Speed arguments against scheming (Section 4.4-4.7 of "Scheming AIs")

Joe Carlsmith Audio

This is section 4.4 through 4.7 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. Speed arguments against scheming (Section 4.4-4.7 of "Scheming AIs") (00:00:00)

2. 4.4 Speed arguments (00:00:29)

3. 4.4.1 How big are the absolute costs of this extra reasoning? (00:02:22)

4. 4.4.2 How big are the costs of this extra reasoning relative to the simplicity benefits of (00:07:06)

5. 4.4.3 Can we actively shape training to bias towards speed over simplicity? (00:09:21)

6. 4.5 The “not-your-passion” argument (00:10:27)

7. 4.6 The relevance of “slack” to these arguments (00:12:46)

8. 4.7 Takeaways re: arguments that focus on the final properties of the model (00:13:38)