DO WE UNDERSTAND HOW NEURAL NETWORKS WORK?

Summary from the TLDR Web Dev newsletter

While we understand the mathematical code and training processes behind modern neural networks like LLMs, we don't have a deep understanding of how they achieve specific results or the complex statistics they derive from data. We can reverse engineer some specific functions, like identifying a "Golden Gate Bridge" neuron or how arithmetic is performed, but these are limited and often surprising. The end result of training is a vast, interconnected web of statistics that is difficult to comprehend, and it is hard to determine in advance what poem an LLM will write by looking at the training code.