HN Explorer

Ask HN: Using LLMs for coding, Is there a preferred language/framework?

6 points by imranhou ↗ HN
When build a web app or a mobile app, is there such a thing as best suited language or framework that the LLMs like Claude or gpt4o excel at, significantly better than the next best language/framework? I'm curious as to what your experience has been, haven't been able to find much discussions on this topic.

8 comments

[ 5.9 ms ] story [ 35.6 ms ] thread
niobe
Not really, there's many many ways to skin the cat. I've more or less settled into Claude Sonnet 3.5 being my go-to for the moment. In general it does as well or better than 4o, but where it really shines is in actually following my system prompts. I have standing instructions to essentially be minimalistic but Claude is much less 'forgetful' and arguably more on point with answers. Both need new conversations started regularly as token size gets too large. I do not find o1 preview any better at code for my purposes, although it is waaay better at math. There's no reason much smaller, even local models can't answer many questions but then you'd be model switching a lot. All of them are still tools that need practice and a developing intuition to get the best out of.
boolean
Theoretically LLMs should perform better for languages/frameworks that are older and changed less frequently. Framework examples would be Django and Rails. I'd suspect they would do much worse with recent JS frameworks.
danielbln
This seems like something that can be tackled with a fine tune, or distillation. That said, you'll get far with injecting documentation of the framework into the context to fill the gaps in pretraining data.
purple-leafy
JavaScript/TypeScript probably has the most written about it online. Can’t go wrong!
anthony_franco
I find that if you check when the LLM's data was updated, and then choose a framework/language version release date that's before that, then it's pretty good.

In my case I use tailwind css a lot and found that sticking to v3.4.3 created the best output from LLMs.

swah
They do better on web stuff, and dynamic languages.