"But now LLMs utterly prove that you can do NLU with zero input from linguistics. The point is, LLMs learned the linguistics. And logic, common sense and world knowledge. And even a lot of human ethics and creativity.
...But LeCun just was absent on it and now seems to be crying sour grapes by referring to LLMs as ‘dog-like’ intelligence and ignoring the clear evidence that the essence of human-like intelligence has emerged as a result of learning to predict the ‘next word’."
If you take the time to think a little more deeply about the issue, you should come away with realizing that the above statement is false. ChatGPT can mimic the output of natural language and outperform humans computationally (every computer outperforms humans), but it simply cannot think. Human linguistic compontence is not tantamount to "natural language processsing"; it's radically different. To avoid going into the intricacies, the first problem is that a human is a living organism (you can call it a survival machine if you want), and all its cognitive competencies are constrained by this fact. Current AI, while sporting a different architecture from traditional computation, is still very much nestled in the comptuational paradigm. All we've done with ChatGPT and transformers is create highly specialized tools that mimic some facet of our behaviour, with the advantage of computational augmentation.
ChatGPT will not soon be thinking about solving the inconsistencies between General Relativity and quantum mechanics. Even if we get it to say very smart things as output, this behaviourist approach to the mind has been disproven over and over again. There's real stuff going on in the mind that is directly relevant to our lingusitic and general cognitive competence. This doesn't mean that we need to necessarily architect a mind to build a machine that thinks, there might be some as yet unknown factor/s that contribute to this competence, but as far as we know our minds are the only blueprint we have. It's possible current AI research might achieve general intelligence by happenstance, but until that happens, we're not even close to engineering thinking machines.
