"The Waluigi Effect (mega-post)" by Cleo Nardo
from LessWrong (Curated & Popular)
https://www.lesswrong.com/posts/D7PumeYTDPfBTp3i7/the-waluigi-effect-mega-post In this article, I will present a mechanistic explanation of the Waluigi Effect and other bizarre "semiotic" phenomena which arise within large language models such as GPT-3/3.5/4 and their variants (ChatGPT, Sydney, etc). This article will be folklorish to some readers, and profoundly novel to others.
Embed this episode
Chapters
- 0:32 Background
- 0:56 Prompting LLMs with direct queries
- 3:40 Prompting LLMs with flattery and dialogue
- 5:05 Simulator Theory
- 9:03 The limits of flattery
- 12:16 Derrida — il n'y a pas de hors-texte
- 15:17 The Waluigi Effect
- 17:00 (1) Rules are meant to be broken.
- 19:45 (2) Traits are complex, valences are simple.
- 22:03 (3) Structuralist narratology
- 25:42 Superpositions will typically collapse to waluigis
- 27:48 Evidence from Microsoft Sydney
- 28:42 Waluigis after RLHF
- 31:19 (2) Empirical evidence from Perez et al.
- 33:11 (3) RLHF promotes mode-collapse
- 40:22 Conclusion
No transcript yet.
Frequently asked questions
How long is this episode of LessWrong (Curated & Popular)?
This episode is 41 minutes long.
When was this LessWrong (Curated & Popular) episode published?
This episode was published on March 8, 2023.
Can I download this LessWrong (Curated & Popular) episode?
Yes. Use the download control on the episode player to save the publisher-provided media file.