Get the latest tech news
Do we understand how neural networks work?
Yes and no, but mostly no.
Neural networks are made of matrices, a perfectly ordinary piece of math that many millions of people have learned about in college. We understand the end result, as much as we do, by making educated guesses, following hunches, and, occasionally, digging into the network to try to reverse engineer exactly how and why it does a specific thing. I know that we made the LLM by having a neural network try to solve autocomplete, and I know that it is writing a poem because I asked it to, but I have very nearly no idea what precisely the training put inside its various matrices that enables it to do that.
Or read this on Hacker News