I’m not sure what that article is supposed to prove. They are using sone computational language and focusing physical responses to visual stimuli but I don’t think it shows “neural computations” as being equivalent to the kinds of computations done by a TM.

One of the chief functions of our brains is to predict the next thing that going to happen, where it's the images we see or the words we hear. That's not very different from genML predicting the next word.

Why do people keep saying this, very obviously human beings are not LLMs.

I'm not even saying that human beings aren't just neural networks. I'm not even saying that an LLM couldn't be considered intelligent theoretically. I'm not even saying that human beings don't learn through predictions. Those are all arguments that people can have. But human beings are obviously not LLMs.

Human beings learn language years into their childhood. It is extremely obvious that we are not text engines that develop internal reason through the processing of text. Children form internal models of the world before they learn how to talk and before they understand what their parents are saying, and it is based on those internal models and on interactions with non-text inputs that their brains develop language models on top of their internal models.

LLMs invert that process. They form language models, and when the language models get big enough and get refined enough, some degree of internal world-modeling results (in theory, we don't really understand what exactly LLMs are doing internally).

Furthermore, even when humans do develop language models, human language models are based on a kind of cooperative "language game" where we predict not what word is most likely to appear next in a sequence, but instead how other people will react and change our separately observed world based on what we say to them. In other words, human beings learn language as tool to manipulate the world, not as an end in and of itself. It's more accurate to say that human language is an emergent system that results from human beings developing other predictive models rather than to say that language is something we learn just by predicting text tokens. We predict the effects and implications of those text tokens, we don't predict the tokens in isolation of the rest of the world.

Not a dig against LLMs, but I wonder if the people making these claims have ever seen an infant before. Your kid doesn't learn how shapes work based on textual context clues, it learns how shapes work by looking at shapes, and then separately it forms a language model that helps it translate that experience/knowledge into a form that other people can understand.

"But we both just predict things" -- prediction subjects matter. Again, nothing against LLMs, but predicting text output is very different from the types of predictions infants make, and those differences have practical consequences. It is a genuinely useful way of thinking about LLMs to understand that they are not trying to predict "correctness" or to influence the world (minor exceptions for alignment training aside), they are trying to predict text sequences. The task that a model is trained on matters, it's not an implementation detail that can just be discarded.