Which is funny since that's a much clearer case of "learning from" than outright compressing all open source code into a giant pile of weights by learning a low-dimensional probability distribution of token sequences.