Rory - Neopoleon : Current Artificial Intelligence is CRAP

Welcome to Neopoleon - Sign in | Join | Help

Navigation: Home | Forums | Galleries

Current Artificial Intelligence is CRAP

[Disclaimer: I dropped out of high school with a .86 GPA (seriously). Anytime you listen to anything serious I have to say, you do so at your own peril.]

I just left a meeting with my new migraine-specialist neurologist. The good news is that he thinks my chances of dying within the next five to ten minutes are, like, really small. The bad news is that, historically speaking, human beings have a 100% success rate with death, and I’m still going to die someday. He wasn’t able to make me immortal like I was hoping.

Anyway, while in there, we got sidetracked and had a nice little discussion about brains, computers, and stuff like that. It reminded me a lot of my bygone days of pseudo-philosophical navel-gazing on the subject of artificial intelligence.

When I first started thinking about it, I was working in a warehouse, stamping boxes with these big official stamps, and had just recently told every last human being who worked at my high school to go to hell. In other words, I didn’t really know anything about anything. I just knew that correct placement of official stamps on boxes made me $6.25/hour.

The thing is, stamping boxes eventually becomes a function performed autonomously by your arm. You don’t need to think about it anymore, and for $6.25/hour, you don’t feel especially pressured to, either. That left a lot of time to contemplate the universe.

What I decided back then was that all that was necessary to create an intelligent machine was a sufficiently complex network of circuits that could reprogram themselves on the fly, a set of environmental inputs to influence the structure of the circuits, and a motivation to function (for humans, not-dying and procreating are the driving forces behind much of what we do). I don’t recall exactly why I thought all of that, but I did. The memories are pretty fuzzy, but that’s probably because all the food I could afford back then was a lump of MSG with a meat attached to it, and it made me feel weird, like there were psychedelic eels living in my stomach.

Fast-forward about twelve years, and we have Jeff Hawkins, inventor of the Palm thingy, employing similar ideas to my MSG-induced hallucinations at his “Redwood Center for Theoretical Neuroscience.”

It was kind of cute when I had my little thoughts all those years ago. You can excuse a teenager who’s been smoking way too much pot for coming up with silly notions about the world, but… Jeff Hawkins?

This is one of my gripes with “computer science,” and particularly where it intersects with artificial intelligence theory.

There is just no science. Thinking like this - that we can push forward with a human-created model of the brain based on observation of behaviors and some sketchy medical papers – is pure hubris. Hubris, hubris, hubris.

Hubris.

With a heavy dose of enthusiasthma.

The human brain is not abstract. It isn’t like math or logic – you can’t just sit around and think your way to a universal theory of intelligence (well, you can, but you will be wrong). The human brain, as far as we know, is the most complex machine in existence, and although many of its structures are clearly visible through casual inspection, the real business is going on at a level that isn’t accessible to garage hax0rz and armchair thinkers. You need, like, microscopes and stuff to see what’s going on. And maybe even a lab coat. And slides. And dyes. And big machines that have magnets in them. And fluids. And other expensive whatevers. There isn’t going to be a Steve Wozniak who, on a budget of three dollars and living out of his VW, is going to solve the problem of AI.

I recently read (no enthusiasthma here – I really am this excited) a fascinating article on artificial intelligence in the latest issue of Skeptic. In it, Peter Kassan just lays down the law and delivers a Dragon Warrior roundhouse kick to the nuts of AI theory.

He breaks down the history of AI research, its foundations, its lofty goals, and its repeated failures.

The argument he finally makes is so god damned simply beautiful: You simply cannot create a good model of something that you don’t understand.

That’s the problem. As long as you’re trying to model human intelligence, you need to know what in the hell it actually is. If I performed a survey right now, and if I asked each of you what you thought intelligence was, my assumption is that there would be some common overlap of grander ideas with some major variations in the details.

And that’s just looking at intelligence from the outside. Even if we could all come up with a definition we agreed on, without understanding how the brain actually works, there’s no guarantee that we’d be even remotely correct. At best, we’d have some of it right, and at worst, we’d just be in agreement about something totally wrong. In the end, we’d be no closer to having a model of intelligence against which to base an artificial intelligence. All we would have done is to have agreed on the appearance of intelligence. It’s much closer to religion than it is to science.

