For me, Barenholtz Brain-LLMs analogy needs more grounding

To embrace ideas beyond pure speculation, I need more science…

A friend pointed me to a fascinating discussion between professors Elan Barenholtz and William Hahn, who claim that human language works like LLMs.

I was captivated, completely carried away, but at the same time the lack of science in their claim bothered me. I listened to a second video where Barenholtz enthusiastically discusses his idea.

At that point, I send a message to Barenholtz: “I’m only just discovering your fascinating work, and one question immediately came to mind. Please forgive me if it sounds simplistic—you may have answered it a hundred times already. When we invented airplanes, we didn’t assume we had understood how birds, insects, or bats fly, since evolution found different solutions to the same problem. Likewise, the fact that large language models can speak in no way proves that we speak the same way they do. To claim similarity, one must first demonstrate that the underlying processes are alike.”

Barenholtz response: “That is just the right question and I appreciate the airplane analogy. Let’s take it seriously because it actually illustrates my point well. We observed the function of flight in nature, applied some basic techniques, and in the process discovered the principles that made flight possible—lift via forward momentum—in machines and birds. The specific means of leveraging this principle may be different (e.g., propellers instead of flapping), but the core principles that made flight possible are the same and were discovered, not engineered.

“The same is true here. LLMs revealed that the structure of language supports internal prediction—with no world model, no grounding, no explicit syntax. That this even works was a massive surprise. Language turns out to be autogenerative: it contains within itself the structure necessary to produce more of itself, purely through predictive continuation. This didn’t have to be true. But it is. That was the key discovery.

“And if this predictive structure is sufficient to generate naturalistic language, the burden is now on those who believe human language is generated through a completely different kind of mechanism to explain why language would evolve with such a rich, predictive structure incidentally. That would be like discovering that the airfoil shape produces lift, then insisting birds must fly some other way.

“How that structure is leveraged—through autoregression or another predictive process—is a separate question. I argue (and present data in an upcoming paper) that humans also generate language autoregressively, in a similarly recursive, context-sensitive way. But that’s secondary. The foundational insight is that language itself contains a self-generating structure, and that is what LLMs have now proven beyond any reasonable doubt.”

I’m not convinced by this answer, so I’m responding in more detail (while publishing the exchange on Substack—and in English, by the way, with the help of LLMs—because this discussion seems important to me and central to the historic transition we are currently experiencing).

My long answer:

I find your theory powerful and it does indeed seem to explain many things about human language. "LLMs revealed that the structure of language supports internal prediction—with no world model, no grounding, no explicit syntax." Indeed, no one can question this assertion anymore. Yes, this is a major discovery: language doesn't need grammar (which Pinker had already demonstrated, I believe). I think, like you, that something of the same order as in LLMs is at play in us, and moreover, as a writer, I've always known this. I don't think my sentences; they emerge word by word (it's more complicated in English).

I'm making an effort to critique your assertions because I sense you're going to face attacks from all sides. When you write to me: "The core principles that made flight possible are the same and were discovered, not engineered," I'm not sure about that. When we were making cannonballs fly, we couldn't conclude anything about bird flight and aerodynamics (and besides, bird flight was really understood only well after we mastered airplanes—we never claimed that birds flew with propellers, for example). In summary: discovering that wings produce lift tells us nothing about how a bird dynamically controls this lift, modulates its wingbeats, adjusts the angle of attack in real time. Function and implementation remain two distinct questions.

Even if the principles of flight were discovered rather than engineered, this doesn't necessarily apply to language. LLMs may use a structure that exists in language without being the one the brain uses (see exaptation below).

But perhaps flight is a bad example, since it's a process whose mechanism is observable, unlike language which occurs in our brains as in LLMs.

Let's take another classic example: vision, rediscovered more than twenty times by evolution with different mechanisms. When we invented photography, it didn't teach us much about the different visions in the animal kingdom. Even better: when we connect a camera to a computer, understanding this process doesn't allow us to assert much about how we see ourselves. Of course, some physical processes are the same, but signal processing is totally different (as it differs across animal species). Knowing a function and being able to reproduce it teaches us little about how it's implemented elsewhere.

Another example: we capture solar energy with silicon photovoltaic cells, but this process has nothing to do with plant photosynthesis. Mastering solar technology reveals nothing to us about the molecular mechanisms of photosynthesis.

Or again: we navigate with GPS, compasses, and maps, but many animals have navigation capabilities that we still don't fully understand. Migratory birds, sea turtles, salmon returning to their natal rivers—they probably use the Earth's magnetic field, stars, odors, but the exact mechanisms remain mysterious (and probably diverse). Our very effective navigation systems don't illuminate these biological GPS systems (we could probably apply LLM logic to this animal navigation).

In the interview with William Hahn, you said you don't believe in the existence of long-term and short-term memory, that all of that no longer makes sense. However, certain lesions destroy one or the other of these memories. This is widely documented. LLMs have long-term memory—their vector space—but the context we carry from prompt to prompt can also be seen as short-term memory. It's impossible to treat the brain as a black box, in the behaviorist manner, and bury neurobiology.

I believe, like you, that we've never been so close to understanding language, and the battle is going to rage with the orthodox. I'm trying to present principled objections to you, more to help than to contradict you. Because if you're right, it's more than refreshing and exhilarating. We must open the black box, connect your model to neurobiology, otherwise we remain in an unfalsifiable metaphysical approach.

It's paradoxical because I'm responding to you at a metaphysical level. But it seems important to me to propose experiments to test or refute your intuition, which good science demands. For example, how to test the autoregressivity of human language? As I said, my literary practice has convinced me of this autoregressivity, but being convinced isn't enough.

When you write "And if this predictive structure is sufficient to generate naturalistic language, the burden is now on those who believe human language is generated through a completely different kind of mechanism to explain why language would evolve with such a rich, predictive structure incidentally," you're thus throwing the ball back into the opposing camp, but this counterattack seems dangerous to me. The argument "why would this structure exist if that's not how we use it?" ignores a fundamental principle of evolution: exaptation. Many biological structures serve functions different from those for which they evolved. Bird feathers first served for thermoregulation before being recycled for flight.

Language could have evolved with a predictive structure for reasons that have nothing to do with autoregressivity—perhaps to facilitate cultural transmission, social learning, or group coordination. This structure could allow LLMs to function without being the mechanism our brain actually uses (just as dinosaurs didn't use feathers for flight).

I'm not seeking to contradict. To embrace your ideas beyond pure speculation, I need more science.