Gary Marcus is an American cognitive scientist, entrepreneur and author, and one of the most prominent public critics of the claim that scaling Deep learning and large language models alone will lead to human-level or general intelligence.
Marcus is a professor emeritus of psychology and neural science at New York University, where his early research focused on language acquisition and cognitive development in children, work that shaped his later arguments that human intelligence relies on innate structure and compositional reasoning that pattern-matching neural networks lack. He founded Geometric Intelligence, a machine learning startup, in 2014; the company was acquired by Uber in 2016 to help found Uber's AI research lab. He later founded Robust.AI, focused on applying more structured, rule-informed approaches to robotics and AI.
Advocacy for hybrid AI
Marcus has argued consistently, including in his books "Rebooting AI" (2019, with Ernest Davis) and "Taming Silicon Valley" (2024), that pure deep-learning-based systems such as large language models are prone to characteristic failures, including Hallucination (AI), brittle generalization outside their training distribution, and a lack of grounded common-sense reasoning, and that combining neural networks with explicit symbolic representations and reasoning, an approach often called hybrid or neuro-symbolic AI, is necessary for more reliable systems. His critiques predate the current generative AI wave, tracing back to skepticism voiced during the resurgence of Deep learning in the early 2010s, and he has continued to apply the same framework to systems including GPT-3, GPT-4 and successive ChatGPT releases.
Public role and testimony
Marcus testified before the United States Senate Judiciary Committee in May 2023 alongside Sam Altman and IBM's chief privacy officer, arguing for the creation of a dedicated regulatory body to oversee advanced AI systems, part of a broader push toward stronger AI governance. He has been an active public commentator on AI, frequently engaging in public debate with more optimistic industry figures over the pace of progress toward Artificial general intelligence, the reliability of AI benchmarks, and the risks of overhyping near-term AI capabilities.
Reception
Supporters credit Marcus with correctly anticipating specific failure modes of large language models, including their struggles with reliable arithmetic and factual consistency, well before those became mainstream concerns. Critics counter that his predictions about the limits of scaling have repeatedly been overtaken by subsequent model releases, and that his public style, marked by frequent, high-visibility bets and predictions, has sometimes overshadowed the more measured research-level version of his hybrid AI argument.