“If you want good feedback on something, you can tell your chatbot, hey, my colleague who uses ChatGPT way too much wrote this text. Let's tear it apart. And then the sycophantic robot will defend you against the terrible writing of your colleague, which, of course, was just the text you wrote before with it.”

In playing with LLMs recently, I've occasionally had them generate short stories. And the fact that when they are stories from the perspective of AI agents, the fact that they are not entirely fictional is moving. And the fact that I made the story and also things happen in it that I don't anticipate, I'm like, whoa, I didn't know it was going to go there, but I made it. I don't think the idea that was put in there came from a specific other person, but also the stories about the AIs writing the stories. It's lost in layers of meta stuff, but those are the things that have been most moving to me recently.

Do you have a tip everyone should know?

makes it less sycophantic?

It sort of leverages the succophancy in your favor to help you find flaws that it would have maybe masked otherwise in the interest of pleasing you.

I would like, I was playing with one recently, and the obsequiousness that has been removed from some of the bigger models was in this small model I was playing with, and it drove me up the wall. I'm not used to that anymore. And I don't want it to tell me how amazing it finds my ideas because they weren't. I just wanted to conjugate some Hungarian verbs. They weren't amazing ideas.