Moving the Goal Posts and Artificial Intelligence

From Mind Hacks:

Mind Hacks: Minsky slams modern AI: [Marvin Minsky makes] a throwaway comment about the 'moving goal posts' problem in the perception of artificial intelligence, that belies much of the problem with how AI is perceived.

It is illustrated by the success of chess computers. In the 60s, it was said that computers will never beat people at chess, because that requires intelligence and computers aren't capable of intelligent thought.

When computers regularly started winning matches in the 80s, it was claimed that playing chess wasn't a test of real intelligence because computers could do it.

As there is no widely accepted definition for intelligence, this is often an example of the No true Scotsman fallacy.

Comments

[Cranky Observer]

> t is illustrated by the success of chess computers.
> In the 60s, it was said that computers will never beat
> people at chess, because that requires intelligence
> and computers aren't capable of intelligent
> thought.

Baloney. Even the computers of the 1960s, and certainly the 1970s/80s, had a clock rate and calculation speed as far beyond the human brain as a GM SD90MAC locomotive is beyond human muscle. Yet it wasn't until the late 1990s that anyone built a chess computer that could regularly beat a human grandmaster (or even master), and even now if Kasparov or Fisher were in their prime they could still provide stiff competition to the machine.

To me, that is the best possible evidence that the AI people are not anywhere near understanding what "intelligence" is or how to model it: they should have been able to produce a winning computer in 1975, they couldn't, and they still don't know why.

And that doesn't even start to touch the fact that with the same brain that plays chess Kasparov can write political rhetoric, drive an automobile, and woo romantic partners. Is there any chess playing computer that can do even one of the hundreds of other functions the silliest human brain can handle? It is easy to build a special purpose machine like a locomotive; it is very hard to build a general purpose machine like a horse - not to mention human intelligence.

Cranky

[jm]

Chess is not a meaningful test of intelligence.

A more meaningful (though still limited) test is optical character recognition. Feed a marginal photocopy a person can read with no trouble at all into any OCR package available today and you'll get nothing useful out. Same if you feed in even a perfect image of text in a highly stylized font. The computing power of the human brain in pattern recognition tasks is enormous.

[Matt Davis]

The thing about *human* intelligence is that the brain can analogize, and thus import existing knowledge into new circumstances. That's not what Deep Blue is doing; it's got an enormous capacity to generate potential outcome scenarios, and it picks the best one.

Kasparov and Fischer could sit down and play Fischer's Bizarro Chess derivative, because Kasparov could intuit why Fischer moved all the pieces around, and could develop new--if awkward--principles of how to play the new game. Deep Blue would be forced to play it as chess.

[JO'N]

Actually, Minsky's completely wrong here. The real problems in present-day AI are precisely those that Minsky's complaining about: trying to get a computer to be as intelligent as an insect. We seem to be getting close to that now.

I think most AI researchers will tell you that if someone gave them an AI as intelligent as a rat, they could make a human-equivalent, Turing-test-passing AI in a very short period of time. If and when we know how to solve the set of problems that a rat can solve -- and we're probably many years away from that -- then people like Marvin will be there to tell us how to take the last step.

[Tom Cecere]

Machine Intelligence is different from human intelligence because they have evolved in enormously different environments. The "problem" with AI is that mimicking (sp?) human intelligence is the primary goal; since they can't get that to work for anything useful, other scientists rightfully dismiss them. This makes it hard to get grants and tenure.

While they're wasting careers on AI, the world is watching machine intelligence evolve. Self-fixing and self-replication are a generation away; size and power consumption are shrinking; cross-communication is growing. But the environmental threat to machines is that people won't find them useful, so only the useful survive.

Once all the pieces are in existence, we could probably cobble together the same set of brain modules that Pinker and others describe, to come up with a bad copy of a human. That would be an interesting novelty, but is really just a side show -- not useful. So it will only be a museum exhibit.

[dogfacegeorge]

I'll believe in AI when I see a computer win the World Series of Poker.

[Joe O]

I'll believe in AI when a computer is doing my job.

[Allan Connery]

Has Minsky moved some goalposts himself?

As I remember it, there was a subtext to the 1960s-era claim that a computer could become a chess master. It went unsaid that of course a chess-master computer could do all the other things that a human chess master can, like understand language, make complex plans and carry on a conversation. Chess was merely an example of how far beyond ordinary mortals a computer would be.

