In Stephen Wolfram's write up on the workings of GPT, he suggests that chat GPT may have identified invariant rules of human language that haven't been formalized yet, ie new linguistic findings, and that a much simpler model taught these rules could perform simiarly:

What does it take to produce “meaningful human language”? In the past, we might have assumed it could be nothing short of a human brain. But now we know it can be done quite respectably by the neural net of ChatGPT. Still, maybe that’s as far as we can go, and there’ll be nothing simpler—or more human understandable—that will work. But my strong suspicion is that the success of ChatGPT implicitly reveals an important “scientific” fact: that there’s actually a lot more structure and simplicity to meaningful human language than we ever knew—and that in the end there may be even fairly simple rules that describe how such language can be put together.

It sounds like in the ~96 layers of processing in GPT, some portion of it might be doing general language processing and some later portion might be involved in knowledge storage/access. I've also heard that sometimes eg in OCR, we can inspect the intermediary layers of the neural net and begin to guess what features each part is looking for.

So, I'm wondering if anyone has tried to inspect the layers of a publicly available chat model eg Llama and see what the intermediary processing stages are doing, or if it is too complex to be understandable in any respect.



You must log in to answer this question.