I think this is related enough and useful to mention. Feel free to downvote if I am wrong.
I am a student in this class taught by Scott Aaronson. Our class blog is linked there too, which may be of some value to LW readers. I think comments from people who are not enrolled in the class are okay, as long as they are sparse. The main point is for students to hash out their reactions to readings and in-class discussions, so if it becomes saturated with remote user comments, it may take away from the cohesive debate aspect.
The course is based around this essay which offers a lot of food for thought. I definitely think LW will get mixed into our discussions; I know a few other students who also regularly read LW.
I really like that essay, and it was linked here before. I would have liked to enroll in the course.
The key insights I took away from the paper are:
No, a waterfall isn’t simulating a chess program, because the mapping from one to the other would be “doing all the work”. IOW, you can only say one computation implements another if there’s a polynomial time reduction between the two that benefits from having the other as an oracle, which is not the case for waterfalls and chess.
Different choices of language have no asymptotic impact on complexity of descriptions, except to the extent that it limits expressive power. If you couldn’t represent the concept of a differential equation, then Newton’s gravitation would be no simpler than Ptolemy’s epicycles: both would require a lot of table lookups to predict motion.
The Turing Test, as well as human-run tests in general (like a quiz in school) can be subsumed into the concept of interactive proofs, where assumptions about the subject plus random probes of their knowledge suffice to prove what they are capable of, without knowing how they implement or represent it.
When considering whether a machine can pass the Turing Test, the relevant question is not if it’s theoretically possible, but if you could do it while only using resources that increase polynomially in the length of the test (rather than e.g. a superexponential lookup table).
IOW, you can only say one computation implements another if there’s a polynomial time reduction between the two that benefits from having the other as an oracle, which is not the case for waterfalls and chess.
That is a very concise and informative way of stating it.
The Turing Test, as well as human-run tests in general (like a quiz in school) can be subsumed into the concept of interactive proofs, where assumptions about the subject plus random probes of their knowledge suffice to prove what they are capable of, without knowing how they implement or represent it.
When considering whether a machine can pass the Turing Test, the relevant question is not if it’s theoretically possible, but if you could do it while only using resources that increase polynomially in the length of the test (rather than e.g. a superexponential lookup table).
These are interesting points of view, which I tried to probe more deeply, or at least offer some counter views here. I think most people felt I was just bashing Watson, which wasn’t my intention. Chapter 7 of the new book The Beginnings of Infinity by David Deutsch also spells out a good reason why hardware and software concerns beyond the guarantees made by an interactive proof Turing test might be necessary to really capture what we might want to mean by intelligence.
I think this is related enough and useful to mention. Feel free to downvote if I am wrong.
I am a student in this class taught by Scott Aaronson. Our class blog is linked there too, which may be of some value to LW readers. I think comments from people who are not enrolled in the class are okay, as long as they are sparse. The main point is for students to hash out their reactions to readings and in-class discussions, so if it becomes saturated with remote user comments, it may take away from the cohesive debate aspect.
The course is based around this essay which offers a lot of food for thought. I definitely think LW will get mixed into our discussions; I know a few other students who also regularly read LW.
I really like that essay, and it was linked here before. I would have liked to enroll in the course.
The key insights I took away from the paper are:
No, a waterfall isn’t simulating a chess program, because the mapping from one to the other would be “doing all the work”. IOW, you can only say one computation implements another if there’s a polynomial time reduction between the two that benefits from having the other as an oracle, which is not the case for waterfalls and chess.
Different choices of language have no asymptotic impact on complexity of descriptions, except to the extent that it limits expressive power. If you couldn’t represent the concept of a differential equation, then Newton’s gravitation would be no simpler than Ptolemy’s epicycles: both would require a lot of table lookups to predict motion.
The Turing Test, as well as human-run tests in general (like a quiz in school) can be subsumed into the concept of interactive proofs, where assumptions about the subject plus random probes of their knowledge suffice to prove what they are capable of, without knowing how they implement or represent it.
When considering whether a machine can pass the Turing Test, the relevant question is not if it’s theoretically possible, but if you could do it while only using resources that increase polynomially in the length of the test (rather than e.g. a superexponential lookup table).
That is a very concise and informative way of stating it.
These are interesting points of view, which I tried to probe more deeply, or at least offer some counter views here. I think most people felt I was just bashing Watson, which wasn’t my intention. Chapter 7 of the new book The Beginnings of Infinity by David Deutsch also spells out a good reason why hardware and software concerns beyond the guarantees made by an interactive proof Turing test might be necessary to really capture what we might want to mean by intelligence.