A Puzzle about Further Facts

Abstract

The late Stephen Hawking has warned that AI could eventually "spell the end of the human race." Elon Musk has predicted that "robots will be able to do everything better than us." Meanwhile, AI systems are starting to outperform people in domains ranging from board games to speech recognition. Is humanity on the way out?For those not working in AI, it can be difficult to interpret highly visible achievements in the field. Take, for example, Watson's 2011 victory over human Jeopardy champions Brad Rutter and Ken Jennings. This was a stunning achievement: while it should surprise nobody that Watson had access to an encyclopedic amount of knowledge, Jeopardy is a game that requires more than that. The hard part -at least for AI systems, but often also for humansisn't having access to the relevant information, but rather understanding the clue well enough to link it to that. Even many AI researchers, myself included, thought this would remain beyond the capabilities of AI systems for a while to come. We were wrong.But does this mean that Watson had obtained a human-level understanding of the world? No. Watson also produced some cringeworthy responses, for example "What is Toronto?" for a clue in the "US cities" category. This is part of a broader pattern of AI systems achieving superhuman levels of performance, and yet making blunders that leave us scratching our heads. For example, researchers from Carnegie Mellon were able to consistently fool a face recognition system that one of them, clearly a man, was actress Milla Jovovich, by wearing carefully designed eyeglass frames [3].In both cases, what causes the mistake is that the AI system solves the problem in a way that is very different from how humans do it. Often, this involves picking up on some statistical pattern that can be used to surprisingly great effect, but that sometimes produces answers that lack any common sense. Moreover, if something changes about how the data is produced, performance may plummet. This is especially so when the change is intended to mislead the system, as in the case of adding the eyeglass frames.This gives some insight into which jobs, or parts of jobs, the AI systems of today and tomorrow are likely to take over from us. Tasks that require responding to the same kind of standardized input over and over again, with a clear measure

Foundations of Cooperative AI

“…As in the Prisoner's Dilemma, this outcome is Pareto-dominated by other outcomes (such as the example outcome in parentheses above). 8 However, now imagine 6 For more discussions on these types of questions, see, for example, Valberg (2007); Hare (2007Hare ( , 2009Hare ( , 2010; Merlo (2016Merlo ( , 2021; Conitzer (2019aConitzer ( , 2020.…”

Designing Preferences, Beliefs, and Identities for Artificial Intelligence

“…6 Recent literature in philosophy addresses the apparent datum that a single subjective experience is present (Valberg 2007;Hare 2007;2009;Hellie 2013;Merlo 2016;Conitzer 2018).…”

The Personalized A-Theory of Time and Perspective

“…He illustrates this with what he calls a "vertiginous question": why, of all subjects, is this subject (the one corresponding to the human being Benj Hellie) the one whose experiences are "live"? In other work (Conitzer, 2019), I explore whether the "liveness" of one particular perspective is a further fact -a fact that does not follow logically from the physical facts of the world -by considering the analogy to looking in on a simulated world through a virtual reality headset: besides the computer code that determines the physics of the simulated world, there must be additional code that determines which simulated agent's perspective to show on the headset.…”