I don’t see how you can write the law such that it allows training ai on copyrighted data without making it possible to train a special llm on a single github instead of the entire universe, and essentially treat it as a full compression of the source.