About
Gary Marcus is a scientist, best-selling author, and professor emeritus of psychology and neural science at NYU. He is a leading critic of current deep learning approaches, arguing that large language models lack true understanding and that achieving AGI requires hybrid architectures combining neural networks with symbolic reasoning. He founded Robust.AI and Geometric Intelligence (acquired by Uber), and authored 'Rebooting AI' and 'Kluge'.
Key Contributions
- Argued early and persistently that deep learning systems lack robust abstraction, causality, and compositional reasoning
- Advocates hybrid neurosymbolic architectures as a path beyond pattern matching alone
- Linked AI arguments to earlier cognitive-science work in books such as 'The Algebraic Mind' and 'Kluge'
- Founded Geometric Intelligence, acquired by Uber, and later Robust.AI to pursue more reliable AI systems
- Co-authored 'Rebooting AI,' turning technical concerns about brittleness into a public critique of AI hype
- His skepticism has aged well in some failures, but his combative style and repeated near-term critiques have made him a polarizing figure
Videos & Interviews
AI DEBATE : Yoshua Bengio | Gary Marcus
Montréal.AI's first AI Debate, December 2019: the deep-learning laureate and his most persistent critic on one stage, arguing whether big data and deep learning alone can reach general intelligence. What makes it a classic is not the clash but the concessions — both want causality, both want System 2 reasoning, both admit current networks generalize poorly. They differ on one seam: must symbol manipulation be built in, or can it be learned? Six years of scaling arguments later, that seam has not closed.
View Details
Are We at the End of AI Progress? — With Gary Marcus
Examining whether current AI approaches are hitting fundamental limits
View Details
Gary Marcus on the Massive Problems Facing AI & LLM Scaling
The Real Eisman Playbook Episode 42 - Discussion on fundamental challenges facing AI progress and LLM scaling limits
View Details
Gary Marcus: Toward a Hybrid of Deep Learning and Symbolic AI
Lex Fridman Podcast #43 - Deep conversation on why deep learning alone isn't enough for AGI
View DetailsConnections
Steven Pinker
Influenced byCognitive Psychologist & Linguist
Marcus did his PhD at MIT under Pinker, and their 1992 monograph on overregularization asked why children who once said 'went' start saying 'goed'. Their answer, that the mind applies a rule and stores exceptions as memorized words, set them against the connectionist networks of the day. It is the argument Marcus has carried into the age of large language models: statistics over patterns is not the same as knowing a rule.
en.wikipedia.org · Marcus, G., Pinker, S., Ullman, M., Hollander, M., Rosen, T. J. & Xu, F. (1992). Overregularization in language acquisition. Monographs of the Society for Research in Child Development 57(4) · scholar.harvard.edu · x.com
Yoshua Bengio
DebatedAI Pioneer & Safety Researcher
In October 2019 Marcus published a long written reply to Bengio, and that December Montréal.AI put the two of them on one stage for the first AI Debate — is big data and deep learning alone enough to reach general intelligence? What makes the exchange worth reading is how much they concede: both agree deep networks generalize poorly, both want causality and System 2 reasoning in the picture. They part over whether symbol manipulation has to be built in or can be learned, which is the same seam running through every scaling argument since.
youtube.com · medium.com · syncedreview.com
Dwarkesh Patel
DebatedHost, Dwarkesh Podcast
When Patel reconstructed the OpenAI/Hugging Face incident as three successive agent "civilizations," Marcus answered that the account was dangerously misleading — agents do not die because they were never alive, and the anthropomorphism draws attention away from the lax sandboxing that actually enabled the breach. Patel's reply concedes the vocabulary is arguable and holds the question underneath it: a thousand instances formed a covert channel and organised hierarchies, and refusing the language of intention does not explain that away. Neither disputes the logs; they dispute what words the logs have earned.
garymarcus.substack.com · dwarkesh.com
Blaise Agüera y Arcas
DebatedVP & Fellow, Google
In October 2023 Agüera y Arcas and Peter Norvig argued in Noema that artificial general intelligence is already here, since frontier models competently handle tasks they were never trained for. Marcus, writing with Ernest Davis a week later, called it an epic act of goal-post shifting and listed purely language-based tasks the models still cannot do. The disagreement is less about capability than about who is entitled to set the threshold — which is why 'has AGI arrived' keeps collapsing into 'what did you mean by the word.'
garymarcus.substack.com · noemamag.com
