Disappointment and a sense of deflation, but no longer denial. Following the release of OpenAI’s GPT-5, the internet was soon awash with tweets and posts from industry insiders reluctantly acknowledging the work of Silicon Valley gadfly Gary Marcus. Since the late 2010s, the cognitive scientist has warned of the limits of large language models (LLMs)—much to the chagrin of deep learning enthusiasts and figures like OpenAI CEO Sam Altman, who have publicly championed a more heady narrative.

Marcus’s critique hinges on what he sees as the inherent fragility of deep learning: a data- and energy-hungry, brute-force approach to “understanding” and generating natural language that has proven dazzling yet fundamentally brittle. Those scare quotes are warranted. LLMs do not understand anything—not in the way we ordinarily mean the term. Instead, these vast symbol-manipulating machines use enormous computational resources to predict the most statistically likely next word or token, based on patterns extracted from the collective human corpus. The results are so impressive that talk of general artificial intelligence (AGI) and even conscious machines has re-entered mainstream discourse. But speculation like that rests on the sandy foundations of anthropomorphic projection and philosophical naivety, confusing surface fluency with depth and mimicry with mind.

Marcus’s story resembles that of an earlier figure in the history of AI realists, whose work exposing the inherent limitations of the technology also made him something of a pariah. In the early 1980s, as optimism surged around so-called expert systems—symbolic logic engines, also known as good old fashioned AI (GOFAI), designed to imitate human reasoning—philosopher Hubert Dreyfus declared himself unconvinced. He had already spent over a decade challenging the foundational assumptions of AI research. His 1965 RAND report, followed by his 1972 book What Computers Can’t Do, argued that genuine intelligence is embodied, situated, and context-dependent, and therefore cannot be captured by rule-based systems or computational representations alone.

Drawing on Heidegger and Merleau-Ponty, Dreyfus contended that expertise and meaning arise not from rule-following but from embodied know-how and being-in-the-world—dimensions inaccessible to representational systems. These claims were met with hostility by many in the field, particularly at places like MIT, where symbolic AI was ascendant. Where his critics saw thinking as a problem of abstract symbol manipulation, Dreyfus insisted that such manipulation could never approximate the pre-reflective, intuitive grasp of meaning that characterises human being.

Just as Marcus objects to the flawed assumptions underlying modern LLMs, Dreyfus warned that no machine, no matter how powerful, could achieve genuine insight or expertise like a human so long as it remained disembodied, disembedded, and blind to the meaningful whole within which human cognition always operates. Despite these limitations, Marcus sees hope in revisiting GOFAI. Its fundamentally different approach, he argues, may offer much-needed resilience and improved reasoning capabilities where deep learning has surpassed the point of diminishing returns. He may well be correct. Nevertheless, there are reasons to suspect that Dreyfus’s half-century-old assessment remains as relevant as ever, particularly when it comes to the loftier aspirations of contemporary AI devotees.