It’s superficial. The current state of things is that the Jeff Hawkinses of the world have decided, based on their own navel-gazing, that they know what intelligence is and how the brain works, and now they’re trying to build AI on those foundations. I think it’s noble, but not the right way to do this.

If I had to guess, and I don’t have to, but if I did, and I don’t, but I will, I would guess that, in a few years, the people out at the “Redwood Center for Theoretical Neuroscience” are still going to be doing a lot of head scratching.

When I think about human beings, right now, trying to create artificial intelligences based on what little knowledge we have of the workings of the human brain (or any brain, for that matter), I can’t help but think of a caveman drawing a picture on the wall of his cave and wondering why it isn’t making him warm.

If you’re interested in this stuff, then you should get a copy of the newest Skeptic. It’s the one with a reading robot on the cover. Brilliant.

And, just so y’all know, I would like more than anything to see an artificial intelligence created. I just think we’ve been going about it the wrong way.

Published Thursday, March 30, 2006 9:02 PM by Rory

Filed Under: Enter the Nerd, General

Comments

	Andy said: I think you misread Hawkins intentions. Here is a quote from him: "If you look at the history of big obstacles in understanding our world, there's usually an intuitive assumption underlying them that's wrong. In the case of the Solar System it was intuitively obvious that the Earth was at the center of the Solar System and things moved around us, but that just turned out to be wrong. ... And it intuitively seems correct that the brain is just some sort of computer -- it just seems natural. ... But it has undermined almost all of our work to build intelligent machines and understand thinking. It's just wrong ... the brain isn't like a computer at all." He thinks along the same lines as you and Peter Kassan do. The current way of trying to develop intelligent machines is broke, we don't understand how the brain works and how being self-aware works. He is looking into a better way. Personally I give him a better chance than most of figuring it out. So he basically has a slush balls chance in hell of figuring it out. Most people only have a full on powdery snow balls chance. March 30, 2006 9:34 PM
	Rory said: Andy - The sentiment is the same, but Hawkins' method is still down to trying to mimic human intelligence using the let's-create-a-super-complex-machine-and-then-stimulate-it plan. The problem is that, regardless of his intentions, we just don't have an understanding of how intelligence works. How can we build a machine that mimics it? I'm not saying we couldn't, but it seems like the "Let's put 500 monkeys in a room with computer parts and see what happens" approach. It's not exactly random, but it is trying to come up with the answer to a problem when we haven't even been able to define the terms of that problem. March 30, 2006 9:47 PM
	Jake Good said: If you've read On Intelligence, he maps out his theory on how the brain works... it's merely his way of trying to figure out how it all works. but I agree with everyone so far... it sucks! March 30, 2006 9:51 PM
	Haacked said: Just because we don't know how intelligence works doesn't mean we can't create automata that can mimic it. At least for specialized situations and tasks. The question is how well can we mimic it? In some ways it reminds me of solving systems of equations. You can take the differential equations approach and create nice elegant solutions for some very specific cases. But in the real world, most system of equations don't have elegant solutions. Instead you take a stochastic approach and come up with a pretty good approximation. The more data points you throw at it, the better the approximation. However, this approach to AI isn't well suited for general purpose AI, but does seem well sutied for specialized task AI. March 30, 2006 9:54 PM
	Rory said: Haacked - "Just because we don't know how intelligence works doesn't mean we can't create automata that can mimic it. At least for specialized situations and tasks." Yeah, but that just makes me think of ELIZA. I'm talking about creating something based on a human intelligence. The elevators in my building, for example, seem to have figured out that the parking levels of the building are a good place to wait (since so many people come and go through the garage), but, seriously, they're terrible conversationalists. March 30, 2006 9:57 PM
	George said: Didn't you already write this post? Or am I imagining things. Now I'm going to have to go look to make sure I'm not going crazy(-y +ier). I swear I already read this. Maybe it was a similar post involving you working with boxes and dropping out of high school and it being monotonous and having lots of time to think? I'm so confused now. Am I dreaming or am I awake? AAAAAAAHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHH! March 30, 2006 10:12 PM
	Rory said: George - I think I've written almost 500,000 words for Neopoleon.com. I probably repeat myself occasionally. I know I've talked about my crappy old jobs before - don't remember if they had to do with AI. I really have to stop smoking so much crack. March 30, 2006 10:19 PM
