Joseph M. Miller

Tony T. Wang, Adam Gleave, Tom Tseng et al. · berkeley

We attack the state-of-the-art Go-playing AI system KataGo by training adversarial policies against it, achieving a >97% win rate against KataGo running at superhuman settings. Our adversaries do not win by playing Go well. Instead, they trick KataGo into making serious blunders. Our attack transfers zero-shot to other superhuman Go-playing AIs, and is comprehensible to the extent that human experts can implement it without algorithmic assistance to consistently beat superhuman AIs. The core vulnerability uncovered by our attack persists even in KataGo agents adversarially trained to defend against our attack. Our results demonstrate that even superhuman AI systems may harbor surprising failure modes. Example games are available https://goattack.far.ai/.

Brycen D. Pearl, Joseph M. Miller, Hang Woon Lee

Earth observation satellites (EOSs) play a pivotal role in capturing and analyzing planetary phenomena, ranging from natural disasters to societal development. The EOS scheduling problem (EOSSP), which optimizes the schedule of EOSs, is often solved with respect to nadir-directional EOS systems, thus restricting the observation time of targets and, consequently, the effectiveness of each EOS. This paper leverages state-of-the-art constellation reconfigurability to develop the reconfigurable EOS scheduling problem (REOSSP), wherein EOSs are assumed to be maneuverable, forming a more optimal constellation configuration at multiple opportunities during a schedule. This paper develops a novel mixed-integer linear programming formulation for the REOSSP to optimally solve the scheduling problem for given parameters. Additionally, since the REOSSP can be computationally expensive for large-scale problems, a rolling horizon procedure (RHP) solution method is developed. The performance of the REOSSP is benchmarked against the EOSSP, which serves as a baseline, through a set of random instances where problem characteristics are varied and a case study in which Hurricane Sandy is used to demonstrate realistic performance. These experiments demonstrate the value of constellation reconfigurability in its application to the EOSSP, yielding solutions that improve performance, while the RHP enhances computational runtime for large-scale REOSSP instances.

Lee Sharkey, Bilal Chughtai, Joshua Batson et al. · deepmind

Mechanistic interpretability aims to understand the computational mechanisms underlying neural networks' capabilities in order to accomplish concrete scientific and engineering goals. Progress in this field thus promises to provide greater assurance over AI system behavior and shed light on exciting scientific questions about the nature of intelligence. Despite recent progress toward these goals, there are many open problems in the field that require solutions before many scientific and practical benefits can be realized: Our methods require both conceptual and practical improvements to reveal deeper insights; we must figure out how best to apply our methods in pursuit of specific goals; and the field must grapple with socio-technical challenges that influence and are influenced by our work. This forward-facing review discusses the current frontier of mechanistic interpretability and the open problems that the field may benefit from prioritizing.