Slashdot Mirror
Whatever Happened To AI?
stinkymountain writes to tell us NetworkWorld's James Gaskin has an interesting take on Artificial Intelligence research and how the term AI is diverging from the actual implementation. "If you define artificial intelligence as self-aware, self-learning, mobile systems, then artificial intelligence has been a huge disappointment. On the other hand, every time you search the Web, get a movie recommendation from NetFlix, or speak to a telephone voice recognition system, tools developed chasing the great promise of intelligent machines do the work."
I remember makeing a small program in basic back in "the day" on my apple II+ that would allow others to call my computer via my 300baud modem and ask questions of the "AI" program I was developing. Of course it was nothing more than a magic 8 ball type system that allowed me to preformat a line or three of text to be thrown in at will while I was watching the screen to make it seem smarter. Yes it was a stupid joke, but it supplied me with a week or two worth of laughes.
I Need someone to rebuild a Digitech Digital Delay pedal for me....for me...for me...for me.
I got my B. Sc. in Computer Science with a concentration in Intelligent Systems. The state of academic AI seems to me like a field looking directly for purpose and direction. The problem with AI is that stuff which was once considered part of AI is now considered an algorithm. This is especially true for graph search algorithms such as A* and heuristics. Classification algorithms, from primitive algorithms such as K-Mean to more complex Bayesian models seem to be going down the same path of "just an algorithm."
Nowadays, it seems like planning is the big thing in AI, but once again, it's just a glorified search in a graph, be it a state or plan graph.
AI is an intuitively 'simple' concept, but there's no clear way to 'get there.'
When any particular subset of what we do with our brains (chess, machine vision, speech recognition, what have you) yields to research and produces commercial applications, the critics of A.I. redraw the line and that domain is no longer part of "A.I." As this continues, the problem space still considered part of "artificial intelligence" will get smaller and smaller and nay-sayers will continue to be able to say "we still don't have A.I."
Simpletoneity, n. -- The phenomenon of many people all doing the same stupid thing at the same time.
AI has always been surrounded by a lot of hype, as the idea of creating non-human life has always been an exciting one.
But we're probably as far from creating a true AI as we are from creating biological life from scratch (by synthesizing DNA sequences to build an organism from the molecular level).
AI research is providing useful gains in computer science, and some of those gains trickle down into the real world.
But contrary to what you may have been sold, we're not 10-15 years away from creating Skynet. We've got a long, long way to go, and scientists that aren't trying to get publicity have always known this.
AI hasn't "gone away"... it's just that the false marketing for it has.
Erik
no, that's not an insult or to call AI a pseudoscience
what i mean is: the ancient alchemists goal was to turn lead into gold. which they thought possible, because they did not perceive magic in gold, it was just stuff. surely, with the right manipulations, some stuff could be turned into other stuff, right?
and from that basic fantasy thought came the groundwork for centuries of hard work, the discovery of the fields of chemistry, physics, all the subfields...
such that one day in the middle of the last century, some dudes with some extra time at a cyclotron said "hey, why don't we bombard some lead atoms, i have a feeling about what the decay product will be (snigger)"
and there, as a completely forgotten afterthought, was a fulfillment of the ancient alchemist's original goals, many generations before
to me, i think this is the fate of AI: it will be a formative motivation. just as the ancient alchemist's looked at gold and saw just stuff, we look at the brain and just see neurons. and all of the ffort to replicate the human brain will spawn incredibly sophisticated fields of information science we can only begin to grasp at the foundations of right now. look at databases, for example: that's an effort at mimicking the brain. and look at all of the unintended and beneficial consequences of database reesearch, as a superficial example of what i am saying about unintended benefits being better than the original goal
so perhaps, many centuries from now, some researchers will say "hey, remember the turing test"? and they will giggle, and make something that is exactly what we now envisage as the ultimate fruit of AI research, a thinking computer brain
but in that time period, such a thing will be but an after thought, and much as the rewards of physics and chemistry so dwarf the fruits of turning lead into gold, so whatever these as-of unimagined fields of inquiry will reward mankind with will turn the search for a thinking computer into an equally forgettable sideshow
the search for AI will lead to much more rewarding and expansive fields of knowledge than we can imagine now. jsut like the guys arguing about "phlogiston" could never imagine things like organic chemistry and radiochemistry. just imagine: fields of inquiry more rewarding than thinking computers. that's a future i want to glimpse, and looking for AI will lead us there
intellectual property law is philosophically incoherent. it is your moral duty to ignore it or sabotage it
The robots are coming.
