@_dbrogle on Instagram: "https://arxiv.org/pdf/1301.3781 All LLMs use embeddings, and the concept of putting words into lists of numbers was so bad that it was unusable until Google did the classic throw more compute and training data at it, but also used two simple tasks to induce the AI into being properly aligned into encoding words into meaningful lists of numbers. The finer points of how they’re generated has since changed, but the fundamental concept of training a model to encode words int