HAL in 2001: A Space Odyssey (circa 1967) was just such a computer.

In real time, computers can play chess, all right, but we're still waiting for Longhorn, never mind HAL.

[felixrayman]

>Even the computers of the 1960s, and certainly the 1970s/80s
>had a clock rate and calculation speed as far beyond the human
>brain as a GM SD90MAC locomotive is beyond human muscle. Yet it
>wasn't until the late 1990s that anyone built a chess computer
>that could regularly beat a human grandmaster

You are doing exactly what the article says, moving the goalposts. The original claim was that computers would not ever be able to play chess. Not that they wouldn't be able to win or to beat grandmasters, just that computers would not ever be able to play the game of chess. Once they could play, the claim was that they wouldn't be able to beat people. Not grandmasters, just ordinary people. Once they did that, the claim was that they would never be able to beat grandmasters. Not beat them all the time, just beat one. Once they did that, the claim is that the fact that grandmasters might still have a chance against AI and the fact that it took a few decades to develop the hardware and software to beat grandmasters is offered as a criticism of AI.

This is definitely moving the goalposts. Which is fine, it helps the development of the art, but you should be aware of what you are doing.

>And that doesn't even start to touch the fact that with the same
>brain that plays chess Kasparov can write political rhetoric,
>drive an automobile, and woo romantic partners. Is there any
>chess playing computer that can do even one of the hundreds of
>other functions the silliest human brain can handle?

You can run anything written in C on the computer that first beat Kasparov in a game in 1997. So the answer to your question is "yes". More importantly, the software that beat Kasparov could fairly easily be ported to any machine capable of compiling C code, so any computer as fast as the one it originally ran on can quickly be taught to play chess as well as the original machine. Can Kasparov immediately show anyone as intelligent as himself how to play chess as well as he can?

Critics of AI, the goalpost moving ones, seem to define intelligence as "that which a computer can not yet do". Which is fine too, but it means that the realm of intelligent behavior will be a constantly shrinking one. Personally I think that if a computer exhibits behavior that you would see as a sign of intelligence were your child to exhibit it, that is artificial intelligence.

[Cranky Observer]

> You are doing exactly what the article says,
> moving the goalposts.

Felix,
I was there at the time, in the 1970s and 1980s. I read quite a bit of the original literature. I can assure you I am _not_ moving the goalposts.

Cranky

[paul]

Minsky made the same point about moving the goalposts at a speech in about 1985, at one of the international conferences in the heyday of AI. There he made the more succinct analysis that "intelligence is what we admire but do not (yet) understand how to program." There's really no question that each bit of ostensibly intelligent behavior, from chess to medical diagnosis to logistal planning to image recognition to language translation, becomes "oh, that doesn't really require intelligence" as soon as somebody figures out how to make a machine do it.

What this calls into question, of course (as Minsky also pointed out almost 20 years ago) is what our intelligence really consists of, and what might constitute a dividing line between human (or human-style) intelligence and all the supposedly-intelligence-requiring behaviors that people will eventually figure out how to program. Preferably a line that most humans don't end up on the wrong side of.

It's interesting to see this rehashed at such a late date -- perhaps just as it's agreed that each major new generation of computer hardware sets software engineering back 10 years, so each new major mode of information dissemination sets the conventional wisdom back 20 years.

[Warren F]

"It was said that computers will never beat people at chess, because that requires intelligence and computers aren't capable of intelligent thought"

I worked as a junior programmer in an AI lab at Carnegie Tech in the 1960's, and nobody said this. This is baloney. Most people in the lab said "The machine will eventually beat the human". And most people outside the lab just asked "When?", not doubting the eventual outcome at all.

Minsky is defending his turf, as is his right, but his revisionist history is a crutch he doesn't need.

It is a weakness of the AI field that we still don't have a scientific agreed-upon definition of thinking. Our word usage is inconsistent, we say airplanes fly but we don't say submarines swim. So without a definition of thinking, it doesn't matter if we say computers think, except at budget time, when we hype things out the wazoo, anyway.

[felixrayman]

>It is a weakness of the AI field that we still don't have a
>scientific agreed-upon definition of thinking. Our word usage is
>inconsistent

This is a weakness of...AI?

