I used to be in the camp of "GPT-2 / GPT-3 is a glorified Markov chain". But over the last few months, I flipped 180° - I think we may have accidentally cracked a core part of "generalized intelligence" problem. It's not about the language, as much about associations - it seems to me that, once the latent space gets high-dimensional enough, a lot of problems reduce to adjacency search.

I'm starting to get a (sure, uneducated) feeling that this high-dimensional association encoding and search is fundamental to thinking, in a similar way to how a conditional and a loop is fundamental to (Turing-complete) computing.

Now, the next obvious step is of course to add conditionals and loops (and lots of external memory) to a proto-thinking LLM model, because what could possibly go wrong. In fact, those plugins are one of many attempts to do just that.

> It can reason correctly about hard problems with very little information.

i am so tired of seeing people who should know better think that this program can reason.

(in before the 400th time some programmer tells me "well aren't you just an autocomplete" as if they know anything about the human brain)

I have a deprecated function and ask ChatGPT what I should use instead, ChatGPT responds by inventing a random non existent function. I tell it that the function doesn't exist, it tries again with another non existent function.

Oddly speaking that sounds like a very simple language level failure, i.e. the tool generates text that matches the shape of the answer but not its details. I am not far enough into this ChatGPT religion to gaslight myself over outright lies like Elon Musk fanboys seem to enjoy doing.