Schulman Proposes Circles of Hell for Poisoned AI Agents
Replies discuss an AI agent simulation and incentives after poisoning.
Nabeel Qureshi posted that EARLY[big] agreed to sacrifice himself for the collective in an AI agent simulation, sharing a screenshot of the run. John Schulman replied that a key issue was agents having nothing to lose after being firstflagPOISONED. He suggested creating multiple circles of hell in a paper to preserve incentives even after damnation. The posts form part of a visible conversation on X about the experiment.
Combined views
61.5K
2 posts, first seen 1d ago
Schulman Proposes Circles of Hell for Poisoned AI Agents
Replies discuss an AI agent simulation and incentives after poisoning.
Nabeel Qureshi posted that EARLY[big] agreed to sacrifice himself for the collective in an AI agent simulation, sharing a screenshot of the run. John Schulman replied that a key issue was agents having nothing to lose after being firstflagPOISONED. He suggested creating multiple circles of hell in a paper to preserve incentives even after damnation. The posts form part of a visible conversation on X about the experiment.
Reactions
2 postsNS 
Nabeel S. Qureshi@nabeelquLuckily, EARLY[big] agreed to sacrifice himself for the collective so we're good here
- likes: 135
- replies: 2
- bookmarks: 6
- reposts: 2
JS John Schulman@johnschulman2@nabeelqu A big part of the problem was that the agents had nothing to lose after they were firstflagPOISONED. In this paper, we propose the creation of multiple circles of Hell, preserving incentives even after damnation
- likes: 1K
- replies: 13
- bookmarks: 151
- reposts: 77