Playing Schelling games is a key dangerous capability for schemers: it's much harder to control AIs that are very capable at Schelling coordination. In this post, I discuss the threat model and describe concrete projects for empirical work.

Suppose you wish to control models that adversarially try to cause a catastrophic outcome (“schemers”). A useful class of techniques is monitoring: before letting any action taken by the potentially-scheming policy be executed, you use another model as a monitor to verify the actions are not dangerous.

A key question about monitors is whether they are trustworthy or not, i.e. whether the monitor strategically tries to deceive humans and let dangerous actions be executed. Untrusted monitoring is a technique that uses possibly-untrustworthy models for monitoring. The benefit is that we have no restrictions on the models we may use for monitoring, and in particular we can use our most [...]

---

Outline:

(04:20) Scope of this post

(05:07) Setup

(06:51) Challenges

(08:35) Concrete projects

(08:44) #1: Identify the Schelling element

(12:29) #2: Identification vs. classification

(13:58) #3: Training and generalization of Schelling coordination

(15:22) #4: Adversariality and non-obvious Schelling points

(20:27) Conclusion

The original text contained 9 footnotes which were omitted from this narration.

The original text contained 1 image which was described by AI.

---

First published:
October 8th, 2024

Source:
https://www.lesswrong.com/posts/n2c62YS8RJtD4Aqhh/schelling-game-evaluations-for-ai-control

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

“Schelling game evaluations for AI control” by Olli Järviniemi

LessWrong (30+ Karma)