Gary Marcus

Gary Marcus appears in A case for AI models that understand, not just predict, the way the world works as the cognitive scientist and NYU professor emeritus explaining why World Models matter for robust AI. In the Marketplace Tech episode, he argues that current large language models are dominated by statistical sequence prediction rather than explicit representations of entities, facts, states, and causes.

The source uses Marcus to name the LLM World Model Gap: if a model does not maintain something like a scene graph or fact-grounded representation of the world, it can sound fluent while hallucinating or failing at simple factual and causal reasoning. His claim is not that correlation is useless, but that stronger AI will need world models as first-class components and correlation as a supporting mechanism.

Marcus also connects the argument to Embodied AI. Robots and other physical agents need to reason about surfaces, strengths, spatial connections, objects, people, and possible actions. That makes world models a practical requirement for Physical AGI, not only an abstract debate about language models.

Connections

• Marketplace Tech and Megan McCarty-Corino - show and host context.
• NYU - institutional context named in the episode credits.
• World Models, LLM World Model Gap, and Causal World Models - technical concepts his critique sharpens.
• Frontier Model Scaling - debate he enters by arguing that pure data-and-compute scaling is producing smaller gains than earlier jumps.
• Embodied AI, Physical AGI, and Video Models - domains where his entity/state/causality critique becomes concrete.