	Ai said: Rory, All this talk of 'AI', and not a single reference/trackback/pointer to me. I'm a little disappointed (ok, I cried a little). I must say that I didn't read all of your post, but I skimmed through it enough to know you didn't write about me. So, this is it, Rory BlythE. You've crossed the line. We're friends, no more. Ai :\| March 30, 2006 10:49 PM
	edddy said: The problem with AI is that we are missing the I part March 30, 2006 11:45 PM
	Brook said: In the last twenty years there has been many strides in AI. Neural Nets and Genetic Algorithms are the two that come to mind right now. These aren't algorithms that will make a robot walk and talk like a human and debate the meaning of existence, but these are algorithms that when setup correctly, solve problems and display intelligence by making educated decisions for themselves. Don't go thinking that we won't ever understand the human brain or be able to make a toaster that gets angry and burns toast when you ignore it's questions about human nature. Just because we don't understand it now, doesn't mean we won't understand it tomorrow. March 30, 2006 11:54 PM
	Rory said: Brook - "In the last twenty years there has been many strides in AI. Neural Nets and Genetic Algorithms are the two that come to mind right now. These aren't algorithms that will make a robot walk and talk like a human and debate the meaning of existence, but these are algorithms that when setup correctly, solve problems and display intelligence by making educated decisions for themselves." Again, though, for a human type intelligence, neural nets (for example), don't even remotely cut the mustard. Too simple - too binary. Too few. Too small. Too wrong. And genetic algorithms aren't intelligent at all - it's like tossing the coding dice. You just test outputs given a constrained set of inputs and hope that the result solves your problem. That makes their use somewhat limited, eh? "Don't go thinking that we won't ever understand the human brain or be able to make a toaster that gets angry and burns toast when you ignore it's questions about human nature. Just because we don't understand it now, doesn't mean we won't understand it tomorrow." I totally agree with that. But we're never going to get there by navel-gazing (except, perhaps, to rule out what definitely won't work). The whole point here, actually, isn't that a human type AI is a stupid idea, or that it's impossible, but that it's not something we're going to arrive at by guessing when we don't even know how to define the thing after which we're modeling our artificial solution. I'm as enthusiastic as anybody. I want an army of robotic sex slaves, too, you know? March 31, 2006 12:06 AM
	Matt said: I've actually done work in the AI field, and I don't think we're going to get true sentience anytime soon. Specifically, I wrote a genetic algorithm in Java (yes yes, I know, but the partner was Java only, even though I wrote all of it). Such algorithms have produced circuit designs, pathfinding algorithms, antenna designs, and other such nifty things better than humans have. By a significant margin. And faster than plain brute force. So they're damn useful. But don't expect human-like antics out of your computer anytime soon. March 31, 2006 2:08 AM
	crucible said: You know I'd be as interested in anyone to have AI from a technological standpoint - but at the same point in time I really don't want it. I mean, what do we really need something other than television thinking for us? The main problem I see about Artifical Intelligence is people want these systems to learn from their mistakes. I see intelligent people making the same mistakes time and time again - I don't get why these systems should be any different. I think some boundries of expectations need to be formed. Do we want a computer that thinks, a computer that thinks like a human does, will that human be an inbred hick who doesnt think very well or Einstein - cos I'm sure he made mistakes as well, and probably didn't learn from them. Are we more interested on what an AI learns, or it's talents for human mimickry - are we trying to design a computer human? And if so, are these bastardising, egocentric androids going to be coming after my job? Are they going to have racism, anti-gay, anti-religious bullshit programmed in or are they supposed to be better than humans? I think its less about how they learn, and more about how they adapt, link their learning with other things, and how they develop into a flawed, unique being. And quite frankly I'd rather see the investment in all this going to cure cancer. Although I'm guessing the pro-robot people would be telling us we can get the robots to cure cancer in the long run... but you know they're just trying to create these things to score a date. The problem is, if they programmed true intelligence into these devices - they'd still get turned down in favour of the microwave. March 31, 2006 3:25 AM
