This is one of the most interesting articles I've read ... maybe this year. The author also has another one that is super cool: https://vgel.me/posts/representation-engineering/

I wish I could take the time to learn more about this because it looks like a way to make LLMs more useful beyond additional prompting techniques.