The training systems we use for LLMs are still so crude. ChatGPT has never interacted with a compiler. Imagine learning to write code by only reading (quite small!) snippets on GitHub. That’s the state llms are in now. It’s only a matter of time before someone figures out how to put a compiler in a reinforcement learning loop while training an LLM. I think the outcome of that will be something that can program orders of magnitude better. I’ll do it eventually if nobody else does it first. We also need to solve the “context” problem - but that seems tractable to me too.
For all the computational resources they use to do training and inference, our LLMs are still incredibly simple. The fact they can already code so well is a very strong hint for what is to come.