If it's repeated enough times in the training data, it will become statistically the most likely continuation of that response. The difference between a database and statistical response is... not very clear. Just like you could learn a poem in a language you don't know by repeating various parts of it enough times.
Uh, obviously it can't recite the entirety of Dune, for example. But just as obviously, the Litany Against Fear is famous enough to likely exist verbatim thousands of times in its training corpus.
I don’t think this is obvious at all. Ignore the algorithm and temperature and all that for a bit and just have it show you the word probabilities. I wouldn’t bat an eye if the entirety of dune was in there.
Maybe obvious was too strong a word, but I very much doubt that it contains a lossless representation of any copyrighted novel-length work. Just due to the pigeonhole principle if not anything else. The model is a vastly compressed approximation of the full training corpus.
Even though vastly compressed, in terms of lossy vs. fidelity, it seems to be able to recite pretty much any bible verse you ask it for.
The intersection space of 10^15 parameters is a lot of pigeon holes though.
LLMs have crazy good emergent properties. Even the stupidest model you use for learning NN, you chuck a transformer on it and it makes a huge difference, and I have a hard time knowing really, really, really ... WHY? Yes keys/values/queries, words talking to each other and stuff. But why!
In a sense, it contains a compressed database of its entire training data. That's where its knowledge comes from.
at some points it starts to resemble more of a lossy compression algorithm.
The 'next most likely word' aspect of it all seem to encode entire sequences of text. Especially if it is a bit of a niche domain. Then these texts can be 'retrieved' from the 'compressed' set of examples.