marginalium

A note on

www.lesswrong.com

May 21, 2023

On the alien characteristics of LLMs: the Waluigi effect.

Short version:

After you train an LLM to satisfy a desirable property P, then it’s easier to elicit the chatbot into satisfying the exact opposite of property P

Why?

When you spend many bits-of-optimisation locating a character, it only takes a few extra bits to specify their antipode.

Join over 2000 of us. Get the newsletter.