Marvin Minsky On AI

An anonymous reader writes "In a three-part Dr. Dobbs podcast, AI pioneer and MIT professor Marvin Minsky examines the failures of AI research and lays out directions for future developments in the field. In part 1, 'It's 2001. Where's HAL?' he looks at the unfulfilled promises of artificial intelligence. In part 2 and in part 3 he offers hope that real progress is in the offing. With this talk from Minsky, Congressional testimony on the digital future from Tim Berners-Lee, life-extension evangelization from Ray Kurzweil, and Stephen Hawking planning to go into space, it seems like we may be on the verge of another AI or future-science bubble."

Comments

Coordination Lacking

[Tablizer]

I think the biggest problem with AI is lack of integration between different intelligence techniques. Humans generally use multiple skills and combine the results to correct and hone in on the right answer. These include:

* Physical modeling
* Analogy application
* Formal logic
* Pattern recognition
* Language parsing
* Memory
* Others that I forgot

It takes connectivity and cordination between just about all of these. Lab AI has done pretty well at each of these alone, but has *not* found way to make them help each other.

Oh, the bogosity

[Animats]

In the 1980s I believed that "strong AI" was forthcoming...

In the 1980s, I was going through Stanford CS, where some of the AI faculty were indeed saying that. Read Feigenbaum's "The Fifth Generation", to see how bad it got. It was embarrassing, because very little actually worked.. Expert systems really were awfully dumb. They're just another way to program, as is generally recognized today. But back then, there were people claiming that if you could only write enough rules, intelligence would somehow emerge. I knew it was bogus at the time, and so did some other people, but, unlike most grad students, I was working for an big outside company, not a professor, and could say so. At one point I noted that it was possible to graduate in CS, in AI, at the MSCS level, without ever actually seeing an expert system work. This embarrassed some faculty members.

There was a massive amount of self-delusion in Stanford CS back then. When the whole AI boom collapsed, CS at Stanford was moved from the School of Arts and Sciences to Engineering, to give the place some adult supervision. Eventually, the Stanford AI Lab was dissolved. It's been brought back in the last few years, but with new people.

We're making real progress today, finally. Mainly because of a shift to statistical methods with sound mathematical underpinnings, plus enough compute power to make them go. Trying to hammer the real world into predicate calculus was a dead end. But number crunching is working. Computer vision actually sort of works now. Robots are starting to work. Automatic driving works. Language translation works marginally. Voice recognition works marginally. There are real products now.

But the AI field really was stuck for over a decade. The phrase "AI Winter" has been used.

Re:WHAT COMPUTERS STILL CAN'T DO

[hahiss]

I'm afraid you've misunderstood Dreyfus's work. His work, like Searle's, does not deny that our minds are *like* (to use your locution) computers. What he denies is that our minds engage the world in a way that is (totally) capturable in propositional form and so are formal programs of the sort

What Dreyfus argues is that there are parts of human experience that aren't capturable in in an unambiguous and propositional form, and so the sort of artificial intelligence that proceeds by trying to code frame systems will fail (unless the AI is specialized for a task that can be brute forced, like chess playing). Put another way: having a theoretical grasp of an activity isn't the same as knowing how to do it (you can be brilliant with fluid dynamics theory and suck at swimming); it is this latter element that Dreyfus calls "skillful coping" and he argues that this isn't capturable by traditional AI programs. Moreover, there is a difference between the cognition of expert humans and such AI systems; chess masters, for example, don't brute force the computations.

Notice that this doesn't mean he argues that it is impossible that machines could think or that robot doppelgangers couldn't be built---just that the mainstream approaches won't work. I believe that Dreyfus would be pleased with the approaches that Mark Tilden and Rodney Brooks have taken to AI, for example.

(None of this is to say that he's right, though I suspect he is. )