AI's most persistent skeptic
Gary Marcus
Profile
Gary Marcus is the AI community’s most persistent skeptic — and he has been right often enough that the field can no longer ignore him. A cognitive scientist by training, he is Professor Emeritus of Psychology and Neural Science at NYU, where he spent decades studying how the human mind actually builds representations. That background shapes everything he writes about AI: he treats intelligence as an engineering problem with cognitive-science constraints, not a scaling law waiting to be solved.
In 2014 he co-founded Geometric Intelligence, a machine-learning startup that became the seed of Uber AI Labs after Uber acquired it in 2016. Since leaving Uber in 2017, he has become the loudest independent voice arguing that pure deep learning is the wrong bet for AGI. His position — that scaling LLMs will plateau and that symbolic reasoning, world models, and structured knowledge must be re-integrated into neural systems — is the neurosymbolic thesis he has been defending since the 1990s. For years it made him a punching bag for Sam Altman, Yann LeCun, and Elon Musk. The hallucination problem, the reasoning failures on unfamiliar tasks, and the DeepSeek-era efficiency shift have aged his arguments well.
Marcus is also a policy voice. In May 2023 he testified alongside Sam Altman before the U.S. Senate Judiciary Committee, calling for a dedicated AI oversight body — one of the few times a researcher and a CEO publicly agreed that their industry needed regulation. His 2024 book Taming Silicon Valley turned that testimony into a policy program: data rights, transparency rules, layered oversight, and meaningful tax reform aimed at preventing Big Tech capture of AI.
For developers learning AI, Marcus is essential reading precisely because he is not a booster. His Substack catalogues the failure modes — brittle reasoning, confabulation, benchmarks that don’t generalize — that marketing decks skip over. You don’t have to agree with him on AGI timelines to benefit from reading him; you just have to want to see the failure cases clearly before you build on top of them.
Books
Taming Silicon Valley: How We Can Ensure That AI Works for Us 2024 policy manifesto on how Big Tech captured AI and what concrete regulation could look like. Rebooting AI: Building Artificial Intelligence We Can Trust Co-authored with Ernest Davis — the canonical statement of the neurosymbolic critique of deep learning. Kluge: The Haphazard Evolution of the Human Mind Cognitive science for engineers: why the mind is a messy hack, and why AI built on clean abstractions keeps missing the point. Guitar Zero: The Science of Becoming Musical at Any Age Marcus teaches himself guitar at 40 and uses it as a lens on learning, talent, and neural plasticity. The Algebraic Mind: Integrating Connectionism and Cognitive Science The 2001 academic work where Marcus first laid out the case for symbolic structure in neural systems — the seed of everything he argues today.Key Articles & Papers
Deep Learning Is Hitting a Wall Deep Learning: A Critical Appraisal The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence The Future Is Neuro-Symbolic Marcus on AI (Substack)Videos
Controversies
Marcus’s public feuds are part of his public record. He has traded barbs with Yann LeCun for over a decade over whether deep learning alone can yield general intelligence, and with Sam Altman over scaling claims and safety. Critics argue he moves goalposts when LLMs clear a benchmark he previously cited; supporters argue the benchmarks themselves are the problem. On the Musk-vs-OpenAI lawsuit, he has publicly said “Elon has a point” about OpenAI’s departure from its non-profit mission — a position that made both OpenAI defenders and Musk skeptics unhappy. None of this is scandal, just a researcher who picks fights in public and keeps receipts.
Spotify Podcasts
