There is a longstanding debate between rules-based and non-rules representations of knowledge, learning and intelligence in both humans and AI. The core of the rules-based argument is based in what Chomsky called "Plato's problem" and is in effect a claim that it is impossible for humans to learn open-ended principles. So they must be built-in - for humans, that means innate grammars, and for computers, that means rules-based AI. Systems like MuZero contradict that claim by mastering rules-based systems without ever needing to be told the rules. This has wide implications, because even if rules remain helpful (and they are!) they are not essential, which means we select rules to teach arbitrarily, and not because they are in some important sense foundational. See also: TechCrunch, BBC, and from Julian Schrittwieser a high-level description of the algorithm.
Today: 4 Total: 103 [Direct link] [Share]
