Hacker News Reader

My problem is the that copilot (actually codex) training data and model isn't available. If it was derived from open source code, that work must be released.

I'm bewildered why nobody else has brought this up.

View on HN

If it just took patterns out of open source and isn't overfit (there is evidence that it is overfit though) it seems ok. Viewing code doesn't actually taint you, clean room reimplementations are just done in a strict manner in corporations to remove all doubt legally, but aren't an actual legal requirement.

Is that not the central controversy? That it’s being training on software that has a license that prohibits such use?

Yes it is, when you reduce everything down.

Reply on news.ycombinator.com