I assumed (heard somewhere) that they only include open source repos in the training data.

Turns out I was wrong. They don't care.

https://web.archive.org/web/20210708165143/https://twitter.c...