Is scheming more likely if you train models to have long-term goals? (Sections 2.2.4.1-2.2.4.2 of "Scheming AIs")

Joe Carlsmith Audio

This is sections 2.2.4.1-2.2.4.2 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

บท

1. Is scheming more likely if you train models to have long-term goals? (Sections 2.2.4.1-2.2.4.2 of "Scheming AIs") (00:00:00)

2. 2.2.4 What if you intentionally train models to have long-term goals? (00:00:38)

3. 2.2.4.1 Training the model on long episodes (00:01:23)

4. 2.2.4.2 Using short episodes to train a model to pursue long-term goals (00:04:33)