Kasparov on the Significance of AI Playing Jeopardy!

Over at The Atlantic they have an interesting little exclusive with Garry Kasparov (who has a little experience with the man vs. machine thing) providing his reactions to IBM supercomputer Watson's victory in Jeopardy!:

• A convincing victory under strict parameters, and if we stay within those limits Watson can be seen as an incremental advance in how well machines understand human language. But if you put the questions from the show into Google, you also get good answers, even better ones if you simplify the questions. To me, this means Watson is doing good job of breaking language down into points of data it can mine very quickly, and that it does it slightly better than Google does against the entire Internet.

• Much like how computers play chess, reducing the algorithm into "crunchable" elements can simulate the way humans do things in the result even though the computer's method is entirely different. If the result—the chess move, the Jeopardy answer—is all that matters, it's a success. If how the result is achieved matters more, I'm not so sure. For example, Deep Blue had no real impact on chess or science despite the hype surrounding its sporting achievement in defeating me. If Watson's skills can be translated into something useful, something groundbreaking, that is the test. If all it can do is beat humans on a game show Watson is just a passing entertainment akin to the wind-up automata of the 18th century.

• My concern about its utility, and I read they would like it to answer medical questions, is that Watson's performance reminded me of chess computers. They play fantastically well in maybe 90% of positions, but there is a selection of positions they do not understand at all. Worse, by definition they do not understand what they do not understand and so cannot avoid them. A strong human Jeopardy! player, or a human doctor, may get the answer wrong, but he is unlikely to make a huge blunder or category error—at least not without being aware of his own doubts. We are also good at judging our own level of certainty. A computer can simulate this by an artificial confidence measurement, but I would not like to be the patient who discovers the medical equivalent of answering "Toronto" in the "US Cities" category, as Watson did.

• I would not like to downplay the Watson team's achievement, because clearly they did something most did not yet believe possible. And IBM can be lauded for these experiments. I would only like to wait and see if there is anything for Watson beyond Jeopardy!. These contests attract the popular imagination, but it is possible that by defining the goals so narrowly they are aiming too low and thereby limit the possibilities of their creations.

From all I can tell, the only really interesting technological advances indicated by Watson's victory are in the ability of a computer to discern a spoken question and it's approach to trying to ascertain the meaning of that question. In both cases, Watson's approach is utterly alien to human intelligence, and in many ways just underscores how different human thought and what a computer does are.

A Conversation with Nobody

Back in 1950 Alan Turing wrote a paper entitled "Computing Machinery and Intelligence" in which he proposed a test (which has been since named the "Turing Test") to determine if an artificial intelligence had been successfully created. The basic idea is that if an artificial intelligence is created such that person A can have a conversation (via typed text) with persons B and C, one of whom is another human sitting at a keyboard and one of whom is a computer program, such that A is not able to tell which of the two entities he's talking to is a human and which is a computer, then this would be evidence that the computer is an artificial intelligence to the extent that it can functionally behave like a human.

Of course, it's a bit flattering to think that the way to determine if a machine is a "thinking machine" is by seeing if it can interact with us humans in a social setting. On the flip side, some levels of human interaction are so low level that one can imagine a fairly simple compluter program succeeding at them pretty well. For instance, I would imagine you could probaby be moderately successful in creating a computer program that could present a reasonable facsimile of human communication via Facebook statuses, links and comments. But that's in part because communication on Facebook is pretty simplistic and you can always not respond or throw out non sequitors without seeming like anything other than a fairly normal Facebook user.

Building a computer program capable of producing a passable simulation of a more wide-ranging conversation is, however, a lot trickier, and such a feat is clearly some ways off in the future.

Thinking about this, however, in relation to the question of whole brain emulation which I wrote about last week it occurs to me that there are two questions here:

1) Is it possible to create a computer program which could fool a human into thinking that he is having a conversation with another human.

2) Is it possible to create a computer program which is actually capable of having a conversation for its own sake.

The first of these is achievable if one can build a good enough set of algorithms around how humans respond to questions, common knowledge for conversational grist, etc. The second, however, is much trickier. When we converse with someone we generally do so because we want to communicate something to that person and/or we want to know something about that person. In other words, conversation is essentially relational. You want to know about the person you are talking with, and you want them to know about you. You want to establish areas of common interest, experience and belief. You want to bring that person to share certain ideas about you or about the world.

I'm not sure that it would ever be possible to build a computer program which had these feelings and desires. Oh, sure, you could give it a basic like/not-like function where it tries to achieve commonalities or confidences and if it is rebuffed puts its interlocutor in the "not like" category and relates to it differently. But this is very different from actually wanting to know about someone and wanting to like and be liked by them. How our own human emotions work in this regard is far from clear to us, so I can't imagine that we're in any prosition to understand them so well that we can create copies in others.

