[eeue56]

I wrote a toy language along these lines a while back[0]. Basically, types and function signatures, with comments in English, produce a valid program. You write a type and a comment, and the compiler goes through GPT to run the code. Fun novel idea.

[0] - https://github.com/eeue56/neuro-lingo

[Mathnerd314]

This paper suggests that pseudocode is more understandable to LLM's than natural language. So building off of that, one should not write comments in the method bodies, one should translate a function body (in mixed English and pseudocode) as a unit. LLM's can memorize a lot but there is a limit so I guess it would have to be structured as a multi-level translation. You mention in the Readme that it is persnickety, this is probably due to the use of "auto-complete" in the prompt. I would use an intro like "here is a function body in pseudocode", then prompts "split it into chunks by inserting labels", "translate whole body preserving chunk labels", "re-translate each chunk using whole body as context", "fix type errors". I think the biggest issues besides prompt engineering are cost and latency though.