Hacker News Reader

Comment by jarbus | Hacker News Reader

I use Julia regularly for experimental machine learning. It’s great for writing high performance, distributed code and even easier than Python for this kind of work, since I can optimize the entire stack in a single language. Not sure if it’s growing in popularity but it’s really solid for what it does

View on HN

Me too, and I'd like it to become mainstream. The major problem right now is that it doesn't have anything that is close to Torch or JAX in performance and robustness. Flux et al. are 90% there, but the last 10% requires a massive investment, and Julia doesn't have any corporate juggernaut funding development like Meta or Google.

This is hurting Julia's adoption. The rest of the language is incredibly elegant, as there is no 2-language divide like in Python. Furthermore, it is really performant. With very little effort one can write code that is within 1.5-2x of C++, often closer.

One possibility is that something like Mojo takes Julia's spot. Mojo has some of the advantages of Julia, plus very tight integration with Python, its syntax and its ecosystem. I would still prefer Julia, but this is something to keep in mind.

LLMs massively compound the advantage of existing popular languages, namely python. Any new learner will find it infinitely easier to use sonnet 3.5 to overcome the so called '2 language barrier' for python, while the lacking data for Julia becomes the real barrier.

This issue will remain until LLMs get so smart they can maybe self-iterate and train on a given language. By then though, we'd likely get languages designed and optimized for LLMs.

To back up the sibling comment, I've found ChatGPT quite capable where Julia is concerned. It does hallucinate the occasional standard library function, but a) it gets it right after it's told it was wrong about half the time and b) Julia's documentation is fairly good, so finding what that function is really called is not a big deal.

It can even debug Pkg/build chain problems, which... Julia could use a bit of polish there. On paper the system is quite good, but in practice things like point upgrades of the Julia binary can involve a certain amount of throwing spaghetti at the wall.

There is a paper that claims that Julia performs the best in ChatGPT: https://arxiv.org/abs/2308.04477

Chris blog on that: https://www.stochasticlifestyle.com/chatgpt-performs-better-...

For what it's worth I've found Claude Sonnet to work really well with Julia.

One fun exercise was when a friend handed me a stack of well-written, very readable Python code that they were actually using. They were considering rewriting it in C, which would have been worth it if they could get a 10x speedup.

I had Sonnet translate it to Julia, and it literally ran 200x faster, with almost identical syntax.

More fragmented, calling-some-api based python code actually causes LLM to hallucinate and mix libraries more: https://www.stochasticlifestyle.com/chatgpt-performs-better-...

IMO Mojo's memory model is way too unfriendly for python programmers.

Could you elaborate? As far as I understand, if you treat it like Python (e.g. use defs and stick with the copy-on-modification default), you'll still see performance improvements without even thinking about memory.

Reply on news.ycombinator.com