When Will We Start Using Scientific Fundamentals to Question the Gary Marcuses of the World?
A response to Gary Marcus’s critique that LLMs lack world models. The essay argues from physics, neuroscience and information theory that human intelligence relies on saving and reading, error correction and collective processes, and that AI should be judged on the same footing.
This article is more than a year old. Some of its findings may have changed since.
The Authority Trap in AI Discourse
The artificial intelligence (AI) debate suffers from a troubling pathology: arguments from authority masquerading as scientific reasoning. Gary Marcus's recent critique of large language models (LLMs) exemplifies this pattern—superficially compelling claims that dissolve under rigorous scientific scrutiny. To move beyond the comfort of familiar dogma, we must demand the same empirical rigor in evaluating AI that we apply to any other scientific phenomenon—and be willing to question our deepest assumptions about intelligence itself.
1. The Mammalocentric Fallacy: Mistaking Familiar Intelligence for General Intelligence
To understand both the limitations and potential of AI, let us first dissect our own cognitive architecture. Strip away every external augmentation—writing, reading, databases, communication tools, collective processing—and examine the raw intelligence we possess as isolated biological entities. What remains is a form of cognition remarkably similar across brainy mammals: operating on similar spatiotemporal scales, in similar environments, pursuing similar goals.
This similarity gives rise to what I call the "mammalocentric fallacy" of AI—the tendency to assume that our specific cognitive constraints define the boundaries of intelligence itself. This bias, whether explicit or implicit, profoundly shapes the current discourse around AI.
The pursuit of artificial general intelligence bears the unmistakable imprint of this fallacy. Precisely because LLMs lack direct sensorimotor experience, we construct benchmarks like ARC-AGI that favor spatial reasoning—not because such tasks represent more "general" intelligence, but because they reflect mammalian cognitive strengths.
Consider a simple thought experiment: take any ARC-AGI puzzle, split it into fragments, and arrange them randomly. Although one could argue that the underlying task is unchanged, human performance collapses because it can't rely on familiar spatial patterns.
Similarly, virtually no human can reason fluently about four-dimensional motion. Yet we never propose adding such tasks to our benchmarks—not because they are irrelevant to intelligence, but because they are utterly foreign to our own evolutionary experience.
2. The First Leap: Saving and Reading
A key dimension of human intelligence is our capacity for saving and reading. These abilities enable us to behave a bit more like Turing machines, extending the coherence of our thoughts over complex processes. Saving information—in the broad sense, including writing on paper, saving to a database, or asking another person to remember—allows us to freeze fleeting thoughts, free up working memory for subsequent steps, return later without losing the thread of a complex process (provided that the next steps do not rely heavily on earlier ones), and revisit our work with a fresh perspective, enabling better sampling of possible next steps.
Combined with reading, saving allows us to implement strategies that approximate formal error correction, extending the coherence and accuracy of our reasoning. We can write out our reasoning, spot inconsistencies or imprecisions by rereading, and then modify our text to clarify or correct it. This continuous improvement process is a pillar of intelligence beyond brainy mammals and intelligence "à la ARC-AGI".
Crucially, LLMs and alternative neural networks do not possess intrinsic saving and reading capabilities. At present, they are best used as units of compositional processing. It is only when properly integrated within broader computational graphs that they enable systems that implement human-like saving and reading, making approximate error correction possible.
3. The Second Leap: Error Correction
Another key dimension of modern human intelligence is our access to Turing machines, or their practical implementation in the form of computers. There are no intrinsic error correction mechanisms in our brains—or in most physical systems for that matter. Our ability to follow steps without errors when adding two numbers or playing chess is never assured. If we want to implement error correction, we must know how and make a conscious effort.
Turing machines enable strict error correction, making it possible to extend coherence over an arbitrarily long sequence of steps. This is how we send shuttles to space and how we manage the overwhelming processing required by the scale of modern society.
Again, LLMs and alternative neural networks do not have intrinsic access to Turing machines. It is only when properly integrated into architectures that leverage Turing machines that they enable dimensions of intelligence powered by error correction.
