[xigoi]

> The database of code on GitHub is many terabytes large, but the model trained on it is significantly smaller.

This just means it's a really efficient lossy compression algorithm, not that it learns like a human.