quartztz

Hello!

Being a student interested in PL design, I have had this idea floating around for a while: the gist is finding out what programming languages LLMs might be the most proficient in, to study their design choices and syntactic features with the goal of designing the perfect language for LLMs. This is, of course, gimmicky, but I entertained the idea for a while as a fun afterschool project.

The challenge is: what would be the best way to evaluate programming performance _in specific languages_? There are two main hypotheses here:

1. There are intrinsic syntactic/structural features that the transformer architecture is uniquely able to parse/reproduce/understand best, leading to higher quality code generated. For example: Lisp dialects make parsing code structure and blocks very easy, so one could assume an LLM can "understand their code better" 2. There is so much Python/JS out there that the question isn't even worth asking, and the performance in those will beat whatever other language you throw at it. This is probably not as much of a point thanks to newer transformer architectures but the question is still up.

I suspect the answer can be made somewhat interesting by considering performance relative to language popularity, but the ground question is: is there a general dataset containing different programming challenges, of varying difficulty, in multiple languages, with standard solutions? I couldn't find anything when I looked around, but I might have missed something obvious. It wouldn't be impossible to build a simple website to crowdsource, but I'm thinking that if I missed something obvious I'd rather find out early than late. Also, if you have any input on the project itself, I'd love to hear your ideas!