Can't agree more. The problem isn't generating code. It's articulating a functional specification with enough precision and clarity that you can solve for the

Can't agree more. The problem isn't generating code. It's articulating a functional specification with enough precision and clarity that you can solve for the unknown, which is the program that satisfies the specification. I agree with the conclusion that tools like ChatGPT won't be replacing programmers any time soon. And it won't be helping us to write programs until it can properly reason about the information it is trained on and prompted with. The problem with natural language is that it is often too imprecise. When you get to concurrency or optimizing data layout an LLM isn't going to understand anything about the problem space and the training data probably won't contain enough examples to produce something remotely plausible let alone good or elegant. Writing, thinking, and asking the right questions will still be important for quite some time. We're not in the age of having a magical genie in our computers that can do all the work for us... yet. Although using it as a procedural generation tool for fake data is nice... still not as nice as writing a property test with generators and shrink... but as a little time saver for quick prototyping, neat.