Wed 21 Nov 2007
Thoughts on AI:
I have a hard time believing we will EVER be able to create a functional AI based on current technology and paradigms, at least according to the Turing test. The current method seems to consist of assigning probabilities to EVERYTHING. This will never fool a human. No human decides if the sprinklers were on based on the wetness of the grass. Correction: no human ACCURATELY determines this. see if you can spot the robot in this sample dialog:
speaker 1: the grass appears to be wet.
Speaker 2: It’s probably from the dew.
speaker 1: statistically speaking for this time of year there is only a .0123 percent chance that condensation would still be present on the grass at this time of day. Given the climate, the water bill, and past watering habits: it is far more likely that the sprinklers have been turned on at some point in the night.
speaker 2: Whatever.
More after the jump
Robots can not infer, can not guess, they can only form opinions based on statistical models. That is IF the brains of said robots were constructed in the fashion of today’s views on AI. This is useless. Even so called “learning” AIs that take in new information about things they already know, so called inference networks, are limited in that they can only infer about what they know.
Much better would be a system where the AI could compare unlike things, and evidently there is some work being done on this subject, and come up with a decision based on previous knowledge. Perhaps this is the goal of the probabilistic models. However I feel they are going about it incorrectly.
The question then arises how do you program an AI to compare objects? indeed how do you program it to decide when to make the comparison. Take the example of shopping carts and cars. A robot is pushing a shopping cart, a college student to engrossed in his DAP and the ripping good tunes he just pirated walks in front of said robot. The robot, not having been programed to handle this, attempts to occupy the same space as the human leading to a, not disastrous, but slightly comedic episode where the student ends up in the shopping cart.
Fast forward. The robot having stocked up on all the robot-y things it needed, is now heading home. When an unwitting motoris,t to engrossed in the political musing of the day time talk show to notice our hapless robotic buddy, runs a red light. The robot should be able to recall what happened with the college student and based on that information avoid the motorist. Current systems of AI would not allow for this.
I am a computer scientist. I know how impossible most of this would be to implement. Ultimately I am certain my way of thinking about it is just as wrong as I perceive the currently accepted methods. The real problem is one of understanding. We don’t know how the brain works well enough to form an accurate model. What causes me to think about marigolds when I see biplanes? what is that link? does an AI even need to MAKE such a link? Well that depends on what the AI is controlling.
In the end I think it comes down to evolution and what the robot needs to do. What do humans need to do? Mostly perpetuate the species. So eat, sleep and breed. These things are generally accomplished under harsh circumstances. AIs do what they are programed to do. So make a robot that is designed to regenerate and replicate. Though direct replication isn’t enough there has to be improvement as well. Each generation must be advanced enough over it’s predecessors to enable it to replicate more efficiently. But then what’s the point? The planet is already covered in things that do this. Infact EVERY living thing does this. So why do we need another thing that competes for resources?
The ultimate point here is that true AIs are not needed. What is needed is something that can be given a task that humans don’t want to do, and can accomplish that task with the minimum amount of effort and input from humans. Essentially things like Roombas plug it in and let it vacuum your house. How long until we have robotic custodial engineers? They don’t need to look at the high-school bathroom and decide if it needs to be cleaned, the robot will just clean it. Trash could be emptied, toilets cleaned, halls swept or vacuumed all without human intervention. These things are so close to being possible that it is odd they don’t happen.
It seems to me that in the future, most dull repetitive jobs could be done by machines. RFID tag your trash cans and Rosie will be able to empty it for you. This is close to being real. Heinlein said it in door way into summer, the book where he predicted the roomba by the way, the last person to benefit from technology is the person who does the housework, but the person who does will be wealthy beyond the dreams of avarice. I’ll leave you with this thought: what happens when the custodial engineers are replaced by the Street Sweeper 5000 custodial automaton?
December 25th, 2007 at Tuesday 07:23:58
Hello,
My friend, Thomas Leeb, referred me to this page after receiving my CD, A.I. Awakens. I’m a musician/composer, but I used to be a programmer and did a bit of AI/neural modeling research in the 80s.
Anyway, I agree that the current AI paradigms are sorely lacking. And I don’t think much progress will be (nor has been) made by having programmers program EVERYTHING into an AI.
I am much more hopeful for something along the lines of Ray Kurzweil’s vision put forth in his book, The Singularity is Near. He has the big-picture vision, which requires huge advances in brain scanning, computational speed and nano-technology.
He believes that, the rate of technological innovation is accelerating at an exponential rate and within a bit more than 20 years, we will have the tools in place to replicate, to a level of sufficiently fine-grained detail, the functionally-relevant properties of the human brain in some kind of “computer” based neural network.
Once the form has been replicated, to a sufficiently detailed degree, it can have the potential for human-like learning.
I’m so glad to see that you mentioned something about the importance of inherent biological imperatives that drive human motivation. This whole aspect is little, if ever addressed, in current AI conceptions.
People actually have MOTIVATION to make inferences and the more they make, that are strongly linked to emotions and to satisfying needs/wants, the stronger they are imprinted and build upon each other and eventually that can lead to more abstract mental connections, which may find a high-level, analogous link between marigolds and biplanes.
But the point is that all of these inferences, feats of deduction and induction emerged out of the system “for free,” meaning , programmers didn’t have to think of gnarly algorithms/heuristics that would lead to such far flung connections.
Once the form and function of the neural network is there COMBINED with some instinctual, OS level, motivations of the system to learn and make abstract connections , you will have an evolving, learning system.
This idea of emergent properties of the brain is really delved into in Douglas Hofstadter’s books: Godel, Escher, Bach and Metamagical Themas. If you’re interested in this kind of approach to AI, I’d be glad to send you a paper I wrote on this topic back in 1987, called: Levels of Representation in Mind/Software and Brain/Computer.
Well, this is Christmas, so, Merry Christmas. Wow, I didn’t expect to be pontificating about AI paradigms today…Thanks for the link Thomas.
Chris Armstrong
http://www.SpaceageFurnitureMusic.com
March 22nd, 2008 at Saturday 09:32:02
So what your saying is that we need to focus on creating an algorythm that is capable of making it’s own decisions based on ALL of it’s surroundings, as opposed to trying to program every possible option it would ever encounter. I do think the answer actually lies between the two. Say you create your “Rosie Trash Colelctor 5000″. It would have to have a fair set of options built into it to be able to navigate and not run over paople or other “analog” interferences. Then when you have a basic “AI” for accomplisihning each of several hundred different jobs, all you have to do is start combining them. I think it is ludicrous to try and create an all encompassing AI. What you need to do is make each piece on it’s own, beta test it for years, and then start combining them. Given enough time (and enough monkeys with typewriters) you would have something close to a functional AI.
Or, just like, try and get the roomba to smoke pot. There’s an experiment for ya.