Granted that I don't know what use you're putting the MacGuffin to, but could you do something with nanotech devices evolving a hive mind?

Man, is it Sturgeon’s Law City out there — and even the stuff written by the people who know what they’re doing seems pretty damned simplistic and old-fashioned.

You're absolutely right.

Some of them even still seem to think Deep Blue’s defeat of Kasparov says something about AI, instead of something about chess.

I think it says more about Moore's Law than anything else. But you're right...the impact of Deep Blue's success to exploring true AI, or the roots of cognition, are virtually non-existent.

As a professional programmer, this sounds like total fantasyland to me. I’m aware there are plenty of professional programmers and comp-sci types out there who would disagree, but I’m just not seeing any qualitative change over the last ten or twenty years in the sorts of things we’re making computers do.

Again, you're absolutely right. It isn't about stringing more and more powerful processors together. It's about organizing the software and hardware we've already got in novel ways.

Wherever that 100- or 1000-fold increase in processor power is going, it’s off on some other vector, perpendicular to the one that would point toward conscious, humanoid AI. More transistors just gets you a giant Beowulf cluster. More transistors is boring.

Well, I don't necessarily think more transistors is boring, in the sense that all that raw computing power might be necessary to set up a framework for true AI to evolve in. That is, all that computational power allows for more fine-grained simulated worlds in which to evolve artificial agents, and more power to evaluate their fitness over longer lifespans and more generations (in less realtime).

More and more I find myself leaning toward the Roger Penrose (The Emperor’s New Mind) argument that consciousness may rely on non-computable processes, even if I’m not enough of a mathematician to follow his argument or even believe it.

Sorry, I think ENM is one of the worst, most unscientific pieces of tripe I've ever read. It doesn't matter whether you're a mathematician or not...his arguments consist of profuse hand-waving and nonverifiable nonsense.

The only interesting computer stuff going on is in things like genetic algorithms and cellular automata.

Well, yeah. :)

Either one could certainly lead to systems with behavior sufficiently complex and unpredictable that we might have to take their word that they’re intelligent, but neither of them promises anything like the Mycroft scenario;

Well, I've argued for a long time that AI researchers need to formulate a better working codification of intelligent behavior and how to measure and evaluate it (the Turing Test is a cute thought experiment, but is worthless beyond that).

neither of them promises either humanoid AI or simulation of human consciousness in the foreseeable future.

That "foreseeable future" is a very Rumsfeldian phrase. :) I'm a technological optimist. I tend to see an academic resurgence in previously ghettoized paths of AI research (namely neural networks and genetic algorithms). Predictions of emerging, little-understood technologies are generally folly, but I don't happen to think that it is entirely unlikely that AIs as smart as chimps could be bottom-up engineered within the next 10 years. I think the hurdles are conceptual, not technological.

I’m thinking it’s likely to turn out that Wittgenstein’s Lion isn’t just, like Ben Rosenbaum says, the answer to Fermi’s Paradox, but the answer to the Turing Test as well: if a machine was to become intelligent, we wouldn’t recognize it.

I think that's an oversimplification. I personally think there are probably a near-infinite number of permutations on types of minds. Many, perhaps most, would be inscrutable to us. But I don't think it's a foregone conclusion that any machine intelligence is going to be that way.

Consider the coevolution of humans and dogs. For thousands of years we've evolved in complicated feedback loops, so that even at a genetic level, we're finely attuned to the expressions and behaviors of dogs, while dogs are even more finely tuned to us. We understand each other at some very fundamental level.

I tend to think that if genetic algorithms are the gateway to developing AI, if they're coevolved to be highly attuned to human behavior, then mutual understanding might be drastically mitigated.

So I don't think this necessarily leaves you screwed. I think there are still highly-plausible ways to render AI in fiction (though it is rarely done well), and the sort of AI that you seem to want to use (one that interacts with humans on some level).

Well, yeah, where ENM falls down is the “I know I’m conscious” bit, which is not just nonverifiable but nondisprovable, and leads you to the Ken Macleod scenario where the protagonists blithely commit AI-genocide because they just know the AIs aren’t really conscious.

That said, Penrose is a world-class scientist, and I’m not ready to dismiss his Gödel argument as hand-waving when I can’t actually work through it myself. (Strangely, though, I don’t have any problem doing that with Tipler’s Omega Point. Hmm.) My gut feeling, though, it that eventually we’ll find out that like playing chess, proving mathematical theorems doesn’t have much to do with the roots of cognition.

