Tell YC: Answers from John McCarthy
-----Professor's Interview Questions-----
Q. Can Computers Think?
A. Thinking isn't one thing. It has many aspects. For example, computers have the ability to remember information and the ability to play games. Some aspects of thinking, we have not succeeded in. A notable examples is the analysis of situations. A computer cannot break a situation into parts, analyze the parts separately, and then combine the parts to come to a conclusion. A specific manifestation of this is the game "Go". This type of thinking is necessary in "Go", where it is not in Chess. This is why the best computers are as good as people in Chess, but the best computers are much worse than people in "Go".
Q. Is there anything in principle that would prevent a computer from thinking as a human would?
A. No
Q. Can computers know?
A. This is largely a question of definition. If a camera looked at a table, we could say it "knows" that there are four containers of liquid on the table (which was true).
Q. Is there any definition of "know" in which computers cannot succeed?
A. Well, I suppose the biblical sense.
Q. Ha, well, what makes you think that?
A. They don't satisfy the necessary axioms (laughter)
Q. OK, can a computer have free will?
A. In my paper over free will, I defined "simple deterministic free will," which a computer can have. In fact, modern chess playing computers have this. However, this is not always true for displays of artificial intelligence. Consider two optimal tic-tac-toe playing programs. The first evalutes future situations in order to choose the optimal solution. The other simply looks at the state of the board, for which there are only 3^9 possibilities, and picks a move from a lookup table. The first program exhibits simple deterministic free will, where the second program does not. A chess program cannot have a lookup table because the state is too complex. Thus quantitative considerations are important. Philosophers would have you believe that they are not. That a chess problem and a tic tac toe problem are equivalent. I believe quantitative considerations are important.
Q. Simple deterministic free will does not require that a computer know that it has free will. How would a computer know that it has free will?
A. Well, computers are good at understanding theories. My theory of simple deterministic free will is a theory. You could teach it this theory.
Q. Are there some senses of free will that aren't simple deterministic?
A. (I didn't catch the first part of his response) Some believe that free will is acheived through random aspects of quantum mechanics. This is particularly attractive to people who don't understand quantum mechanics.
Q. Can computers achieve consciousness?
A. Human consciousness starts with being aware of basic things such as hunger. Advanced states of consciousness are simply more elaborate forms of these basic awarenesses. We have a surprisingly limited ability to examine our own state. We ought to remember what we've had for breakfast for the past 30 days, but we can't. Short answer -> yes, machines can have consciousness.
-----Student Questions-----
Q. Why would we want to give computer's emotions?
A. Human emotion involves the state of the blood, and this is inherited from our animal ancestors. Giving a computer this kind of emotion, or "state of the blood", would not be to our advantage.
Q. (Something that led him to talk about his new language Elephant)
A. Elephant was meant to come out in 2005, but 2005 has come and gone and the language isn't ready yet. It is a new way to talk to computers. I/O is done through speech acts. (He said something about the programming language dealing in obligations and promises, and I'm not sure what that means)
My Q: While we're on the same to...
75 comments
[ 4.7 ms ] story [ 130 ms ] threadFor me , this pretty much sums it up--thank you for posting ;)
A. If I hadn't come up with it, someone else would have. Pure Lisp was a discovery, everything that has been done with it since has been an invention.
This brought back thoughts about Gladwell's recent essay which was posted here.
Aston - email me at info@reatlas.com please.
That's too bad. I always noticed the distinction between professors who welcomed interaction with students and those who shooed students away. Stanford has a lot of the latter. (Update: my assumption is that this is truer in general of top-tier schools than average ones, probably because smart, keen students stand out more at average schools and so get more attention. Of course, average schools have fewer smart, keen teachers too. But if you connect with the right one, the experience can be life-changing.)
"comfort level being outgoing" * "likelihood of doing research for fun" + "comfort level being outgoing" * "confusion over the days lecture"
this plus a little bit of reflection on the distribution of socialization styles amongst smart folks is a fairly complete model.
Unfortunate. Next time stand your ground. You might not get another chance to meet someone like McCarthy.
This man is now my personal hero.
Suddenly I understand why so many bright people were drawn to work with McCarthy. :)
This might have been true in the days before anyone had the concept of macroscopic decoherence (popularly known as many-worlds) but quantum physics is pretty normal these days. See the recent series at Overcoming Bias.