[fifi]

Chess computers do well in over-the-board play (2hrs/40 moves, touch-piece rules, etc.) but fare very poorly at postal chess, a version of the game in which accurate calculation of variations and visualization of the board are less relevant. Chess is a game of perfect information and calculation, so we would expect computers to excel at it, perhaps eventually solve it, even, but right now they are quite stupid at the game – they don’t “understand” it -- and any half-way competent amateur can beat them by taking moves back. By analogy, few people would consider an HP calculator to be an “intelligent” mathematician because it can add up numbers real quick-like without making stupid mistakes.

[fifi]

Why aren't there Artificial Energy or Artificial Force branches of philosophy?

[Movie Guy]

In 1988, I had the good fortunate to beat a Finnish chess master (a PhD specializing in fourth dimensional mathematics - his explanation) while visiting Bergen, Norway. It made a friend of mine extremely mad as he was an accomplished ranked chess player, and had lost three games in succession to the Finnish master. I mean the guy wouldn't speak to me. For days.

Yes, it's true that the Finnish master had consumed far more alcohol than I. Further, he was really taken with the lovely blonde who was watching the game. Of course, having won the match, I was fortunate enough to date her. For a few years, actually.

Two weeks later, I lost a game to a German chess computer hooked up to an expensive beautiful wood set designed for the computer which I had purchased for my German landlord, the lead constitutional attorney for the nation of West Germany. A brilliant fellow, that one. He loved that chess set.

I've always assumed that the amount of fine alcohol that evening may have contributed to my defeat by the chess computer. I don't recall that computer ever having anything to drink... Maybe, but I can't be sure.

[PaulC]

The goal posts are moving? If so, I think Minsky is mistaken about the direction. Let's see where Turing set the goal posts in 1950:

http://www.cs.swarthmore.edu/~dylan/Turing.html

"I believe that in about fifty years' time it will be possible, to programme computers, with a storage capacity of about 10^9, to make them play the imitation game so well that an average interrogator will not have more than 70 per cent chance of making the right identification after five minutes of questioning."

I corrected "109" from the link to 10^9, roughly what we'd now call a gigabyte. It's an uncannily accurate prediction, by the way. He might not have realized that there would be many millions of inexpensive computers with that kind of capacity, though.

But on the subject of intelligence, he set a very ambitious goal of having a computer simulate human intelligence by fooling a human in conversation. The goals have been scaled back since then. It's true that you could trick a lot of people for 5 minutes, so in some sense, the overall prediction might be accurate. I think a trained examiner should be able to distinguish between a computer and a human faster than that. I can only guess without actually running such a test.

We're only just coming over some humps that might lead to Turing-test capable AI. For years, the circuitry was just far smaller than the brain. This may still be true for individual computers, but I think it won't be for long. There was also a lack of readily available domain knowledge for some of the most compelling projects, such as natural language comprehension. The Internet has suddenly made huge quantities of text available as well as pictures and multimedia. It wouldn't surprise me to see AI advances coming out as a byproduct of some of Google's work, for instance. The have the people, the computing power, and a world's worth of data.

When Turing made his prediction, he probably picked a storage size of a billion because it seemed unfathomably large for a computer of the day. He had no idea what it would take for a computer to be smart, was probably influenced by a tabula rasa model of the human brain, and conflated intelligence with logical reasoning from axioms without considering the importance of spatial and sensory inputs.

I think I could see AI in my lifetime, but there are a lot of uncertainties. For every goalpost moved away, a number of other goals have been dropped as too ambitious. Language translation was initially thought to be within grasp in the 60s. It is probably getting closer, but it's also now understood that completely accurate translation is as hard as language comprehension.

If Minsky is saying that people ought to have more faith in AI researchers, I think he's blowing smoke. The average person recognizes that computers do amazing things, and keep getting better. I don't think people are skeptical that a computer could be intelligent; they're just waiting to see it. AI research has been making steady advances over the decades, but if anything reduces their credibility, it has been the tendency to hype the significance of toy successes.

[Auros]

Tom: the usefulness of modular understanding of the human brain shows up when you happen to be a human who needs brain surgery, and we need to know where different modules reside and how messing them up might affect you. Having a machine copy -- even a bad one -- that can easily be messed up and repaired, could, in the long run, be tremendously valuable. And this is leaving aside the potential to eventually design brain-enhancing implants. (I know _I_ want a math coprocessor, and some offboard memory for my schedule and addressbook...)