4. The Third Leap: Collective Intelligence
When multiple brainy mammals augmented by strategies like saving, reading, and error correction start interacting, a whole universe of collective intelligence emerges.
Again, LLMs and alternative neural networks do not intrinsically capture this dimension. Even though they are exposed to collective intelligence and compress some of it, it is only when integrated in systems that enable such interactions that the richness of collective intelligence can be realized.
5. The Fourth Leap: A Whole Universe of Unfamiliar Intelligences
Beyond our mammalocentric views of intelligence—collective or not—a whole universe of alien intelligences exists, operating in different spaces and with goals we cannot comprehend.
Biology alone can broaden our appreciation of what intelligence can be: a brain, a cell collective, a swarm, a vine growing around a pole. These systems exhibit memory, learning, problem-solving, and adaptation at various levels of sophistication, often without any central world model or even any nervous system. They succeed by exploiting physics, feedback, and local rules that yield global order. I challenge anyone to go to Michael Levin's lab and maintain an anthropocentric or even mammalocentric view of intelligence.
6. What Gary Marcus's "World Model" Critique Gets Wrong
Let us now take the example of Gary Marcus's recent critique of LLMs (see his Substack post on generative AI's world model problem) to illustrate the wider lack of scientific anchoring in the current debate.
Gary Marcus writes: "Generative AI has a crippling problem that nobody wants to talk about: LLMs lack world models, and the consequences are vast." He defines world models as "persistent, stable, updatable (and ideally up-to-date) internal representations of some set of entities within some slice of the world," giving examples like "a database to track a set of individuals over time, including, for example, their addresses, telephone numbers, social security numbers, etc."
He further claims: "An LLM doesn't maintain structured symbolic systems like databases; it has no direct database of cities and populations." "Every day of the week Wikipedia will correctly tell you that Shearer was born in Los Angeles. But ChatGPT can't leverage that kind of information reliably, and doesn't have its own explicit, stable, internal model. One day it may tell you Shearer is British, another day that he is American."
There are two fundamental points to address here:
First, if AI is meant in the sense of human-like intelligence—which is what most people have in mind—then Gary Marcus's critique warrants deeper examination: AI does not require his notion of a world model, since humans do not have that either. Our brains intrinsically lack reliable (error-corrected) representations—symbolic or not. We augment ourselves with error-correcting tools for that.
Second, even if AI had some sort of intrinsic reliable database about all facts and concepts of the human world, it would be fundamentally impossible for it to retrieve such information efficiently. Searching information requires computational resources—time, energy—that scale with the size of the body of knowledge. If AI is meant to answer on a human timescale (e.g., in a few seconds), then the more facts and concepts we put in, the less reliable its answers must be.
Intelligence lies in knowing how to get and use information, not in hoarding it all. As humans do, AI must be able to fill gaps by interacting with the world—asking, observing, searching records, and so on.
Though LLMs are imperfect units of compositional processing, they make this possible: integrated properly, they can compose inputs and assemble on the fly a tiny "world model" for a task. This requires architectures designed to refine and memorize such task-related world models—like the rules of chess and basic past experience of winning moves.
Yes, systems relying on naked LLMs—without saving and reading, error correction, multi-system interactions, and so on—are limited. They are prone to hallucinations, they cannot solve complex problems. These weaknesses are real. However, they are largely amplified by mainstream platforms that operate on (mostly) naked LLMs.
Based on the above arguments, it is highly unlikely that a Gary Marcus–style “world model” is required for human-like AI—or, indeed, for any other form of practical intelligence. This conclusion holds even if we interpret his notion of a world model—explicit, stable, and reliable—as one that merely relies on approximate error correction mechanisms, rather than perfect symbolic structures.
7. The Limits of Gary Marcus's World Models in AI and Biology
Gary Marcus writes, "People who have built neural networks such as LLMs have mostly tried to do without explicit models, hoping that intelligence would 'emerge' from massive statistical analyses of big data." "One hopes that proper world models will simply 'emerge'—but they don't."
As far as I know, there is absolutely zero scientific evidence that nature has endowed us with explicit models. But since I cannot keep track of everything these days, I will simply note that there is overwhelming evidence against the existence of explicit models.
