AI systems. Those are public positions that others can weigh.

The Case Against Him

Critics disputed his January 2018 paper, and LeCun suggested he confused "deep learning" with "supervised learning." His March 2022 headline, "deep learning is hitting a wall," came one year before GPT-4. Critics note that the essay argues for hybrid systems, and that it does not argue progress will halt. Readers can check the claims in the essay itself rather than its headline.

The Musk wager has no recorded outcome, and Musk is not recorded as accepting. No independent scorecard appears in the sources checked.

Beyond his NYU emeritus title, the sources checked list no current role. He has founded two companies and became director of Uber's AI efforts, while also criticizing AI hype. Readers should weigh that record alongside his criticism.

What Would Change This Verdict

The Brundage bet, once judged, would give a scored result on a named claim. A published independent tracker that scored his resolved predictions would change the picture. A posted resolution of the Musk wager would do the same. Until one of those exists, the verdict stays provisional.

What His Critics Say

A named criticism came in 2018. Thomas G. Dietterich, a former AAAI co-chair, called Marcus's January 2018 paper a "disappointing article." He argued that deep learning is "essentially a new style of programming," which he called differentiable programming.

Yann LeCun, then Facebook's director of AI research, backed Dietterich. He said, "Tom is exactly right." He also suggested Marcus might have confused "deep learning" with "supervised learning."

Unnamed Reddit critics argued that Marcus ignored technical detail and recent advances. They cited generative adversarial networks (GANs) and zero- and few-shot learning. One commenter wrote, "If anything, I came out more convinced DL is the future," using the abbreviation for deep learning (DL).

A criticism also came from outside AI. In 2006, Mameli and Papineau reviewed his nativist book in Biology and Philosophy. They called his position "more sophisticated than any version of nativism on the market," but argued his rebuttals "reconfigure the nativist position out of existence."

The strongest point against him on substance concerns the 2022 headline. The essay argued that progress would need symbol manipulation, not that capability gains would stop. For the investment side of the scaling debate, see our AI bubble explainer. For the wider set of AI critics, see our guide to who is fighting AI.

Sources

• Gary Marcus, Wikipedia
• "Deep Learning Is Hitting a Wall," Nautilus, March 10, 2022
• Senate hearing record, May 16, 2023
• Hearing details on govinfo.gov (S. Hrg. 118-037)
• Marcus and Brundage bet on AI at the end of 2027
• Open letter to Elon Musk, April 9, 2024
• Marcus on AI, about page
• Synced Review on the backlash to Marcus's 2018 deep learning critique