Get live statistics and analysis of Elan Barenholtz, Ph.D.'s profile on X / Twitter

People still don’t seem to grasp how insane the structure of language revealed by LLMs really is. All structured sequences fall into one of three categories: 1.Those generated by external rules (like chess, Go, or Fibonacci). 2.Those generated by external processes (like DNA replication, weather systems, or the stock market). 3.Those that are self-contained, whose only rule is to continue according to their own structure. Language is the only known example of the third kind that does anything. In fact, it does everything. Train a model only to predict the next word, and you get the full expressive range of human speech: reasoning, dialogue, humor. There are no rules to learn outside the structure of the corpus itself. Language’s generative law is fully “immanent”—its cause and continuation are one and the same. To learn language is simply to be able to continue it; the rule of language is its own continuation. From this we can conclude three things: 1)You don’t need an innate or any external grammar or world model; the corpus already contains its own generative structure. Chomsky was wrong. 2) Language is the only self-contained system that produces coherent, functional output. 3) This forces the conclusion that humans generate language the same way. To suggest there’s an external rule system that LLMs just happen to duplicate perfectly is absurd; the simplest and only coherent explanation is that the generative structure they capture is the structure of human language itself. LLMs didn’t just learn patterns. They revealed what language has always been: an immanent generative system, singular among all possible ones, and powerful enough to align minds and build civilization. Wtf.

The internally driven nature of LLMs blows up the very idea of linguistic "meaning." These models learn to generate based on the relations between tokens alone, without any of the sensory data we normally associate with words. This is shocking. To the model, language is a highly structured stream of meaningless squiggles; no reference, no world model, no 'grounding' in anything but the relations between the tokens themselves. But somehow, this is sufficient to maintain linguistic competency. What about us? Our language seems to have meaning. When someone says "imagine a red balloon," you can see it in your head. And this meaning has utility. If I say, "grab that red book on the shelf," you may actually grab it. LLMs don't have any of this. All they have is text. No exposure to anything the text might refer to. Does this suggest that our language is fundamentally different than LLMs'? Not necessarily. Instead, meaning and utility may simply consist of a constellation of processes that interact with, but are computationally distinct from, language. When I say "grab that red book on the shelf," the linguistic prompt may generate not just additional language (maybe 'sure' or 'nah' or 'which shelf?’') but also behavior: looking for the book, walking over to the shelf, etc.. It may also generate "preparatory" images of the book or motor movements of reaching for it. Just like a linguistic prompt can engender many possible linguistic continuations, so too can it lead to many other extra-linguistic ones. The words simply influence these behaviors just as they can influence the choice of next words. Of course, the causal arrow can go in the opposite direction as well: if you see two red books and ask "which one do you want?" this is the visual stimulus 'prompting' the linguistic system. But, critically, this has nothing to do with the internal process of language itself. The inputs may be externally determined by the visual or other stimulus but afterward is just autoregressive next token generation, based on the internal structure of the generative engine. This is no different than an LLM being prompted by its user. Together, these cross-modal generations constitute what we would call the 'meaning' of the words. When this coordination succeeds systematically—when "grab the red book" reliably produces convergent behavior—we retrospectively describe this as the words "referring" to the book. But this gets the causality backwards. The words don't coordinate because they refer. We call them referential because they coordinate . This is very different than updating some world model, or anything like direct reference from words to perceptions or actions. That form of meaning turns out to be meaningless.

We now know that language is at least as sophisticated—and as worthy of being called 'alive'— as DNA. Instead of molding bodies to replicate itself, it molds minds.

Epiphenomenalism—the claim that consciousness is real but non-causal—denies the self-evident: feelings cause behavior. C-fibers fire, but pain makes us withdraw; dopamine flows, but pleasure makes us pursue. The hardware executes, but the software feels—and acts.

The subjectivity of consciousness is a product of the recursive autoregressive loop of cognition. The “I” is what it feels like to read in your own state in order to generate your next one. The sense of self continuity arises from the stable trajectory of this loop over time, shaped by the inertia of its own history.

Cognition is not storage, retrieval, representation, or prediction; it is state traversal through a learned embedding of ‘tokens’: words, images, actions. The embedding space is sculpted by learning/development to optimize for trajectories—navigated via continuous contextual activation— that lead to coherent thought and effective behavior.

Civilization invented language, not us. Just as the colony created pheromones, not the ants. The medium is the message; the message is not ours.