The big breakthrough was the DARPA Grand Challenge. Up until the 2005 DARPA Grand Challenge, mobile robots had been something of a joke. They'd been a joke since Elektro was shown at the 1939 World's Fair. But on the second day of the 2005 Grand Challenge event at the California Motor Speedway, suddenly they stopped being a joke. Forty-three autonomous vehicles were running around and they all worked. The ones that didn't had been eliminated in previous rounds.
Up until the Grand Challenge, robotics R&D had been done by small research groups under no pressure to produce working systems. Most systems were one-offs that were never deployed. DARPA figured out how to get results. There was a carrot (the $2 million prize), and a stick (universities that didn't get results risked having their DARPA funding for robotics cut off.)
The other big result from the DARPA Grand Challenge was that robotics projects became much larger. Nobody had 50-100 people on a robotics R&D project until then (well, maybe Honda). Robotics projects used to be a professor and 2 or 3 grad students. Suddenly stuff was getting done faster.
DoD started pushing harder. Robots like Big Dog got enough money to be forced through to working systems. Little tracked machines were going to battlefields in quantity, and enough engineering effort was put into mechanical reliability to make the things really work.
CPU power helped. Texture-based vision now works. Vision-based SLAM went from a 2D algorithm that sometimes worked indoors to a solid technology that worked outdoors. Much of early vision processing is now done in GPUs, which are just right for doing dumb local operations like convolution in bulk. GPS and inertial hardware got better and cheaper. Some of the mundane parts, like servomotor controllers, improved considerably. Compact hydraulic systems improved substantially.
It's finally happening.
As for the hard stuff, situational awareness and common sense, watch the NPCs in games get smarter.
AI is a Holy Grail. In other words, something we'll probably never get, but we'll create a whole bunch of useful stuff while trying to attain it. "AI" is just a stated goal that gets a bunch of smart people together to develop tools towards that goal. AI research has already given us Lisp and Virtual Machines and Timesharing/Multitasking and the Internet and a bunch of useful data structures and algorithms.
At some point after all that, a computer was developed that can play Grandmaster-level chess, but this was not a necessary development to justify the all research grants.
Not a typewriter
It's not just some keyword matching algorithm thats used. Without going into technicalities you might want to check out the Netflix prize contest, a 1M$ prize to improve the netflix prediction system by 10%.
That's basically the "scruffy" approach to AI, as opposed the "neat" approach, which was to define all the supposed rules that people supposedly follow. There was always a competition between the "scruffies", who thought that neural nets, genetic algorithms, and bayesian nets would enable us to "grow brains in a box" that would eventually be complex enough to think like we do, and the "neats" who could never define all the rules, because they relied on question-answer sessions with the "thinkers" who often thought they were following rules, but often turned out to be using instincts and assumptions that they never consciously thought about.
I was working in an AI research company back in the late 80's, and I remember the fun and "fun" we had back then. I tend to fall more in the "scruffy" category, but I'm coming from an implementation rather than research background, and I saw all the problems with the rule-based approaches. Even getting good probabilities from "experts" concerning their decisions and evaluations, to feed into Bayesian probability nets, was nearly impossible.
Nowadays I think we'll have better luck just following biology augmented with microelectronics. I want my cyberbrain, dammit!
By the taping of my glasses, something geeky this way passes
If it ended with the robot seeing his other selves, realizing he wasn't a beautiful and unique snowflake, and kervorking into the ocean -- THE END -- it would have been a really pretty good movie. Dark, but with a Western message that it is our individualism and uniqueness that make life worth living.
I think Kubrick must have written everything except the ending. He didn't know how to add some inspiring, lifting message to a movie that can't have one.
That's not the same. When there is a success made in any of the fields that you mention it remains part of that field. A solved part of that field. Every success made in AI is no longer AI, so there are no successes or progress made "within the field". It's quite a substantial difference when it comes down to the perception of the field.
Chess was considered the ultimate AI problem back in the 40s and 50s. When we knew little about the game and how to solve it, it seemed that intelligence must be required to solve it. Now that machines are better at chess than humans we've redefined as a problem that is susceptible to brute force. It is not considered a success in the AI field, just another refinement of what is not AI.
Slashdot: where don knuth is an idiot because he cant grasp the awesome power of php
This is the primary point I came in here to say. Whenever I've read anything about AI, it seems to be based on cool science-fictiony ideas, or else it's actually a simpler method to use statistical analysis to approximate human decision-making for particular purposes. If you're talking about real self-aware thinking things, the approaches are all wrong.
