Skip to content
"The Waluigi Effect (mega-post)" by Cleo Nardo episode artwork

Mar 8, 2023 41 min

"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.

Episode details come from the podcast's RSS feed.

Embed this episode

Ready to play

"The Waluigi Effect (mega-post)" by Cleo Nardo

0:00 41:11

Chapters

  1. 0:32 Background
  2. 0:56 Prompting LLMs with direct queries
  3. 3:40 Prompting LLMs with flattery and dialogue
  4. 5:05 Simulator Theory
  5. 9:03 The limits of flattery
  6. 12:16 Derrida — il n'y a pas de hors-texte
  7. 15:17 The Waluigi Effect
  8. 17:00 (1) Rules are meant to be broken.
  9. 19:45 (2) Traits are complex, valences are simple.
  10. 22:03 (3) Structuralist narratology
  11. 25:42 Superpositions will typically collapse to waluigis
  12. 27:48 Evidence from Microsoft Sydney
  13. 28:42 Waluigis after RLHF
  14. 31:19 (2) Empirical evidence from Perez et al.
  15. 33:11 (3) RLHF promotes mode-collapse
  16. 40:22 Conclusion

No transcript yet.

Requests help us decide what to transcribe next.

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.