The behavior of CPython and a few other implementations of Python (such as PyPy) is well documented and well understood. The semantics of the tiny subset of Python that this Python-to-eBPF compiler understands is not. For example, inferring from the fact that it statically compiles Python-ish AST to LLVM IR, you can have a rough idea that dynamic elements of Python semantics are unlikely to be compiled, but you cannot know exactly which elements without carefully reading the documentation or source code of the compiler. You can guess globals() or locals() won't work, maybe .__dict__ won't as well, but how about type() or isinstance()? You don't know without digging into the documentation (which may be lacking), because the subset of Python this compiler understands is rather arbitrary.
And also, having an LLM translate Python-ish pseudo code into C does not imply that you cannot examine it before putting it into a program. You can manually review it and make modifications as you want. It just reduces time spent compared with writing C code by hand.
Writing C for eBPF is cumbersome and you'd like to avoid it. Okay, that's reasonable. But I don't think it would be a good idea to write a compiler that emits eBPF binary from (a tiny subset of) Python. Why not just write code in pseudo-Python (or whatever language you're comfortable with) and have it translated by an LLM, and paste it in the source code? That would be much better because there would be fewer layers and a significant reduction in runtime cost.
This reads like GPT-5 output. Anyone familiar with the model will recognize its distinctive style. While using LLM-generated content isn't inherently wrong, why not share the prompts? It's like presenting a book summary without naming the book.
And also, having an LLM translate Python-ish pseudo code into C does not imply that you cannot examine it before putting it into a program. You can manually review it and make modifications as you want. It just reduces time spent compared with writing C code by hand.