Non-classic stories about scheming (Section 2.3.2 of "Scheming AIs")

Joe Carlsmith Audio

This is section 2.3.2 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

Chapters

1. Non-classic stories about scheming (Section 2.3.2 of "Scheming AIs") (00:00:00)

2. 2.3.2 Non-classic stories (00:00:36)

3. 2.3.2.1 AI coordination (00:00:55)

4. 2.3.2.2 AIs with similar values by default (00:05:57)

5. 2.3.2.3 Terminal values that happen to favor escape/takeover (00:07:51)

6. 2.3.2.4 Models with false beliefs about whether scheming is a good strategy (00:11:59)

7. 2.3.2.5 Self-deception (00:13:33)

8. 2.3.2.6 Goal-uncertainty and haziness (00:15:46)

9. 2.3.2.7 Overall assessment of the non-classic stories (00:18:19)

10. 2.4 Take-aways re: the requirements of scheming (00:20:08)

11. 2.5 Path dependence (00:20:51)