automate creating nix packages

[KiaraGrouwstra]

This is kind of a meta-issue -- many other (dormant) issues here come down to packaging a package for Nix. I wonder if it may be possible to use machine learning to try to package software.

Initial guesses may result in packaging errors, but the point would be to train it on how to deal with different types of errors, by generating sample solutions by deviating from working packages, showing the error, then showing the (original) solution.

Ideas on how to tackle this:

• tree prediction: recursive neural networks, predicting the structure of a whole nix expression?
• reinforcement learning: predicting actions of how to change a file/AST, maybe stepping back to the previous version?

[KiaraGrouwstra]

also, learn from Hydra errors and subsequent fixes.

[KiaraGrouwstra]

notes on this:

auto-nix: use ML to generate nix configs to package software; train by randomly altering existing nix expressions (randomly remove a part), teaching it the proper solution (fixed original) for different error messages - https://github.com/jameysharp/autobake

related: tree-sitter AST diffs for nix (https://discourse.nixos.org/t/announcing-tree-sitter-nix/2483)