Towards shutdownable agents via stochastic choice

Elliott Thornley (Global Priorities Institute, University of Oxford), Alexander Roman (New College of Florida), Christos Ziakas (Imperial College, London), Leyton Ho (Brown University) and Louis Thomson (University of Oxford)

GPI Working Paper No. 16-2024

The POST-Agents Proposal (PAP) is an idea for ensuring that advanced artificial agents never resist shutdown. A key part of the PAP is using a novel ‘Discounted Reward for Same-Length Trajectories (DReST)’ reward function to train agents to (1) pursue goals effectively conditional on each trajectory-length (be ‘USEFUL’), and (2) choose stochastically between different trajectory-lengths (be ‘NEUTRAL’ about trajectory-lengths). In this paper, we propose evaluation metrics for USEFULNESS and NEUTRALITY. We use a DReST reward function to train simple agents to navigate gridworlds, and we find that these agents learn to be USEFUL and NEUTRAL. Our results thus provide some initial evidence that DReST reward functions could train advanced agents to be USEFUL and NEUTRAL. Our theoretical work suggests that these agents would be useful and shutdownable.

Other working papers

Economic growth under transformative AI – Philip Trammell (Global Priorities Institute, Oxford University) and Anton Korinek (University of Virginia)

Industrialized countries have long seen relatively stable growth in output per capita and a stable labor share. AI may be transformative, in the sense that it may break one or both of these stylized facts. This review outlines the ways this may happen by placing several strands of the literature on AI and growth within a common framework. We first evaluate models in which AI increases output production, for example via increases in capital’s substitutability for labor…

The paralysis argument – William MacAskill, Andreas Mogensen (Global Priorities Institute, Oxford University)

Given plausible assumptions about the long-run impact of our everyday actions, we show that standard non-consequentialist constraints on doing harm entail that we should try to do as little as possible in our lives. We call this the Paralysis Argument. After laying out the argument, we consider and respond to…

What power-seeking theorems do not show – David Thorstad (Vanderbilt University)

Recent years have seen increasing concern that artificial intelligence may soon pose an existential risk to humanity. One leading ground for concern is that artificial agents may be power-seeking, aiming to acquire power and in the process disempowering humanity. A range of power-seeking theorems seek to give formal articulation to the idea that artificial agents are likely to be power-seeking. I argue that leading theorems face five challenges, then draw lessons from this result.