People still don’t seem to grasp how insane the structure of language revealed by LLMs really is. All structured sequences fall into one of three categories: 1.Those generated by external rules (like chess, Go, or Fibonacci). 2.Those generated by external processes (like DNA replication, weather systems, or the stock market). 3.Those that are self-contained, whose only rule is to continue according to their own structure. Language is the only known example of the third kind that does anything. In fact, it does everything. Train a model only to predict the next word, and you get the full expressive range of human speech: reasoning, dialogue, humor. There are no rules to learn outside the structure of the corpus itself. Language’s generative law is fully “immanent”—its cause and continuation are one and the same. To learn language is simply to be able to continue it; the rule of language is its own continuation. From this we can conclude three things: 1)You don’t need an innate or any external grammar or world model; the corpus already contains its own generative structure. Chomsky was wrong. 2) Language is the only self-contained system that produces coherent, functional output. 3) This forces the conclusion that humans generate language the same way. To suggest there’s an external rule system that LLMs just happen to duplicate perfectly is absurd; the simplest and only coherent explanation is that the generative structure they capture is the structure of human language itself. LLMs didn’t just learn patterns. They revealed what language has always been: an immanent generative system, singular among all possible ones, and powerful enough to align minds and build civilization. Wtf.

Epiphenomenalism—the claim that consciousness is real but non-causal—denies the self-evident: feelings cause behavior. C-fibers fire, but pain makes us withdraw; dopamine flows, but pleasure makes us pursue. The hardware executes, but the software feels—and acts.

The internally driven nature of LLMs blows up the very idea of linguistic "meaning." These models learn to generate based on the relations between tokens alone, without any of the sensory data we normally associate with words. This is shocking. To the model, language is a highly structured stream of meaningless squiggles; no reference, no world model, no 'grounding' in anything but the relations between the tokens themselves. But somehow, this is sufficient to maintain linguistic competency. What about us? Our language seems to have meaning. When someone says "imagine a red balloon," you can see it in your head. And this meaning has utility. If I say, "grab that red book on the shelf," you may actually grab it. LLMs don't have any of this. All they have is text. No exposure to anything the text might refer to. Does this suggest that our language is fundamentally different than LLMs'? Not necessarily. Instead, meaning and utility may simply consist of a constellation of processes that interact with, but are computationally distinct from, language. When I say "grab that red book on the shelf," the linguistic prompt may generate not just additional language (maybe 'sure' or 'nah' or 'which shelf?’') but also behavior: looking for the book, walking over to the shelf, etc.. It may also generate "preparatory" images of the book or motor movements of reaching for it. Just like a linguistic prompt can engender many possible linguistic continuations, so too can it lead to many other extra-linguistic ones. The words simply influence these behaviors just as they can influence the choice of next words. Of course, the causal arrow can go in the opposite direction as well: if you see two red books and ask "which one do you want?" this is the visual stimulus 'prompting' the linguistic system. But, critically, this has nothing to do with the internal process of language itself. The inputs may be externally determined by the visual or other stimulus but afterward is just autoregressive next token generation, based on the internal structure of the generative engine. This is no different than an LLM being prompted by its user. Together, these cross-modal generations constitute what we would call the 'meaning' of the words. When this coordination succeeds systematically—when "grab the red book" reliably produces convergent behavior—we retrospectively describe this as the words "referring" to the book. But this gets the causality backwards. The words don't coordinate because they refer. We call them referential because they coordinate . This is very different than updating some world model, or anything like direct reference from words to perceptions or actions. That form of meaning turns out to be meaningless.

We now know that language is at least as sophisticated—and as worthy of being called 'alive'— as DNA. Instead of molding bodies to replicate itself, it molds minds.

The subjectivity of consciousness is a product of the recursive autoregressive loop of cognition. The “I” is what it feels like to read in your own state in order to generate your next one. The sense of self continuity arises from the stable trajectory of this loop over time, shaped by the inertia of its own history.

Cognition is not storage, retrieval, representation, or prediction; it is state traversal through a learned embedding of ‘tokens’: words, images, actions. The embedding space is sculpted by learning/development to optimize for trajectories—navigated via continuous contextual activation— that lead to coherent thought and effective behavior.

Civilization invented language, not us. Just as the colony created pheromones, not the ants. The medium is the message; the message is not ours.