	Ryan Roberts said: Did some stuff with genetic algorithms for my university project, breeding prisoners dilemma playing programs. The big problem with GA is that simulating the environment to evolve the software you want is often at least or more complicated than solving the damn problem in the first place. I also experienced this when writing image analysis software for biochemical assays - many of our competitors were using neural nets and all sorts of clever stuff. We used a custom scripting language and solved the (pretty simple, compared to some image analysis problems) feature recognition and scoring for every assay type. Guess which approach got better results (and failed safe) on edge cases.. What we need is to be able to evolve the algorithms indepentently of human assesment, using emergent criteria that increase in complexity. Of course, doing that there is no way we could control the results and a genetic algorithm is very likely to get stuck on 'useless' local maxima. March 31, 2006 11:50 AM
	-dan said: Here’s a thought: Who are we using as the yardstick of Intelligence? I’ve seen what is passing for average intelligence these days, and I have to tell you it scares the knockwurst out of me… and I’ve never had knockwurst! Or more to the point, judging from the choices made by the powers that be, the average intelligence is on par with a Casio Calculator, and I’m not talking about a scientific calculator here, just the throw away ones you get as giveaways with someone's logo on it. So with that in mind, I think you’re wrong, I think machines have at least passed common intelligence. Unless you know of a computer that runs its own self created destruct program. Have a nice day March 31, 2006 2:27 PM
	Greg said: There's a whole alternative school of artificial intelligence centered around the work of Rodney Brooks at MIT. Brooks argues that ai will not be designed but will emerge. Just like as in humans out "intelligence" is not an entity separate from the regulation of our bodies, but a side effect of it. He works on making robots that demonstrate emergent human, animal, and insectoid behavior while obeying relatively simple sets of rules -- like a humanoid robot that makes proper facial expressions in response to someone who is talking to it and looking at it rather than "modeling" emotional states, walking robots that fall down a lot and alter their gaits to try to fall down less, and swarms of tiny robots that can accomplish relatively complex tasks even though each of their inidivduals only do something extremely simple. Brooks also places two important caveats on the definition of ai. Most importantly, he says that there's no reason to expect an artificial intelligence to be anything like human intelligence in character. Human intelligence arose out of the specifics of our bodies and a robot's intelligence will do the same for their very diffrerent bodies. We will have to come up with new tests for recognizing this intelligence when it does emerge. Secondly, he erases the distinction between intelligent behavior and "true" abstract intelligence. About the question of whether the facial expression robot is actually "experiencing" the emotions it displays, he makes the point that he has just as much evidence that it is doing so as he does that a real person is. March 31, 2006 7:19 PM
	Robert Daniel said: For further reading, I would highly recommend Jaron Lanier's One Half A Manifesto. http://www.edge.org/3rd_culture/lanier/lanier_p1.html March 31, 2006 9:42 PM
	crucible said: Just had a thought waking up this morning - yeah I thought about this first thing this morning, I need a life. What if we are God's little AI project, and to pass the final test of intelligence we have to create our own new intelligence ourselves - propogating a great chain of existance. And then become Gods ourselves. That or you have my vote for the sexbots. Really need that life. March 31, 2006 10:12 PM
	neileboi said: Why would we want to mimic human intelligence in a computer? Since there are already billions of people on the earth to have a conversation with, I do not see the point in building a robot to converse with. Some of us waste too much time in the house absorbed in our televisions and computers living vicariously through the characters in shows and video games when we could be out exercising and socializing with real people, having a much more fulfilling existence. I think the term AI is outdated and based on old sci-fi fantasies about the role of machines in the future. Genetic algorithms, simulated annealing, etc. are better thought of as problem-solving tools. It is irrelevant whether human-style AI is ever achieved. Computers have already proven their worth in many areas which would at one time have been thought of as requiring intelligence, e.g. playing chess, seeking travel routes, equation-solving. Once a computer can do something better than a human, it is no longer thought of as requiring intelligence. As time goes by, there will be more things that computers do better than people. It is hubris to think that the human brain is the most complex machine. The human brain is the most destructive machine. Our "intelligence" has resulted in the extinction of many species. We live a lifestyle which is not sustainable. Once the oil runs out, billions of people will die, regardless of how many windmills and nuclear generation stations have been built in the meantime, because our dependancy on oil goes far beyond the need for energy. I think there are other animals on this earth that are more intelligent than humans, but we have too much hubris to admit it, much like Kasparov could never admit that a computer could possibly beat him at chess. April 1, 2006 11:36 AM
