OpenAI Codex Python to C++ Code Generator | Hacker News Reader

OpenAI Codex Python to C++ Code Generator

github.com

36 pointsArdelleF22 comments

I'm writing a Python to C++ translator for https://www.oilshell.org. The shell itself is less than 40K lines of code and the translator is ~4-8K lines.

And I even have a 50K euro NLNet grant to hire someone!

Brief Descriptions of a Python to C++ Translator https://www.oilshell.org/blog/2022/05/mycpp.html

---

But I don't think there is anything interesting in this repo. It doesn't show anything that works, and doesn't really make any claims.

But maybe it would be a fun illustration of the limitations of AI. So I would propose that if you can get even 1000 lines of Oil's code translating and running correctly with AI, I will award you a big part of my grant.

I think it's impossible with the state of the art. I don't have much interest in current ML techniques, but I'm credited on a deep learning paper with thousands of citations, so it's not a totally uneducated claim.

---

On the other hand, I'm looking for a C++ / Python / compiler engineer, particularly in the European Union. Please take a look at Oil Is Being Implemented "Middle Out" [1] and send leads my way :) (e-mail in profile)

[1] https://www.oilshell.org/blog/2022/03/middle-out.html

Anything you can learn or use from Shedskin?

https://github.com/shedskin/shedskin

It's mentioned in the blog post

So it is, I missed the link, thanks.

Great post BTW. That is cool you got to cross paths with John Plevyak. That IFA paper is pretty damn great. He worked on Wasm support for Envoy you know. Maybe Plevyak is your compiler writer? :)

Ah OK I haven't looked at IFA, but I wonder how it would scale to 40K lines of code? As far as I know that is beyond the state of the art, i.e. a research level problem.

Yeah I should get in touch for sure... I remember talking with John about the Apache Traffic Server (formerly Yahoo) and his Python-to-C++ pyc compiler almost 10 years ago!

Right now I'm contacting a bunch of people in the EU, since the funders have some preference there. But yes Shed Skin is one of the closest things!

You might want to contact the author of Pythran [1], maybe something can be learned from what they do.

[1] https://github.com/serge-sans-paille/pythran/commits/master

Yeah interesting, I have seen most of these projects (mentioned in the blog post) but not looked closely at Pythran.

It does say it's specifically for numeric code. That is what I hinted at in the blog post around use cases for Shed Skin vs. string- and dict-oriented code.

There are quite a few numeric accelerators for Python, and I even chatted with Travis Oliphant (of NumPy) about another one he was prototyping.

But there are fewer for the kind of program a shell is. mypyc and shed skin are the 2 I mentioned that are most similar. We already have an approach that works, but it just needs "labor".

You want to translate fully type annotated python to C++ to be able to compile as a stand-alone program?

From what I’ve read in the past inferring the types is the hard part (the scheme stuff I saw used abstract interpretation) then just did a relatively simple translation step to end up with C. Well, continuations were also troublesome as were dependency graphs but IIRC that’s because they needed types to compute them. So, yeah, easy peasy if you already have the types.

Interesting idea.... I wonder if the same could be done for c/c++ -> rust :D

C++ is already very safe if written in a modern style. Ive not had a real modern C++ program crash or UB on me in a while. It would be interesting to translate old C++ to modern C++, which should be pretty easy

I am happy for you.

I once liked c++ but after years of writing it professionally. I am done.

You have to fight c++ to get things working which obfuscates problems in the code. Safety comes from simplicity -- c++ is never simple.

I could rant about this for hours. But in my opinion rust is a better language and the one we as software engineers should get behind for safe systems.

C++ is old and needs to be retired.

Sure, we can start by rewritting LLVM/GCC or CUDA/HIP/SYCL in Rust then, or maybe get some foundation to sponsor that work.

Not to be that guy, but tell me one successful rust rewrite that ended up being less buggy or more reliable

Lack of function overloading in Rust makes this fun!

Haha you are right there! That does make it a whole bundle of fun :)

Shameless plug for Copilot Labs, which offers arbitrary language-pair translations.

https://githubnext.com/projects/copilot-labs/

Warning, Codex is nondeterministic and sometimes nonsensical! Good prompts are hard to find.

So my understanding is it works like this: You take some Python you throw it at OpenAI and you see if the resulting C++ compiles. Doesn't the fact that you need to check the result even compiles also imply you probably want to check if the result is functionally correct also?

Yes certainly you may need to check that even if the code compiled

A question for the computer scientists here: are there applications for this where Codex could be used to automatically optimize complex systems?

Possibly, but if Codex is like similar AI text generation tools, complex systems or large software architectures with many interacting components are going to specifically be the corner cases where results are less reliable.

The real problem is how do you ensure that code alterations preserve observable behavior. Complex systems generally have complex specifications.

I think it would be simpler to convert Python to Nim, which itself compiles to C. That's assuming this is for performance gains only.

Reply on news.ycombinator.com