5 ms·
That's assuming the consequent: you are assuming that such an enumeration strategy is requisite for strong AI to occur. [EDIT: I misunderstood what you meant b
by Dn_Ab 14y ago
That's assuming the consequent: you are assuming that such an enumeration strategy is requisite for strong AI to occur. [EDIT: I misunderstood what you meant by enumeration, my statement of your assumption is wrong. What you do assume is that in order to be able to outperform humans at any task a Strong AI must not be tripped by diagonalization. Or that humans would not be any such output, that is that the human can somehow step outside your conundrum.]
Yet it works just as well to consider humans to be char programs which map strings to strings. But then, they somehow manage exhibit Strong AI. Actually I believe Scott Aaronson has used this table look up type argument to show how you could theoretically pass the Turing Test.
Now
imagine a lookup table that stores every possible history H of A and B’s conversation, and next to
H, the action fB (H) that B would take next given that history. Of course, like Borges’ Library
of Babel, the lookup table would consist almost entirely of meaningless nonsense, and it would also
be much too large to fit inside the observed universe. But all that matters for us is that the lookup
table would be finite, by the assumption that there is a finite upper bound on the conversation
length. This implies that the function fB is computable (indeed, it can be recognized by a finite
automaton!). From these simple considerations, we conclude that if there is a fundamental obstacle
to computers passing the Turing Test, then it is not to be found in computability theory.
But you are going in the right direction. More developed versions of your argument lead to why a Bayes Optimal AI is not possible.
Note also that as Strong AI is a vague term I have defaulted to the nearest solid metric: prove it is conscious.
- Dove 14y agoI'm not assuming an enumeration strategy is a prerequisite for strong AI. An enumeration strategy already exists. Heck, I don't even have to construct a mapping--a compiled program, when you get right down to it, is a number. You are assuming human beings--with all of their code and inputs--can be mapped to numbers. I'd say that's the more unlikely assumption, given that we don't completely understand how they work. Heck, I don't think we're even sure they're deterministic.
- Dn_Ab 14y agoNo I am not assuming anything other that what you wrote. Consider an AI to be a chat program which maps strings to strings I could just as well have written a chat program that for any question (including this), merely looked up the correct form of an answer (wrong or not) without attempting to run anything. But before any kind of meaningful conversation can continue I need to know what your definition of Strong AI is. Because what you have given is certainly not proof against the common definition of Artificial General Intelligence. Since I cannot prove the Church Turing Thesis I will not try to argue against your belief. I will only note that it is simpler to believe that there is nothing special about what collections of neurons compute than to assume something..more.. is going on. Neural implants should soon enough settle the question.
- Dove 14y agoWould you be satisfied if, instead, I claimed to be proving that it is impossible to construct a program -- in the sense of a real, honest-to-god compiled executable -- that can respond to statements in English text at least as well as a human being would?
- Dn_Ab 14y agoI have already quoted a strong argument against such an impossibility: http://www.scottaaronson.com/papers/philos.pdf http://www.scottaaronson.com/papers/philos.pdf Practicality is another matter for which, what you have is not a proof against. Like I said, you can use arguments from computability to show why a Bayes Optimal AI is impossible. But there is nothing stopping an arbitrarily close approximation. If you are interested in this I strongly urge you to familiarize yourself with the current literature. It pays to start new ideas from a point that does not cover past work. http://www.hutter1.net/ai/uaibook.htm http://www.hutter1.net/ai/uaibook.htm
- Dove 14y agoI disagree. Here is the resolution. The paper suggests that a finite table could be compiled which exhaustively enumerates all the possible conversations I could have with an AI (to be more generous than he is, I'll say) in my lifetime. Some of those will persuade me that the AI is conscious. We can make a (finite) program that follows that mapping, and hence it has to be possible at least in terms of a finite-length program. I disagree with the middle assumption: some of those will persuade me the AI is conscious. I think it is entirely possible, if I pose the question I did up above, there is no possible response the AI could give me which would persuade me it was conscious. No matter how clever its response is, the program is just a lookup table, and hence whatever response it gives me will match the output of the specified program, demonstrating that the AI indeed did not understand the request. In other words, I take issue with the author's assumption that there is some subset that will work. It is possible the sparse set the author is hoping to fish out has size zero. (The difference between the AI and the human, in the chat logs, is that the AI has source code I can reference. The human does not. The human can run any program and produce a different output. An AI cannot do this with its own program.)