Towards shutdownable agents via stochastic choice

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

GPI Working Paper No. 16-2024

Some worry that advanced artificial agents may resist being shut down. The Incomplete Preferences Proposal (IPP) is an idea for ensuring that doesn’t happen. A key part of the IPP 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 suggest that DREST reward functions could also train advanced agents to be USEFUL and NEUTRAL, and thereby make these advanced agents useful and shutdownable.

Other working papers

Longtermism in an Infinite World – Christian J. Tarsney (Population Wellbeing Initiative, University of Texas at Austin) and Hayden Wilkinson (Global Priorities Institute, University of Oxford)

The case for longtermism depends on the vast potential scale of the future. But that same vastness also threatens to undermine the case for longtermism: If the future contains infinite value, then many theories of value that support longtermism (e.g., risk-neutral total utilitarianism) seem to imply that no available action is better than any other. And some strategies for avoiding this conclusion (e.g., exponential time discounting) yield views that…

Minimal and Expansive Longtermism – Hilary Greaves (University of Oxford) and Christian Tarsney (Population Wellbeing Initiative, University of Texas at Austin)

The standard case for longtermism focuses on a small set of risks to the far future, and argues that in a small set of choice situations, the present marginal value of mitigating those risks is very great. But many longtermists are attracted to, and many critics of longtermism worried by, a farther-reaching form of longtermism. According to this farther-reaching form, there are many ways of improving the far future, which determine the value of our options in all or nearly all choice situations…

Will AI Avoid Exploitation? – Adam Bales (Global Priorities Institute, University of Oxford)

A simple argument suggests that we can fruitfully model advanced AI systems using expected utility theory. According to this argument, an agent will need to act as if maximising expected utility if they’re to avoid exploitation. Insofar as we should expect advanced AI to avoid exploitation, it follows that we should expected advanced AI to act as if maximising expected utility. I spell out this argument more carefully and demonstrate that it fails, but show that the manner of its failure is instructive…