Why focus on schemers in particular? (Sections 1.3-1.4 of "Scheming AIs")

Joe Carlsmith Audio

This is sections 1.3-1.4 of my report “Scheming AIs: Will AIs fake alignment during training in order to get power?”

Text of the report here: https://arxiv.org/abs/2311.08379
Summary of the report here: https://joecarlsmith.com/2023/11/15/new-report-scheming-ais-will-ais-fake-alignment-during-training-in-order-to-get-power
Audio summary here: https://joecarlsmithaudio.buzzsprout.com/2034731/13969977-introduction-and-summary-of-scheming-ais-will-ais-fake-alignment-during-training-in-order-to-get-power

Kapitler

1. Why focus on schemers in particular? (Sections 1.3-1.4 of "Scheming AIs") (00:00:00)

2. 1.3 Why focus on schemers in particular? (00:00:36)

3. 1.3.1 The type of misalignment I’m most worried about (00:01:14)

4. 1.3.2 Contrast with reward-on-the-episode seekers (00:04:27)

5. 1.3.2.1 Responsiveness to honest tests (00:04:46)

6. 1.3.2.2 Temporal scope and general “ambition” (00:07:54)

7. 1.3.2.3 Sandbagging and “early undermining” (00:11:17)

8. 1.3.3 Contrast with models that aren’t playing the training game (00:17:13)

9. 1.3.4 Non-schemers with schemer-like traits (00:23:13)

10. 1.3.5 Mixed models (00:25:20)

11. 1.4 Are theoretical arguments about this topic even useful? (00:28:35)