“The most significant near-term limit to scaling large language models is the availability of training data.”