The one really useful thing about what Penrose is doing, is that he’s trying to put a bullet in the head of the popular (in SF) pseudo-mystical interpretation of the Uncertainty Principle, the one that says there really is something special about observation by a conscious observer.

I’m not sure I believe in consciousness anyway — I tend to lean toward what Tom Maddox says Minsky said about consciousness, that it’s not in charge, it’s just a debugging trace. (If anyone’s got a reference to Minsky actually talking about that, let me know.) But if you’re talking (which I was intending to) not just about AI but about running human minds inside simulated environments and that sort of thing, and you’re writing from the point of view of the simulatee, it kind of presupposes consciousness. :)

Good point about GA-based systems being deliberately involved to communicate, by the way.

I suppose we’ll know we’ve gotten somewhere when we get a generation that evolves to the point of being able to plead for its life. :)

As you may remember Dave, much of what you said were things we were discussing at UCSC when I was studying neural modeling and cog sci at UCSC. Then and now, most of the AI field is spending time and energy throwing more computation cycles at the problem, which is just monkeys, typewriters and shakespeare as far as I'm concerned.

Until we get far enough past Descartes in terms of our own understanding of our own consciousness, I just don't see solving the problem.

If someone can tell me that genetic algorhythms are going to get us farther than leaving crap in a glass bowl and waiting for life to occur, I'd love to hear it. Until then, I don't think that's any better of a method of studying/generating AI.

If you ask me, I think we're on the wrong track, and that we're either going to have to rethink the problem, or hope that it spontaneously and accidentially occurs.

If someone can tell me that genetic algorhythms are going to get us farther than leaving crap in a glass bowl and waiting for life to occur, I'd love to hear it. Until then, I don't think that's any better of a method of studying/generating AI.

Well, crap already has plenty of living bacteria in it, so you wouldn't have to wait long (but I'm guessing that's not what you're getting at).

Genetic algorithms show promise because they harness the power of bottom-up design inherent in evolutionary processes (or are we to assume from your blithe dismissal that you don't believe evolution is the means by which all species have originated?).

GAs have been around since the advent of computing, but weren't really seriously formalized until the 70's, and even since then haven't been explored very thoroughly.

If you were involved in neural modeling, then you know how Byzantine even the simplest brains are, and the hopes of forward-engineering something that complex seems nearly impossible. But GAs allow for optimization without having to forward-engineer solutions. What does have to be crucially-defined for a GA to work well is the fitness function (the evaluation of an individual at a particular task or group of tasks in a particular environment). That's why it's crucial to explicitly define what we're talking about when we're talking about intelligent behavior.

And the key word there, B, is behavior. To heck with intelligence, whatever the hell that is; let’s see if we can build a system that exhibits intelligent behavior. It’s an end-run around the whole consciousness problem.

some of the currebt projects coming out of robotics can exhibit some very advanced intelligent behavior, but I don't see how that gets us any closer to the problem. They can sense and respond to external stimulai using any number of "senses', but that doesn't mean cognition. Not that I really want to argue this point, since I still think it's the wrong approach to AI.

Derek - I totally agree with your description of GA - my analogy was perhaps a little too offhanded. What I meant was the same thing that you said - waiting for a GA to spontaneously evolve into AI is a fools game, IMHO. And that's all you can really do with them until we understand our own consciousness/cognition better.

I guess where I’m coming from is, if we don’t even know what the hell we mean when we say “cognition”, then it’s kind of pointless to argue about whether a machine’s cognittin’, or to try to build one that cognits.

David writes: To heck with intelligence, whatever the hell that is; let’s see if we can build a system that exhibits intelligent behavior.

I agree...but then "intelligent behavior" still has to be described in some way. I had a recent blog in which I talked about some possible aspects of intelligent behavior, including, but not limited to:

--spatial awareness
--self awareness
--progressive learning
--symbol acquisition and manipulation
--causal reasoning
--inference
--mathematical ability
--memory and recall
--physical dexterity

Of course, some of these attributes overlap, and certain tests could measure multiple behaviors (e.g., the agent must navigate a 3-D maze in which doors may only be opened by completing certain tasks, like pressing symbols in a particular sequence, toward some goal...and then timing their progress on multiple trials). But I'm making all sorts of inherent assumptions about what intelligence is by the design of the test.

Dave - here you go!

cog·ni·tion ( P ) Pronunciation Key (kg-nshn)
n.
The mental process of knowing, including aspects such as awareness, perception, reasoning, and judgment.
That which comes to be known, as through perception, reasoning, or intuition; knowledge.

:) SO, now what?

