It uses a combination of LLMs, selecting the best output. (from the blog post)