Empirical work that might shed light on scheming (Section 6 of "Scheming AIs")

Joe Carlsmith Audio

This is section 6 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

บท

1. Empirical work that might shed light on scheming (Section 6 of "Scheming AIs") (00:00:00)

2. 6. Empirical work that might shed light on scheming (00:00:33)

3. 6.1 Empirical work on situational awareness (00:05:34)

4. 6.2 Empirical work on beyond-episode goals (00:07:03)

5. 6.3 Empirical work on the viability of scheming as an instrumental strategy (00:10:29)

6. 6.4 The “model organisms” paradigm (00:12:14)

7. 6.5 Traps and honest tests (00:13:29)

8. 6.6 Interpretability and transparency (00:16:49)

9. 6.7 Security, control, and oversight (00:18:35)

10. 6.8 Other possibilities (00:21:08)