Brandon writes: Not that I really want to argue this point, since I still think it's the wrong approach to AI.

With all due respect, what do you think is the right approach to AI?

What I meant was the same thing that you said - waiting for a GA to spontaneously evolve into AI is a fools game, IMHO.

You say things like this, and it gives the perception that you don't understand either GAs or evolutionary theory very well. You speak as if evolution, either biological or artificial, is a completely random process. It's not.

The genotypes and phenotypes produced from any evolutionary process are a direct result of the given fitness function (which is a function of all the selection pressures applied by a given domain).

Take the case of a rabbit. Evolutionary theory doesn't posit that the rabbit either originated or developed randomly or spontaneously. A rabbit's fitness function can be broken down into smaller fitness parameters (such as its speed, ability to sense predators, resist harsh weather conditions, resist parasites, etc.) which combine into its overall ability to reproduce and persist its genes.

If an AI designer using GAs can develop good ways to measure incremental improvements in the type of behavior we define beforehand as intelligent, then that's half the battle. The point is, there is still an awful lot of care and thought that go into creating an environment for the agents to evolve in, and developing good fitness functions. It's not just generating a bunch of random bit strings and hoping that after trillions of cycles you get something that can compose an opera, libretto and all.

You say things like this, and it gives the perception that you don't understand either GAs or evolutionary theory very well.

Derek - it's dangerous to make public assertions about people who you don't know from Adam. :)

Okay, you two, settle down. :)

I didn't assert anything about what you know or don't know. I said your words give that perception.

As I was trying to point out, when it comes to GAs, you don't sit around and wait for anything to evolve spontaneously. You called using GAs to develop AI a "fool's game". Since I work with GAs and neural nets, I could just as easily scold you for calling people you don't know "fools". :)

I'm still curious to know what sort of approach to AI research you think is viable, though.

Derek - did he call developing GA's a fool's game, or did he call expecting GA's to yield what the general public thinks of when they hear the words 'artificial intelligence a fool's game? I took it to be the latter. And, if so, I think he's right.

Genetic Algorithms are tremendously useful as a tool for developing solutions to problems that we couldn't solve through a more straightforward development process. And they're clearly capable of developing intelligence in a certain sense. But they're not going to develop 'artificial intelligence' in the sense of conscious, self-aware machines, in my lifetime.

But they're not going to develop 'artificial intelligence' in the sense of conscious, self-aware machines, in my lifetime.

Hmm...well, as I said, there are too many factors for such speculation to really carry any validity.

Let me as you this: Do you think any method will yield AI in the sense of "conscious, self-aware machines" in your lifetime?

I tend to think that if we do develop AI on par with the dolphins, chimps, or humans, the method by which we do will be bottom-up, GAs or something very similar. We're talking about systems that are orders of magnitude more chaotic and complex than any system we've ever forward-engineered. I just don't think any person or group of people are going to understand enough how a brain works to sketch out a working blueprint and build it.

I think the design will have to be evolved through indirect processes (computational mutation, reproduction, and selection), and that the end result will be something no one completely understands.

I completely missed this whole discussion while I was offline. It's probably too late now, but I wanted to toss in something that seems self-evident to me but that I never hear people talking about so there must be something wrong with it:

We have one example of a physical structure that correlates with something that everyone is willing to agree to call "consciousness": the human brain in the adult healthy normally-socialized human body.

If we want to produce an artificial human-like intelligence, it seems like the most straightforward way to do it is to create an artificial human brain (out of silicon or bioengineered cells or whatever) and put it in an artificial human body and raise it like a human child. If that didn't result in something that looked like consciousness, I'd start to look more seriously at metaphysical notions of souls and such.

I assume that the reason we can't do this experiment right now is that the brain is really complex and has a whole heck of a lot of cells and we don't quite know how it works. But if we believe that consciousness is all physical rather than metaphysical (and let's just say that's true for the sake of argument), then it seems like we should eventually be able to figure out the physical processes the human brain uses and duplicate them.

Of course, if the match becomes too exact—a biological brain in a biological body that's indistinguishable from a human brain and body—then there are other questions, like whether we call it "artificial" or "human." But it seems quite plausible to me that the chemical (and even quantum if necessary) processes of the brain could be simulated and reproduced artificially with a high enough degree of accuracy to make it work.

Is this a long-discredited idea? Is it just so impractical that nobody cares about it?

Straightforward; maybe not practical. But yeah, I agree, in principle that ought to work. It’s just not the direction all the people who seem to think the Singularity’s just around the corner are going.