They already built uv, which works extremely well for that

Yeah, and uv2nix is already pretty good! I wonder if pyx will be competing with uv2nix.

It's easy to compete with Nix tooling, but pretty hard to compete with the breadth of nixpkgs.

I'm guessing it's the right PyTorch and FlashAttention and TransformerEngine and xformers and all that for the machine you're on without a bunch of ninja-built CUDA capability pain.

They explicitly mention PyTorch in the blog post. That's where the big money in Python is, and that's where PyPI utterly fails.

There are a bunch of problems with PyPI. For example, there's no metadata API, you have to actually fetch each wheel file and inspect it to figure out certain things about the packages you're trying to resolve/install.

It would be nice if they contributed improvements upstream, but then they can't capture revenue from doing it. I guess it's better to have an alternative and improved PyPI, than to have no improvements and a sense of pride.

There is a lot of other stuff going on with Pyx, but "uv-native metadata APIs" is the relevant one for this example.