It seems hard to imagine, if its training has been on small chunks of text, that the model has a way of understanding a large codebase. But this stuff keeps on

It seems hard to imagine, if its training has been on small chunks of text, that the model has a way of understanding a large codebase. But this stuff keeps on surprising me.