Transcript Week 10
Philosophy 4610 Philosophy of Mind Week 9: AI in the Real World The “Chinese Room” The Chinese Room ► In the Chinese Room, there is a rule book for manipulating symbols and an operator who does not understand any Chinese ► The Room produces perfectly good Chinese answers and could pass a Turing Test conducted in Chinese ► But nothing in the room actually understands Chinese The Chinese Room ► According to Searle, in the Chinese Room there is intelligent-seeming behavior but no actual intelligence or understanding. There is syntax (rules for the manipulation of meaningless signs) but the semantics or meaning of the signs is missing. This shows, Searle argues, that rule-governed behavior is not enough to give real understanding or thinking. The Chinese Room: The “Systems” Reply ► Even if there is no single element in the Chinese Room that understands Chinese, perhaps the understanding of Chinese really is in the whole system itself. ► What are the criteria for “really understanding” as opposed to just seeming to understand? What role (if any) does experience, consciousness, or self-awareness play? How might we test for these qualities? The Loebner Prize ► ► ► Every year, philanthropist Hugh Loebner sponsors a “real-life” Turing Test He offers $100,000 to any computer program that can successfully convince a panel of judges that it is “more human” than at least one human subject Every year, $2,000 is offered for the program judged “most human”. The Loebner Prize ► Which of the transcripts seemed “most human”? Which did not seem “human” at all? Why? The Loebner Prize: Things to Look For ► Ambiguity. Many words in English have multiple meanings. For example: “He put a check on the board” (here ‘check’ can mean either a monetary instrument, or a mark). ► ‘Canned’ responses. Many of the responses that a computer might give seem “automatic” or inappropriate to the situation (how can you tell?) ► Jokes and puns. It is difficult for computers to understand jokes or puns that depend on the difference between literal and metaphoric meaning (why?) Dreyfus and What Computers Can’t Do ► ► Like his colleague Searle, Dreyfus thinks that it will be much harder than many have assumed to build a real thinking machine. He argues that it is much more difficult than it seems to “program in” ordinary, practical intelligence of a kind we exhibit constantly and everyday. Artificial Intelligence: Two Approaches ► The “frame” approach (Minsky): To get a computer to exhibit actual intelligence, we just have to program it with an appreciation of the “frame” or context of ordinary human situations. ► The “script” approach (Schank): To get a computer to exhibit intelligence, we just need to represent a “script” or plan for handling ordinary situations (sitting in a chair, ordering at a restaurant, cooking an egg, etc.) Scripts and Frames: Trying it Out ► Let’s try to “program” an AI system to handle some ordinary tasks. ► We’re allowed to specify any RULE that we want, provided that the rules are welldefined in terms of the information available to the system. Dreyfus: how do you sit in a chair? ► “Anyone in our culture understands such things as how to sit on kitchen chairs, swivel chairs, folding chairs, and in arm chairs, rocking chairs, deck chairs, barbers’ chairs, sedan chairs, dentists’ chairs, basket chairs, reclining chairs, wheel chairs, sling chairs, and beanbag chairs – as well as how to get off/out of them again. … (p. 163). Dreyfus: The assumption of traditional AI ► There is a great deal of knowledge that we rely on everyday and use in a wide variety of situations that is not explicit. ► Traditional AI research assumes that this knowledge is all representable – that it can be programmed into a computer by inputting a finite set of rules. ► But Dreyfus argues that there is no reason to think that this knowledge must be representable this way. Real-World AI: Summary ► Classical AI research, following Turing, assumes that it’s possible to get a computer to be intelligent by programming it with some finite set of rules. ► But passing a Turing test – or even being able to function in everyday situations – requires a vast amount of knowledge that is not generally explicit. ► Is it possible to represent this knowledge at all? If it is not representable, then how do we acquire it? Might an artificial system or robot be able to acquire it as we do, even if it cannot be ‘programmed in’ explicitly?