It's qualitatively different to go through source code and specifications to understand how something works than to look at a database with all the weights of an LLM and pretend like you could predict the output.