The fundamental bit I was aiming at here is that running gradient descent or similar training algorithms on a thing doesn't mean you will understand the end result fully - and we don't understand it fully. Its entirely possible that to predict the next word correctly, ChatGPT had to build / encode some kind of model of the universe (as written down by humans) within its weights.

We know its not possible to encode the full probabilities of what words will follow what other words, as the article itself states [1].

So how do you best "compress" these probabilies? By trying to find the most correct generalizations that are more widely appplicable? Perhaps even developing meta-facilities in recognizing good generalizations from bad?

[1]: https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-....