[This is not to say that those researchers are unintelligent, just that a certain bias may exist.]
http://www.overcomingbias.com/2008/04/on-being-decohe.html
Of course, I have a Ph.D. in semiconductor lasers, so I've been trying to think about such stuff off and on for years... which means I'm a terrible test case for articles like this. ;)
?
I'm game if you are, though. You might start by thinking about a modern digital computer. It's made up of a very large number of very small devices (transistors) that sling populations of electrons back and forth. The brain is made of a much larger number of less-small devices called neurons, that sling bigger populations of bigger objects (organic molecules) back and forth. So according to the basic understanding of QM that tells us relative indeterminacy increases when considering smaller objects, the functioning of the human brain is marginally more predictable than the insides of a modern computer.
People get confused because it's easy to look at a quantum mechanics problem from the wrong angle and see nondeterminism. We're really well trained in intuitive mechanics ("an electron is like a tiny tennis ball, and tennis balls are always either here or there, right?") so at first glance quantum mechanics seems wacky and random: the electron might be here, or it might be there, with equal probability, and we can't tell which! Whereupon your head explodes. Seriously: The discoverers of quantum exploded in horror, and they started ranting about that crazy cat in the box, or "God playing dice". (Colorful but misleading metaphors. One reason that Bohr and Einstein's early confusion persists today is that they were just so darned eloquent.)
In fact, God does not play dice: God sees all the outcomes of the dice roll at the same time and doesn't understand why we think it's a game, and not a static work of art.
People also think that nondeterminism is somehow an important ingredient in whatever it is we mean by "free will". This doesn't make much sense to me. If you want it to not make sense to you as well, read Daniel Dennett's Freedom Evolves. You might want to budget more than a couple hours for that book, though.
Edit. Nevermind, I see what you mean.
From [http://www.fourmilab.ch/hotbits/how3.html]
"But hidden variables aren't the way our universe works—it really is random, right down to its gnarly, subatomic roots. In 1964, the physicist John Bell proved a theorem which showed hidden variable (little clock in the nucleus) theories inconsistent with the foundations of quantum mechanics. In 1982, Alain Aspect and his colleagues performed an experiment to test Bell's theoretical result and discovered, to nobody's surprise, that the predictions of quantum theory were correct: the randomness is inherent—not due to limitations in our ability to make measurements. So, given a Cæsium-137 nucleus, there is no way whatsoever to predict when it will decay. If we have a large number of them, we can be confident half will decay in 30.17 years; but if we have a single atom, pinned in a laser ion trap, all we can say is that is there's even odds it will decay sometime in the next 30.17 years, but as to precisely when we're fundamentally quantum clueless. The only way to know when a given Cæsium-137 nucleus decays is after the fact—by detecting the ejecta. A Cæsium-137 nucleus which has “beat the reaper” by surviving a century, during which time only one in a thousand of its litter-mates haven't taken the plunge and turned into Barium, has precisely the same chance of surviving another hundred years as a newly-minted Cæsium-137, fresh from the reactor core."
The short answer is that in the many-worlds interpretation, this too is a deterministic process. The total wavefunction of the whole system (atom + observer) evolves deterministically. What isn't deterministic is "your" subjective view of it, but "you" only view a vanishingly small slice of reality.
Sorry, that's the best three-sentence explanation I can come up with right now, and I admit it's only a shade better than "trust me, I'm a physicist". But trust me, I'm a physicist.
I also think he's wrong, but at least he's not infatuated about it.
edit: I love his books. Well worth the read, and for those concerned he puts his speculations in chapters very clearly marked "speculations". From his books I first learned the details about Turing machines and lambda calculus, and a lot more.
It all hangs, not on quantum mechanics, but on his initial assertion. I've read his explanation in Shadows of the Mind, but the lightbulb didn't go on for me. For McArthy to dismiss the question offhand was a little disappointing. I'd expect him to have a deeper insight.
[1] Spivey, M., Grosjean, M. & Knoblich, G. (2005). Continuous attraction toward phonological competitors. Proceedings of the National Academy of Sciences, 102(29), 10393-10398.