[Tom Cecere]

Auros - Diagnostic models are a terrific application, and I suppose one would need to gather them together to watch interactions. Good point! This would then be a "useful" development, and would be continued; chess playing for chess playing's sake is certainly not.

[RSA]

I think that there are a few versions of the "moving the goalposts" issue that are being conflated in the original article and in this discussion.

It's clear that Minsky thinks AI has abandoned to a large extent its original goal of building agents with general intelligence. He might say that this is a kind of goalpost moving, in that we count as successes systems that are in most ways pretty unintelligent. I've heard some senior AI researchers comment about a lack of vision in recent AI research, but it's certainly open to debate.

Another kind of goalpost moving involves the recognition that AI is much harder than we originally thought. As another AI founder, Robin Popplestone, once remarked, "We used to think that playing chess was hard and playing football was easy. Now we know it's the reverse." It's only reasonable to set up new goals that are more likely to be achieved in the near term, if the original task is found to be too hard.

Finally there's the man-on-the-street perception issue. Martha Pollack, in the end of the Wired article, has it right, that stuff that was initially core research in AI has turned into commodity software, and thus no longer AI. Work on expert systems is inappropriately dismissed, despite results being incorporated into tax preparation systems and other everyday software. Bayesian reasoning is carried out in your Windows Help system, and so it's obviously not AI. Neural networks in elevator controllers and washing machines? Also not AI. Game-playing agents that do path planning and other tasks originally considered intelligent behavior? Not AI. I think that even if someone were to build a system that could pass the Turing Test without fail, for 15 minutes at a time, someone would say, "A *real* AI system would be able to go for at least 16 minutes."

That said, AI remains a thriving field.

[BillCross]

Mark Tilden at Los Alamos National Labs has some interesting ideas on robotics/artificial intelligence based around analog rather than digital design.

here is an interview

http://www.exhibitresearch.com/tilden/

and gere a proposed documentary (with trailer)

http://www.robosapiens.org/Robosapiens.html

[RT]

I don't see that it's an instance of goalpost-moving; it's an instance of some people being wrong in a limited instance.

I'm old enough to remember when the best computers couldn't beat a 1500-rated chessplayer. The claim that computers would eventually be able to beat the champs was agreed with by some, disagreed with by others, and many of us just didn't care.

But whichever camp you were in, this wasn't thought to have anything to do with AI; it was about computing power and/or programming. Nobody claimed that the computer's eventual victory would represent 'thought' or 'intelligence' as we understand it.

Computers can beat the best humans at chess played within a certain timeframe. But whether one was right or wrong about its eventual occurence, it has zilch to do with questions of AI.

[Rob Sperry]

"As there is no widely accepted definition for intelligence, this is often an example of the No true Scotsman fallacy"

While its not widely accepted, here is a formal definition of intelligence

http://www.vetta.org/documents/42.pdf

The paper starts with:
"Intelligence measures an agent’s general ability to
achieve goals in a wide range of environments."

And then adds maths.

[Alex]

In a sense, this isn't a problem: at every step, we've created things that are closer to intelligence without achieving it yet. There's obviously some irreducible X-factor that constitutes intelligence, and as we go we're cutting the cruft out of the definition. In fact, there's a good logical case that intelligence ought to be defined as that property of living creatures that would distinguish them even from a self-replicating machine; so yes, intelligence is what the computer can't do yet.

[John Aspinall]

Alex says: "There's obviously some irreducible X-factor that constitutes intelligence..."

It's worth pointing out the opposing viewpoint, to wit: "It's algorithms all the way down." The point about many of the previously cited examples of skills/tasks, the performance of which used to (but not longer) imply intelligence, is that it's the appearance of an algorithm that provokes the rejection of that task as an intelligence test.

That says more about human xenophobia than it does about the nature of implementation-independent intelligence. If you disagree here's a challenge: conjure up an example (probably science-fiction-y) of something intelligent but not living.

[Alex]

Not necessarily. If it's algorithms all the way down, then presumably one of them would finally tip the scale and generate undeniable intelligence (say, passing the Turing test) - and the difference between that and the irreducible X-factor is a matter of wordplay. My point is that this may just be an example of eliminating the options until one arrives at the answer.