People tend to act treat the subject as though dumping enough raw information into a fast enough processor will yield intelligence, and then as that intelligence grows and develops, things like "sensible responses to answers" or "appropriate emotional responses" will emerge. Or else they think grouping enough "appropriate responses" will eventually yield intelligence.
It seems to me that that's all backwards. If you want to design an artificial intelligence, you first need a good philosophical understanding of how intelligence works, which will tell you straight-off something that AI researchers don't seem to consider: intelligence is an animal trait.
I think the absolute first thing you need to do is to figure out how to give machines emotions, to approximate pleasure/desire and pain/aversion. The second thing you need to do is give it "senses", and the ability to draw a very basic sensory conception of its world based on those senses, which includes a sense of time and objects. Also, you'll have to give it the ability to interact with its world in such a way that it is able to pursue its desires, encounter obstacles, and experience "pain". Finally, you'll have to figure out a way to give it the ability to adapt, to "rewrite its programing", preferably in a way that allows it to reproduce and evolve.
So in a way, the most obvious answer is that if you want an artificial intelligence, you'll have to design an artificial/virtual animal and place it into an environment where it can evolve intelligence. There may be some shortcuts on growing/evolving it faster, but you shouldn't be quick to discount the animal nature of intelligence as we know it.
And the reason for these things are bound up with the fact that, like I said, the only model for real intelligence we have to base anything on is animal intelligence. Animals develop and express their intelligence by being self-motivated in a world that presents obstacles. If there's nothing you want, there's no point in figuring anything out. If there's no way to get what you want, then there's no point in figuring things out. If there are no obstacles in your way, then there's nothing to figure out.
So if you don't have a self-motivated desire and the ability to move towards achieving that desire, then you can't make self-determined intelligent decisions. If course, this also presents a scary twist to the whole AI thing, because it suggests one of the chief scifi fears of AI will turn out to be correct: If we're successful in creating AI, we may not be able to control it.
Indeed, there was a time when binary search trees were called "artificial intelligence".
Remember that program to catalogue animals? It started with something like "Is it a dog?", then you say no, and since the database is seeded with only one animal, it would respond with "I don't know the animal, what is it?" ("a bird"). Then it would ask what question would make the difference between the two clear ("Does it fly?"), and next time you run the program, it starts with "Does it fly?". If you say yes, it would ask "Is it a bird?" and so on, and so on.
It's a fun little project while learning how to program, but it's not really counted in the AI-domain anymore.
"It's too bad that stupidity isn't painful." - Anton LaVey
Actually, I detest them, but I do think there is a lot of untapped research potential there, because of the sheer number of people who are willing to sit there for hours on end, waiting for npcs to respawn. With a good learning algorithm and enough entropy (causing 'genetic mutations'), those npcs will eventually find a few optimal ways to react to their environments, prolonging their own lives. They just need people to coach them through it.
With that many users, you'd get enough variation between the newb that's killing them for experience to the maxed-out-character that blows up everything in his way.
It would be really cool if Blizzard let some serious AI programmers go nuts, so that the NPCs try to maximize their own lifespans, rather than just dying and respawning.
Maybe for enough money, they'd let you set up a few thousand bot-controlled characters?
"The cup is in turn designed for holding hot or cold liquids, and has an open rim and closed base." --US Patent #5425497
Check out http://www.20q.net/ . It's a neural network that's been put online for quite some time and does exactly what you describe. It's very interesting to note the final question that determines your answer; Here's me playing vs 20q: I was thinking of a lampshade.
Q20. I am guessing that it is a lamp shade? Right, Wrong, Close
19. Does it weigh more than a duck? No.
18. Is it found on a desk? Sometimes.
17. Is it larger than a microwave oven (or bread box)? Sometimes.
16. Do you use it at night? Sometimes.
15. Is some part of it made of glass? No.
14. Is it worn? No.
13. Is it decorative? Yes.
12. Is it pleasurable? No.
11. Does it move air? No.
10. Is it black? Sometimes.
9. Is it square shaped? No.
8. Can it be easily moved? Yes.
7. Does it beep? No.
6. Can you talk on it? No.
5. Does it usually have four corners? No.
4. Is it larger than a pound of butter? Yes.
3. Does it get wet? No.
2. Do you hold it when you use it? No.
1. It is classified as Other.