[g051051]

Why not have the LLM go straight to LLVM IR? What would a program look like when you remove all (or most) of the layers of abstraction needed by humans? Or are LLMs too contaminated by the training data to do this? I almost wish I could try this.

[Daishiman]

You probably could but the thing is that low level languages do not encode intent as well as high level constructs. You can trivially make an iterator in assembly but in a high level language a `for`, `map`, and `reduce` have specific meaning that helps catch bugs.