Users spend tokens on DeepSeek, Claude, and others -> generate variations of projects uploaded to GitHub -> Microsoft (GitHub's owner) -> trains LLMs on this -> ends up with better models.

No matter what model a user may use at any price point, microsoft can distill it.

It's distillation at a higher level

most good programmers aren't going to commit garbage solutions; they'll push it to a git remote once they see it actually solves a problem they are trying to solve, and since the git commits and their messages are also AI generated, you've good visibility into the user's intention and resulting edits to the codebase.

They want to capture it all.