Instead, there are dozens of studies suggesting that our behavior emerges from our exposure to the world, rather than from ad hoc "explicit models." And when neural networks or other sufficiently capable learning algorithms are exposed to similar stimuli, the representations that they build are often compatible with those in the brain.
Gary Marcus writes, "A friend asked ChatGPT to draw an upside-dog and got back an upside puppy with 5 legs. If an LLM had a proper world model, it wouldn't draw a puppy with five legs." When we draw things, we focus on very few salient features. Is a world model more proper when it draws what Gary Marcus draws, or when it gets right a much richer set of features—some of which we do not even process consciously in our brain?
He also observes, "The fact that they can get by as far as they can without explicit world models is actually astonishing." What's more remarkable is that, despite the incompatibility of explicit world models with both human cognition and practical AI, some still consider them essential.
The most generous interpretation I can offer is that critics are judging generative AI solely by the capabilities of naked LLMs or ChatGPT-like architectures. Disregarding the fact that LLMs are powerful compositional processing units that can be embedded in computational graphs to enable grounding in facts, error correction, rule enforcement (as in chess), and much more.
Also, can we once and for all stop saying that LLMs have a statistical nature and we do not? Physics tells us that any system that interacts with the rest of the world has, from the viewpoint of the system alone (i.e., without full information about the rest of the world), a statistical nature. This is a fundamental empirical fact of nature, not a design flaw.
8. I Have a Dream
I would welcome a world where major AI actors optimize for human augmentation, rather than mimicking. A world where AI systems are designed for beneficial human-IA interactions, rather than raw problem solving with human-like language glued on top.
We often ask how to augment AI itself, but what kinds of training, interfaces, tools, etc. would best augment us? Saving and reading, leveraging approximate error correction mechanisms, integrating Turing machines, optimizing for human augmentation instead of mimicking would all be possible starting points—elements that are largely absent from current AI systems and will be challenging to do right in future ones.
I would also welcome a world where basic scientific literacy prevails over authority arguments, confirmation bias, and other weaknesses of humans as brainy mammals.
One example of an authority argument I see frequently is the alternatives to LLMs put forward by people like Yann LeCun, such as hierarchical JEPAs.
JEPAs make more coherent compositional processing units. Do they fundamentally change what is missing when we do not have saving and reading, error correction, and so on? No. They are also quite simplistic, as renormalization group theory—a pillar of modern physics—provides a much more powerful framework for designing hierarchical representations. See, e.g., Gökmen et al., Nature Communications (2024).
Today, most people find anthropocentrism reassuring. Most people respond to authority arguments, and are not familiar with our most profound descriptions of reality—quantum circuits, information theory, and beyond. As long as this remains true, voices like Gary Marcus's will unfortunately continue to shape the majority view.
9. Conclusion: More Science, Less Rhetoric
Gary Marcus writes: "The lack of stable, robust world models matters enormously not just for business, and legal writing, agents, math, everyday understanding, video comprehension and video generation (to name a few) but, also, crucially, for AI safety." This much I agree. Though the solution is not more superficial engagement with AI concepts.
The way forward is not to cling to comforting myths about human superiority or to mistake authority for evidence. It is to confront the reality that intelligence—biological or artificial—emerges from statistical processing compensated by error-correction mechanisms and relentless information-gathering from the world.
Progress in the coming years will likely come from LLMs and more advanced neural networks encapsulated in ChatGPT-like interfaces, but this is not the best possible future. The less likely but more promising path is to ground ourselves in scientific fundamentals and build systems that truly augment human intelligence, moving beyond the shadow of the Gary Marcuses of the world.
Note: I deliberately omit citations because this analysis rests entirely on fundamental principles from physics, neuroscience, and information theory. The language used to express these principles is irrelevant. Look beyond rhetorical satisfaction, focus on the science.
Finally, to be fair, Gary Marcus raises legitimate concerns about AI safety and regulation that transcend what I believe to be technical misconceptions. His calls for international coordination, transparency requirements, and accountability mechanisms address real risks and inspire me every day. The irony is that his safety concerns actually strengthen the case for more scientific approaches to AI, and more human augmentation than mimicking in AI products.