Speaker

Gary Marcus

Gary Marcus

Gary Marcus is a cognitive scientist, AI researcher, and author — emeritus professor of psychology and neuroscience at NYU — best known as the most prominent sceptic of large language models and of the claim that the current paradigm leads to artificial general intelligence.

Marcus trained in cognitive science, the study of how minds represent and reason about the world, and that discipline shapes his critique: he treats intelligence as multi-dimensional rather than a single quantity that scale increases. He founded the machine-learning company Geometric Intelligence, acquired by Uber in 2016, and has authored six books, most recently Taming Silicon Valley (2024), which warned of the capture of policy by a handful of tech firms. His May 2023 US Senate testimony, seated beside Sam Altman, brought his warnings about AI risk to a wide audience. He has anticipated many of the current limitations — unreliability, hallucination, the absence of stable world models — years in advance, dating his published warnings about LLMs to 2019.

Core positions

Marcus’s technical objection is that pure large language models are ‘next-token predictors’: they approximate human output by predicting what word comes next, which captures part of cognition but not the understanding, rule-following, and stable world models that intelligence requires. Pushed outside their training distribution they fail in revealing ways, and Hallucination is not a bug to be patched away but a consequence of the architecture — a direct challenge to the maximalist reading of Scaling Laws.

His economic thesis follows from the technical one. Because the leading labs build on near-identical architectures, there is no moat; commoditisation drives a price war and chronic unprofitability. He has called OpenAI ‘the WeWork of AI’ since 2023 and rates Anthropic the sounder bet, while holding that whether any lab justifies its valuation is genuinely undecided. He distinguishes his stance from the wholesale-doom view (that no lab survives), placing himself a step more optimistic.

On policy, Marcus argues for mandatory, FDA-style pre-release screening of models that weighs benefits against documented harms — sycophancy, delusions, unreliability — and against government equity stakes in unprofitable labs, which he reads as backdoor bailouts. He frames the strategic error as an ‘explore versus exploit’ failure: committing the whole economy to one architecture rather than funding alternative approaches to machine intelligence. See Over-Attribution of Intelligence for his organising idea.

In the wiki