I agree with these observations; I just don't think they tell us much about human cognition. For instance, you can get V1 simple cells in all kinds of different ways mathematically. If you did independent component analysis first, and got simple cell receptive fields and concluded "oh, V1, and therefore the rest of the cortex, just does ICA!", you'd be quite wrong. (We don't know what the right answer is, but we do know it isn't that.)

We've done transformer-based generative LLMs first to create plausible language output. It does tell us something about the statistical structure of language as emitted by humans. But I don't think it necessarily tells us much about human cognition. (Not even the function of language-related cortices.)