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
Non-additive axiologies in large worlds – Christian Tarsney and Teruji Thomas (Global Priorities Institute, Oxford University)
Is the overall value of a world just the sum of values contributed by each value-bearing entity in that world? Additively separable axiologies (like total utilitarianism, prioritarianism, and critical level views) say ‘yes’, but non-additive axiologies (like average utilitarianism, rank-discounted utilitarianism, and variable value views) say ‘no’…
Evolutionary debunking and value alignment – Michael T. Dale (Hampden-Sydney College) and Bradford Saad (Global Priorities Institute, University of Oxford)
This paper examines the bearing of evolutionary debunking arguments—which use the evolutionary origins of values to challenge their epistemic credentials—on the alignment problem, i.e. the problem of ensuring that highly capable AI systems are properly aligned with values. Since evolutionary debunking arguments are among the best empirically-motivated arguments that recommend changes in values, it is unsurprising that they are relevant to the alignment problem. However, how evolutionary debunking arguments…
Prediction: The long and the short of it – Antony Millner (University of California, Santa Barbara) and Daniel Heyen (ETH Zurich)
Commentators often lament forecasters’ inability to provide precise predictions of the long-run behaviour of complex economic and physical systems. Yet their concerns often conflate the presence of substantial long-run uncertainty with the need for long-run predictability; short-run predictions can partially substitute for long-run predictions if decision-makers can adjust their activities over time. …