I guess you have completely missed people running language models on WebGPU then.
I think webasm + webgpu will be a target for a lot of new ml libraries.
I think the key difference is that it’s hard to setup good governance for JavaScript. You’re going to need a fascist linter that is actually enforced, and you’re going to need a tight grip on your developers to force them to really, really, think about their dependencies, but once you have that it’s quite honestly the best language I can think off. The ability to use isolated functions instead of putting them into “classes” is just such a great way to do a solid mix of functional OOP programming, which is obviously heresy to hardliners of either, but it’s just so magical when you do it right.
I think Python is getting there, it was such a great language for such a long time that it sort of forgot to improve. But now it has copied the NPM/Yarn package handler, and hopefully it’ll soon be possible to actually do a Typescript sort of Python, so maybe it’ll be able to win me back. Or to be fair, I think it’s a great language until you have to work together. It’s just so hard to get the codebase governance up with Python that it only really becomes worth it in ML shops where your developers want to work with Python. I’m not sure how Instagram managed, and there are certainly the projects that fit into the Django box which absolutely should be put into the Django box, but the only general purpose language to me personally is currently JavaScript.
Part of that is because we need initiatives like this one. We need “React” in Python or Rust or whatever if we want small dev teams in non-tech enterprise to be able to work with other languages than Typescript. Yes I know we have some C++, and a little Rust, but unlike the rest of our many different projects I’m the only one who can maintain them. Which is actually the primary reason we work with JavaScript, because if we don’t, then the React developer won’t ever be able to go on a vacation. :p It helps that JavaScript has become such a great language, and it likely has exactly because the React dev wants to go on vacation in a trillion IT departments. But I’m all for Python having this React Python so ML heavy shops don’t need that React dev.
But to say JavaScript is atrocious is sort of silly to me and I’m not sure you would have that opinion if you gave it a real chance.
As for static types to be honest I don't see how it's useful in the context of ML. You might use a Pandas dataframe, which already comes with types enforced inside of it. You shouldn't ever use a for loop on a Pandas dataframe or whatever because then you're running Python instead of C++, Pandas has inbuilt functions and operators.
However Python does have type hints but probably not strict enough, Mojo may improve in this if it really supports both AOT and JIT.
Package management in Python isn't that bad with requirements.txt. The real problem is Python versions are breaking and whatnot (very often ML libraries are months behind latest) but the main Python installation you have only supports virtual environments for packages. Really it should support something like conda out of the box where you create an environment with a Python version.
But it’s not too terrible, it’s it’s not Node or Cargo, but it is really hard to build governance around.
TypeScript requires you to compile, the whole point of Python for ML and data science is you're running in an interpreted environment where you delegate as much code out as you can for both data processing and algorithms to invisible C++.
The interpreted part is key as ML is about experimentation. It's not like there's any overlap between current web developers and ML people anyway to need an unified bridge.