“Large language models are fundamentally a compression technology that can distill the entirety of human knowledge down to a few hundred gigabytes by eliminating redundant information.”