|
The Irrelevance of AI -- the Reverse Turing TestFirst version of the test: Suppose you are a regular participant in a newsgroup, and someone claims that one of the correspondents on the list is an AI (i.e. an AI which has gone beyond the level of human equivalence). You ask it questions and try to judge by its answers whether it is really an AI or not. In addition to asking direct questions, you also post messages on various subjects and see what kind of reply you get. The "AI" keeps up its end of the conversation, brilliantly. It doesn't give you the usual Turing Test nonsense. Its replies are on point, not evasive or tricky. It obviously has an understanding of the world at least as rich as yours. There is no doubt that you are conversing with an at-least-human intelligence. The only question is whether it is more than human. It isn't omniscient, but it does have most of math, science, and engineering at its fingertips, and it exhibits a high degree of originality in its answers. In the newsgroup discussion, the purported AI doesn't just reply to topics raised by others, it initiates its own threads, and raises points that had not occurred to anyone else. It moves effortlessly from embryology to fixed point theorems to the Yoga Sutras. It's marvelously inventive, and it seems to be drawing on some new kind of ontology that you can only guess at. Many people on the list are a little unnerved by it. It is the dominant presence on the list, even though some of the humans on the list are brilliant themselves. Does this imply that it really is an AI? Not at all. Your correspondent could actually be a team of people and computers. Suppose IBM decided to play this game, and they put as much money and energy into it as they put into the Deep Blue- Kasparov match. They assemble a team of very bright humans with a variety of talents -- on the level of René Thom, Roger Penrose, Erwin Schrödinger, Northrop Frye, and Graham Cairns-Smith, for example [David Jones has been removed from this list, by the way] -- and provide them with a suite of high end workstations with a high bandwidth connection to the internet, plus whatever software they want (or whatever software they can create), plus whatever assistance they need from any part of IBM (including the team that created Deep Blue). In other words, they have the whole company backing them up. They practice playing the game until they can work together seamlessly. Could you distinguish between this team and a real AI? How? |
|
Second version of the test: suppose you have a purported AI that doesn't claim to be intelligent in general, it merely claims to be a "full fledged automated engineering system" that can design machines 10,000 times faster than a human engineer. IBM could pass that test easily. They wouldn't even have to put together a special team to take the test. IBM can already design a new microprocessor 10,000 times faster than a human engineer working alone. So can Sun, Intel, etc. Anything a "full fledged automated engineering system" could do can be simulated by a corporation, i.e. a team of humans and computers. Going back to the first version of the test, suppose you ask your correspondent to translate something written in Hungarian. It hesitates only briefly, and then produces a translation that makes sense. You give it something written in Italian, with the same result. It doesn't have to spend years learning a new language. It can read anything in any language, at least any major language, off the top of its head. Does this imply that it must be an AI? Not at all. There could be a human using a semi-automatic translator, instead of an AI program with a fully automatic translator. In other words IBM could have anticipated this question, and written software that would enable a bright human to look words up fast enough, and check grammar fast enough, to translate any (major) language in real time. (Obviously there would be a steep learning curve here. We are assuming a very bright human who has practiced using the system long enough to be good at it.) Now the interesting question is: which would come first? Which would be easier, to write a semi-automatic translation system, or a fully automatic translation system? I think the answer is obvious: the semi-automatic translator would be easier, and will be written first. When a translation system finally appears that really does what it is supposed to do -- i.e. it produces translations that make sense -- it will be semi-automatic. It will be a translator's workbench, not a free-standing program that tries to translate by itself. Which would be easier, to write a CAD program that helps human engineers and architects do their work, or to write an automated engineering system that designs things on its own? Which would be easier, to write an IDE that helps human programmers do their work, or to write a program that writes programs? CAD programs and IDEs already exist. We are still waiting for their fully automated counterparts. What about chess? Suppose Kasparov had a program to assist him, and he (plus his assistant) played Deep Blue (by itself). Kasparov would win, obviously. But the question I have raised in the other cases has a different answer this time: which came first, the semi-automatic system (the chess player's assistant) or the fully automatic system? Programs that play chess fairly well have been around for a long time, but programs to assist a human chess player have not yet appeared. For decades, ever since Turing and von Neumann, computer scientists have been stuck on the idea of creating fully automated systems. The idea of semi-automatic systems is just now coming into its own. Now that programmers have started to think in terms of IA (intelligence augmentation), semi-automatic systems will be created before fully automatic systems. There is no reason why fully automated systems would ever pull ahead, because a fully automatic system could always be used as the back end of a semi-automatic system. For example, Deep Blue could be modified so that instead of making its own moves, it assists a human player. It shows you what it thinks is the best move, and why. You can ask it to explore certain directions of play, and see the results in illuminating ways. Suppose Kasparov, assisted by a semi-automated chess program whose underlying chess engine is the same as Deep Blue's, played the original Deep Blue. The team would win, 6-0. This remains true if any kind of AI software is substituted for Deep Blue. The fundamental inequality: For any given task that an AI could do, a combination of humans and computers could do it easier and better, and (assuming that the idea of IA catches on) the human-plus-computer team will always get there first. Whatever capability AI has at any given time, humans assisted by computers will have already reached that point and moved ahead. To paraphrase Laplace: AI is an unnecessary hypothesis. This is the end of Part 3 of Nanotechnology without Genies, as originally written. The next logical page is the Singularity page, which was added later but really should have been the culmination of Part 3. |