Something intelligent but not-living? Well, the obvious example is a highly advanced computer. I tried, but I can't imagine nonliving intelligence in nature in any remotely plausible fashion, chiefly because nonlivingness means no replication with mutation and natural selection, and hence no plausible route to complexity without design.

The question is how we would know the computer was intelligent when it got there - which is another way of wondering how to define intelligence. I would offer that genuine intelligence isn't task-dependent. Sure, the computer plays a mean chess game, but the chess program is shit at word processing.

[PaulC]

> Alex says: "There's obviously some irreducible X-factor that constitutes intelligence..."

> It's worth pointing out the opposing viewpoint, to wit: "It's algorithms all the way down." The point about many of the previously cited examples of skills/tasks

My understanding of "irreducible X-factor" in the above context was some emergent property of "algorithms all the way down". Obviously, an intelligent computer is going be running algorithms--in the sense the M. Jourdain had been speaking in prose all his life without knowing it--but trying to understand intelligence in terms of algorithms may be as elusive as understanding all of biology in terms of quantum mechanics.

I think that intelligence may turn out to be an aggregate of algorithms that do not have strictly bounded running times nor even guaranteed correctness, but just happen to increase the probability of making the correct decision in many circumstances. We may find ourselves slapping one poorly understood algorithm on top of another until the computer is clearly intelligent, but we don't know how we got there. When we ask what part of it makes it intelligent, there may be no clear answer. (Of course, we might build an even more intelligent machine that could understand it better than we do.)

I've been struck by a number of amazing things that computers can do with relatively simple algorithms. One is the use of brute force search to find solutions to sets of constraints (particularly in the search for Game of Life oscillators, but it's a more general phenomena). The solutions can be so ingenious that they appear to be the product of a creative mind. It's most fruitful to think of this process in declarative terms: in go the constraints, out come the solutions that satisfy them. A focus on the search algorithm does not give you much understanding, though it may be a challenging task to optimize it. If you imagine a system filled with many steps of this sort, it is reasonable to suspect that the outcome is so far removed from the designer's understanding of how it works that there is an "X-factor".

Another case more closely tied to human intelligence is the uncanny accuracy of modern web search algorithms. Of course, search can sometimes be frustrating, but more often I find that I can get to a piece of long neglected trivia--say about some TV show from my childhood--just by free associating into a search engine. I suspect that you could design a simple Markov process to search iteratively--feeding back keywords culled from the last search--that came up with truly remarkable connections of the sort usually associating with letting the mind wander over memories. This would be much more interesting than many classic AI hacks (such as the Markov-chain prose) because it would be drawing on a much richer domain.

[Ian Montgomerie]

"Baloney. Even the computers of the 1960s, and certainly the 1970s/80s, had a clock rate and calculation speed as far beyond the human brain as a GM SD90MAC locomotive is beyond human muscle."

Best estimate I've seen is that the processing power of the human brain is still something like five or six orders of magnitude greater than that of the most powerful supercomputer ever created. We just happen to be slower at arithmetic because computers use that as their fundamental operation on top of which everything else is built, and humans don't.

[RSA]

". . .trying to understand intelligence in terms of algorithms may be as elusive as understanding all of biology in terms of quantum mechanics."

See Newell's article/book, The Knowledge Level, for one early treatment of this issue. Just as computer programmers using high-level languages don't (usually) worry about register allocation, AI researchers needn't remain at low levels of abstraction in working toward intelligent behavior.

[paul]

Not that anyone cares any more, but the original Turing test was about gender: five minutes with a teletype connection to decide whether your interlocutor was a man pretending to be a woman or vice versa, or one of the real things. And the prediction was that a computer program would be approximately as good as a human at making people guess wrong. The version involving far-extended conversations and a judge trying to figure out whether the interlocutor was human or computer (which some programs have also managed roughly as well as humans) got tacked on later. For the first version, bots in chat rooms do just fine hundreds of times every day.

Just by the way, there's also a moral dimension to the Turing test, especially given the then-much-more-recent history of "advanced" peoples believing in the existence of lesser human races who, although they might for a time provide a more-or-less convincing imitation of a white man, were ultimately lacking in the spark necessary for intelligence and civilization.