Of course, in real conversation you often can't tell if the other person you're talking with, even face to face, is actually interested in you or actually likes you. This question is a source of considerable concern in human interactions. So perhaps the question of whether a computer can care about who you are or what you talk about is irrelevant to the question of whether a computer could be designed that could pass the Turing Test. But I do think that the question is probably quite relevant to whether a computer could ever be a person.

The Materialism of Limited Toolset

I make a point of always trying to listed on the EconTalk podcast each week -- a venue in which George Mason University economics professor Russ Roberts conducts a roughly hour-long interview with an author or academic about some topic related to economics. A couple weeks ago, the guest was Robin Hanson, also an economics professor at GMU, who was talking about the "technological singularity" which could result from perfecting the technique of "porting" copies of humans into computers. Usually the topic is much more down-to-earth, but these kinds of speculations can be interesting to play with, and there were a couple of things which really struck me listening to the interview with Hanson, which ran to some 90 minutes.

Hanson's basic contention is that the next big technological leap that will change the face of the world economy will be the ability to create a working copy of a human by "porting" that person's brain into a computer. He argues that this could come much sooner than the ability to create an "artificial intelligence" from scratch, because it doesn't require knowing how intelligence works -- you simply create an emulation program on a really powerful computer, and then do a scan of the brain which picks up the current state of every part of it and how those parts interact. (There's a wikipedia article on the concept, called "whole brain emulation" here.) Hanson thinks this would create an effectively unlimited supply of what are, functionally, human beings, though they may look like computer programs or robots, and that this would fundamentally change the economy by creating an effectively infinite supply of labor.

Let's leave all that aside for a moment, because what fascinates me here is something which Roberts, a practicing Jew, homed in on right away: Why should we believe that the sum and total of what you can physically scan in the brain is all there is to know about a person? Why shouldn't we think that there's something else to the "mind" than just the parts of the brain and their current state? Couldn't there be some kind of will which is not materially detectable and is what is causing the brain to act the way it is?

(Or to use the cyber-punk terminology which seems more appropriate with this topic: How do we know there's not a ghost in the machine?)

Hanson's answer is as follows (this section starts around minute 32 of the podcast):

"I have a physics background, and by the time that you're done with physics that should be well knocked into you, that, you know, certainly most top scientists, if you ask them a survey question will say, 'Yeah, that's it.' There really isn't room for much else. Sorry. It's not like it's an open question here. Physics has a pretty complete picture of what's in the world around us. We've probed every nook and cranny, and we only ever keep finding the same damn stuff.

We have enormous progress on seeing the stuff our world is made of. Almost everything around you is the same atoms, the same protons, electrons, the rare neutrino that flies around. And that's pretty much it. You have to get pretty far off to see some of the strange materials and things that physicists sometimes probe. Physicists have to make these enormous machines and create these very alien environments in order to find new stuff to study because they've so well studied the material around us. The things our world is made out of are really, really well established. How it combines together in interesting ways gets complicated and then we don't get it, but the stuff that it's made out of, we get.

Your head is made out of chemicals. We've never seen anything else. It's always theoretically possible that when something's really complicated and you don't know how to predict the complexity from the parts, you could say, 'Well therefore, it could be this whole is different from the parts, because it's too difficult to predict.'
...
We should separate two very different issues here. One is technological understanding and knowing how things work and how to make things, and the other is knowing what the world is made of. So, I make this very strong and confident claim: We know what the world is made of, and we know what pieces they are and how they interact at a fine grain. But at higher levels of organization, we don't know how to make other things like, even, photosynthesis in cells. We don't know how to make a photosynthesis machine. You could take your cell phone out of your pocket and take it apart and you wouldn't know how to make a phone like that.... We don't know how it works, but we're pretty sure what it's made out of."

Now, this line of thinking seems fairly familiar to me from talking with materialist/atheists of a scientific bent: We have all these great scientific tools, and all they've ever detected is matter and energy, never a "will" or a "beautiful" or a "soul", and so therefore it's pretty clear that when we talk about our minds we're really talking about our brains and there just isn't anything there except chemicals and electricity.

However, it seems to me that this presents a rather obvious blind spot. We, as human persons, experience all sorts of things which would seem to be evidence of having a will which decides things in a non-deterministic fashion. We also respond to ideas such as "beautiful" or "justice" or "good" in ways that would suggest that there is something there that we're talking about.

When we say, "Physicists have done all this work, and all they've ever found is matter and energy," you are really saying, "Given the tools and methodology physicists use, all they are able to detect is matter and energy." But I'm not clear how getting from that to, "Therefore there is nothing other than matter and energy," is anything other than an assumption.

Is there any valid reason why we should accept the jump from, "Tools that scientists use to detect things can only detect the existence of material things," to "Only material things exist"?

This seems particularly troublesome given that the project here is supposedly to create an emulation program which can be given a brain scan and then act like an independent human. If our experience of being human is that there is something in the driver seat, something which decides what is beautiful or what is right or who to marry or whether we want rice pudding for lunch today, then unless there is some active, non-deterministic thing within the brain which can be measured by this scan, then what you get is going to be, for lack of a better word, dead.