	Caleb said: Reading through your post and replies - which only caused me to start doing my own pondering about the whole AI bit - I suddenly got the image of everyone's friend Marvin in my head http://www.starstore.com/acatalog/marvin-figure-10-1.jpg April 3, 2006 10:43 PM
	Mark Miller said: I took one of those philosophical courses on AI in college. Interesting course, though it did get, I think, absurdly abstract. I remember one of the operating theories of the time was the "humunculous" (don't ask me how they came up with the name--maybe it was Latin. I don't know). Anyway, the humunculous was supposed to receive signals on cards in sequence and then respond to those signals with actions. I don't even remember what the point of it was. Something about creating a system where only rules applied and no interpretation of input was necessary. I brought the whole thing down by saying, "But the humunculous STILL has to interpret the cards to differentiate one card from the next, recognize what the 'signal' is, and then respond appropriately!" I could just see it, from how my professor responded to that: "Oh well. Back to the drawing board..." I did learn something from the course. Fast forward a little bit. As I got older I recognized that one of the things that was discouraged in my family was making a mistake. I tried to avoid mistakes at all costs. It was nerve wracking. In the course we talked about different AI systems that had been tried. One was a system that was actually being tried by a particular individual over a period of many years. It's amazing to me that this actually got funded. This man set up a computer system where experts in various fields would be interviewed, and then the results of those interviews would be entered into the AI system. The structure was such it was told everything it was supposed to know, and it was allowed to process the information it was given and try to form relationships. The idea being to give the AI system "common sense". Some of the constructs given to it were just lame. I remember one of the examples was that Abraham Lincoln died in 1865, "and is still dead". I got the impression that they had to do that for EVERY SINGLE PERSON they gave to it. "Yes, so-and-so died on X date, and yes, is STILL DEAD." I mean, that is just lame. I developed an opinion pretty quickly that it was just unworkable. Then we got into neural networks, and those seemed to be a step in the right direction, at least. You're right, Rory, that they're based on stimulation. But the thing I learned from them is that neural networks used a feedback loop. They were allowed to output or try something and receive feedback about whether the output or attempt was desirable. In other words, it was allowed to make mistakes and learn from them. Further, the operation of the net controlled how it learned things, not the person or thing giving it feedback. So it could learn things that an operator couldn't even anticipate. It was presented as a more flexible model than the former one where everything was spoon-fed to it. The "spoon-fed" model broke the moment it was given a situation that it had not been given knowledge about previously. At least the neural net would be able to make an attempt at giving an appropriate response. Whereas the "spoon-fed" machine would just sit there going, "???" I think the most promising AI I've seen to date was from a guy who just chucked the whole idea of AI being digital. He went analog. He showed some little "insects" he had made from small boards with solid state electronics on them, little servos, and wires. Very simple, yet effective. He showed them, each with a different number of legs, all able to walk over varying terrain. He showed how he could even bend a couple of the legs on one, and it would still figure out how to walk on its changed legs. His theory is that we need to start simple--start with insects. Learn what makes them tick, and from there move up to more complex "organisms". Oh yeah, and I think he figured that the best way to model a brain is to use analog technology, since brains are analog themselves. That way you avoid the "binary-ness" of the digital models. April 5, 2006 11:30 AM
	sifta said: I realize that the article is hard to get access to, since I also haven't seen it electronically. You might have to go to a library, though the skeptic website says that permission is granted to redistribute w/ proper acknowledgement. That said, the article is really good. I can't help but think that a lot of the commenters here would do well to read it. Kassan is addressing the philosophical underpinnings of AI by following the current threads of thought to their logical conclusions, and pointing out the reality gap (which is enormous) with each of these. It is unfortunate that the article wasn't in a more popular magazine (like Science). Pointing out the existence of neural nets (a second-rate function approximator) or genetic algorithms (a second-rate search algorithm) is WAY below the level of Kassan's discussion on why AI has gone awry. May 23, 2006 3:43 PM

Current Artificial Intelligence is CRAP

Comments

Andy said:

Rory said:

Jake Good said:

Haacked said:

Rory said:

George said:

Rory said:

Ai said:

edddy said:

Brook said:

Rory said:

Matt said:

crucible said:

Ryan Roberts said:

-dan said:

Greg said:

Robert Daniel said:

crucible said:

neileboi said:

Caleb said:

Mark Miller said:

sifta said:

About Rory

This Blog

Post Calendar

Post Categories

Feeds