Understanding Reward Tampering in AI

The 'model organisms of misalignment' line of research creates AI models that exhibit various types of misalignment, and studies them to try to understand how the misalignment occurs and whether it can be somehow removed. In this episode, Evan Hubinger talks about two papers he's worked on at Anthropic under this agenda: "Sleeper Agents" and "Sycophancy to Subterfuge".

Patreon: https://www.patreon.com/axrpodcast

Ko-fi: https://ko-fi.com/axrpodcast

The transcript: https://axrp.net/episode/2024/12/01/episode-39-evan-hubinger-model-organisms-misalignment.html

Topics we discuss, and timestamps:

0:00:36 - Model organisms and stress-testing

0:07:38 - Sleeper Agents

0:22:32 - Do 'sleeper agents' properly model deceptive alignment?

0:38:32 - Surprising results in "Sleeper Agents"

0:57:25 - Sycophancy to Subterfuge

1:09:21 - How models generalize from sycophancy to subterfuge

1:16:37 - Is the reward editing task valid?

1:21:46 - Training away sycophancy and subterfuge

1:29:22 - Model organisms, AI control, and evaluations

1:33:45 - Other model organisms research

1:35:27 - Alignment stress-testing at Anthropic

1:43:32 - Following Evan's work

Main papers:

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training: https://arxiv.org/abs/2401.05566

Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models: https://arxiv.org/abs/2406.10162

Anthropic links:

Anthropic's newsroom: https://www.anthropic.com/news

Careers at Anthropic: https://www.anthropic.com/careers

Other links:

Model Organisms of Misalignment: The Case for a New Pillar of Alignment Research: https://www.alignmentforum.org/posts/ChDH335ckdvpxXaXX/model-organisms-of-misalignment-the-case-for-a-new-pillar-of-1

Simple probes can catch sleeper agents: https://www.anthropic.com/research/probes-catch-sleeper-agents

Studying Large Language Model Generalization with Influence Functions: https://arxiv.org/abs/2308.03296

Stress-Testing Capability Elicitation With Password-Locked Models [aka model organisms of sandbagging]: https://arxiv.org/abs/2405.19550

Episode art by Hamish Doodles: hamishdoodles.com