Q. Is there anything in principle that would prevent a computer from thinking as a human would?
A. No
IOW, there's still no recognition today, on the side of the purveyors of "classical" AI, that anything except digital processing might be needed for a computer to think "as a human would". So the big money is likely to continue being thrown at attempts to emulate animal brains using purely digital means. And I suspect that these funds might largely be better spent elsewhere.
What the quantum intelligence does is take one unresolved problem (how do we think) and replace it with another (we think with quantum computing, but we don't know exactly how). Its only result is to take away the unknown and replace it with an incomplete explanation.
I also don't think Penrose would have come to this conclusion now. Cognitive psychology has taken enormous strides in the past years, and it's already pretty clear it's on the right track. Take a look for example at "The Emotion Machine" by Marvin Minsky (who by the way is on par with McCarthy in AI but chose cognitive psychology as a main field: http://en.wikipedia.org/wiki/Marvin_Minsky).
It's not as if there aren't many, many great physicists who have failed to understand quantum mechanics. Max Planck invented the field but never understood it. And Einstein, if he did indeed understand it, never stopped wishing that he didn't have to.
Anyway, I haven't read Penrose's arguments -- geez, I barely take time to read credible science these days -- so I'm not dismissing him: I'm outsourcing my dismissal of him to guys like Dennett. Meanwhile, I'll happily accept your assertion that Penrose understands quantum mechanics perfectly well even if he enjoys abusing it for philosophical kicks, kind of like how the screenwriters who wrote Gladiator claim to have actually read real Roman history before they reinvented it for the screen.
From now on I'm outsourcing all my dismissiveness to Daniel Dennett. I just don't have the time to be properly dismissive, and he's so much better at it than I am anyway.
But that's like the 10th time hard-to-read-text has happened to me today, and it gives me an idea.
On-the-fly, stored in the cloud site specific CSS customizations would be a killer feature for a Firefox extension. Something that useful must already be in Firebug ;) /goes to check
Also forgive me about wanting to be a better programmer, especially here on hacker news. What was I thinking.
I wouldn't mind AI questions too, if only they were a bit more original.
He pointed that out in the answer.
(That's an interesting question: what makes a table a table? Not all tables are made out of the same material or have the same color. Not all tables have four legs (or legs at all!) and not all have a flat surface. Not all tables are the same height or width or are used for the same purposes. What, then, makes some particular table "a table"?)
We don't say that a camera "knows" there's three glasses of water on the table when it takes a picture any more than we say a newborn baby "knows" when he looks at a chessboard that Kasparov has a mate in three.
I guess I shouldn't take McCarthy's comments at a freshman seminar as representative of his most thorough theories on AI, but I found that one answer particularly naive.
But I also think there are some very important ideas from the more recent work that began with connectionism and led to embodied/ enactivist approaches. That has led to a definition something along the lines of knowing being an organism's ability to interact effectively with its environment (which would involve being able to predict correctly the results of actions performed on the object). So that would imply that both the organism and the environment are involved in the knowledge.
The camera has no knowledge of the table because it has not had the experience of lifting the table and feeling its weight, being aware of its ability to throw it (and how far), to set it down (and how its weight will affect how quickly it will hit the floor), its surface as a stable place for setting other objects, etc. All of these interactions lead to the perceptual skills necessary to know and understand the table, which is to say to have a trained neural network controlling, planning behavior, categorizing experience with these trained expectations.
(That seems to imply some guiding principles for implementing AI: 1. basic locomotion and physical interaction with the world is an important and non-trivial problem 2. there needs to be a linking theory to extend basic-level knowledge to novel, abstract categories of knowledge grounded in the earlier type. For 1. lots of neuroscience work is relevant including constructivist/ modeling approaches, including the behavior-based AI paradigm and for 2. one such major linking paradigm is the one that started with Rosch/ Lakoff /Faconnier/ Gibbs etc currently under the headings of conceptual metaphor and blending theory, cognitive linguistics, embodied cognitive science)
(So we can think to build computers, but computers can't think to build humans...)
Neither animals nor computers will ever be able of that.
Although, I find his responses quite superficial sometimes.
You find them (freshly created) inside yourself, but you can't tell where they were coming from.
Bullshit - I've had Cocoa Puffs.