Photo by SHVETS production on Pexels

While Large Language Models (LLMs) present a linear output in the form of human language, their internal workings are far from linear. These sophisticated AI systems leverage parallel processing, performing complex matrix operations that involve embedding prompts into high-dimensional vector spaces. This allows LLMs to simultaneously update probabilities across countless dimensions, resulting in efficient and powerful language generation. The key takeaway is that although the result is linear, the inference process is massively parallel. This observation was recently shared in a thought-provoking Reddit discussion: [https://old.reddit.com/r/artificial/comments/1p6x12q/llms_do_not_think_linearlythey_generate_in/]