Packaging AI/ML models as conda packages

I have a love and hate relationship with conda. For a long time, it was the only way to reliably install ML stuff on Windows or difficult to package dependencies in general. They've generally been a force for good in the AI/ML development space, for packaging and supporting pandas/numpy/etc if I'm not mistaken.

But jeez, the DX is bad now, they also avoided implementing most of the PEPs that make Python packaging less of an issue now. The way conda work is also really invasive, it changes the PYTHONPATH left and right, sometimes does dynamic linking with the system libs, sometimes not... It pushes dark patterns like constant virtual env re-use, that inevitably breaks.

They should also have jumped on the issue of model versioning, data versioning and overall artifacts management in the ML space but didn't, not sure why.