Gary Marcus has been making the same argument for years. Sitting across from Brian Greene, he made it again, and the longer you watch, the harder it gets to dismiss.
There is a comfortable story the AI industry has been telling itself for the better part of a decade, and it goes like this: we are on a smooth, almost mechanical road to general intelligence, and the only thing standing between us and the destination is more, more data, more parameters, more compute. The road is paved. The car is fast. We just need to keep driving.
Gary Marcus does not think the car is on the road. He thinks it is in a field.
In a recent conversation with Brian Greene, the physicist who, when not explaining string theory to skeptical audiences, has been doing some of the most useful long-form interviewing in popular science, Marcus laid out his case again. It is a case he has been making, with admirable patience, for years. The version he gave Greene was sharper than usual. The conversation runs over an hour, and it rewards the time.
The objection in one sentence
Marcus's argument, distilled, is that large language models are statistical approximators rather than reasoning engines. They are extraordinarily good at predicting plausible next tokens given patterns observed in their training data. They are not, in any architectural sense, doing the thing humans do when we reason about a problem we have never seen before.
This sounds like a small distinction. It is not. It is the difference between a system that has learned the surface of language and one that has learned the world the language is about. The former can produce fluent prose about object permanence. The latter actually believes the cup is still there when you put a napkin over it.
Fluent output is not evidence of a fluent mind. We keep confusing the two because we have never before lived in a world where anything but a mind could produce fluent output.
Why people keep being surprised
One of the more interesting threads in the Greene conversation is Marcus's diagnosis of us, the users, rather than the systems. Why are intelligent people so easily convinced that LLMs understand them?
His answer is a kind of statistical malpractice that we all commit. A model writes a working Python function. A model summarises a passage cleanly. A model passes a logic puzzle. The user, observing these successes, generalises to a global claim about the system's competence. This is, of course, a small-sample error. The model may have failed catastrophically on neighbouring tasks the user never thought to test.
The other half of the diagnosis is older and harder to fight: anthropomorphism. We evolved in a world where only other minds produced language. Our intuitions for "who is on the other side of this sentence" were trained on a million years of data in which the answer was always another person. When a machine produces grammatical, contextually appropriate prose, those intuitions misfire in a way that no amount of conscious reasoning entirely corrects.
The proposed alternative
Marcus is sometimes characterised as an AI pessimist, which gets him exactly wrong. He is not opposed to artificial intelligence. He is opposed to a specific bet about how to build it. The alternative he proposes, and has been proposing, with increasing vindication, for many years, is neurosymbolic AI: a hybrid architecture that fuses the pattern-recognition strengths of neural networks with the formal logic and explicit rule-handling of symbolic systems.
The wager is straightforward. Neural networks are good at perception and bad at reasoning. Symbolic systems are good at reasoning and bad at perception. The failure modes of one are precisely what the other handles well. A correctly designed hybrid would not have to choose.
This is not a new idea. Symbolic AI is older than deep learning, and neurosymbolic approaches have been around for decades. The field largely abandoned them in the 2010s when deep learning's empirical wins made the symbolic side look unfashionable. Marcus's argument, increasingly, is that the abandonment was premature, and that the field is now slowly, expensively, re-discovering why hybrid systems matter.
What infants know
The part of the conversation I found most striking is the one that draws on Marcus's original research home: developmental psychology. The argument has a beautifully simple shape.
Decades of careful research on infant cognition suggest that human babies are not born as blank slates. They appear to come pre-equipped with rudimentary models of objects (they persist when occluded), agents (they act with goals), number, and space. This bundle of innate apparatus is sometimes called core cognition.
If even biological intelligence requires built-in scaffolding to bootstrap learning, the expectation that AI systems will learn everything, including the basic furniture of reality, from raw data alone may be a profound architectural mistake. Marcus's proposal is that AI should be built with explicit world models: internal representations of how objects, causes, and effects work. Not learned from scratch. Built in. The way evolution built them into us.
A baby knows objects don't vanish when hidden. We are still building AI systems that, in a meaningful sense, do not.
The risks that don't require AGI
The final third of the conversation widens out, and this is where Marcus departs from the standard AI doom script. He is, of course, worried about long-term risks. But he is more focused on the dangers already in motion, risks that do not require AGI to materialise, and are not in any meaningful sense speculative.
Industrial-scale misinformation. The erosion of shared epistemic ground. The use of generative systems by actors with no incentive to care about truth. These are present-tense problems. They do not require the machines to wake up. They only require the machines to be cheap enough, fluent enough, and indifferent enough.
This strikes me as the right calibration. The most likely near-term harms of AI are not science fiction. They are familiar harms; fraud, propaganda, social manipulation, running at industrial speed, conducted by systems that face no consequences and feel no shame.
Where the conversation lands
Despite the sharp critique, Marcus ends in qualified optimism. Human-level intelligence in machines is not impossible. It is simply not on the road we are currently driving. If the field embraces some architectural humility, neurosymbolic foundations, explicit world models, structured priors borrowed from the way biological intelligence actually works, the destination remains reachable.
This is, I think, the right place to leave the question. The interesting argument is not whether machines can think. The interesting argument is whether the things we are currently building are on the right path to get there. Marcus's claim is that they are not. He may be wrong. He has been wrong before. But the people he is arguing against have also been wrong, and the cost of taking the comfortable story for granted is the cost of building, at vast expense, an enormous tower that turns out to be in the wrong place.
Worth the hour. Watch the conversation:
