Gary Marcus

In 'The Algebraic Mind' (2001), Marcus argued that human cognition relies on symbolic rules that cannot be captured by pure connectionist (neural network) models. This set up a decades-long debate with the deep learning community that the rise of LLMs has made more, not less, urgent.

Co-authored 'Rebooting AI' (2019) arguing that deep learning, despite impressive results, lacks the reliability, common sense, and causal reasoning needed for real-world AI applications. Proposed hybrid architectures combining neural networks with symbolic reasoning as the path forward.

One of the most persistent and documented critics of LLM capabilities (2022–present). Collects and publicizes systematic failure cases: reasoning errors, hallucination, compositional failures, inconsistency. His argument: impressive performance on benchmarks hides fundamental inability to reason reliably about novel situations.

Documents gaps between AI companies' public claims and actual system performance. Argues that AI hype creates dangerous deployment of unreliable systems in high-stakes domains (medicine, law, autonomous vehicles). His blog and substack provide ongoing documentation of AI failures that companies downplay.