Slashdot Mirror
Why Ray Kurzweil's Google Project May Be Doomed To Fail
moon_unit2 writes "An AI researcher at MIT suggests that Ray Kurzweil's ambitious plan to build a super-smart personal assistant at Google may be fundamentally flawed. Kurzweil's idea, as put forward in his book How to Build a Mind, is to combine a simple model of the brain with enormous computing power and vast amounts of data, to construct a much more sophisticated AI. Boris Katz, who works on machines designed to understand language, says this misses a key facet of human intelligence: that it is built on a lifetime of experiencing the world rather than simply processing raw information."
The old `Chinese Room` again.
The Complete 1 Atlantic Recordings 1956-1961
It's Penrose vs Hofstadter! (Seriously, haven't we done this before?)
You can draw a distinction between experiencing the world and processing raw information, but how big of a line can you draw when I experience the world through the processing of raw information?
It is no longer uncommon to be uncommon.
It won't be perfect, but "fundamentally flawed" seems like an over statement to me. A personal AI assistant will be useful for somethings, but not everything. What it will be good at won't necessarily be clear until it's put into use. Then, any shortcomings can still be improved, even if certain tasks must be more or less hard-wired into its bag of tricks. It will be just as interesting to know what it absolutely won't be useful for.
Kurzweil is delusional. Apple's Siri, Google Now and Watson are just scaled-up versions of Eliza. Circus magic disguised as Artificial Intelligence is just artifice.
The data vs IRL angle isn't in and of itself an important distinction, but an entirely valid concern that is likely to fall out of this distinction (though needn't be a necessary coupling) is that the brain works and learns in an environment where sensory information is used to predict the outcomes of actions - which themselves modify the world being sensed. Further, much of sensation is directly dependent on, and modified by, motor actions. Passive learners, DBMs, and what have you are certainly able to extract latent structure from data streams, but it would be inadvisable to consider the brain in the same framework. Action is fundamental to what the brain does. If you're going to borrow the architecture, you'd do well to mirror the context.
I've always thought it was about information combined with wants, needs, and fear. Information needs context to be useful experience.
You need to learn what works and doesn't, in a context, with one or many goals. Babies cry, people scheme (or do loving things), etc. It's all just increasingly complex ways of getting things we need and/or want, or avoiding things we fear or don't like, based on experience.
I think if you want exceptional problem solving and nuance from an AI, it has to learn from a body of experience. And I wouldn't be surprised if many have said so, long before I did.
We have always assumed that humans are essentially a very sophisticated and complex version of the most sophisticated technology we know. Once it was mechanical clockwork, later steam engines, electrical motors, etc. Now it is digital logic - put enough of it in a pile, and you'll get consciousness and intelligence. A completely non-disprovable claim, of course, but I doubt that it is any more accurate than previous ideas.
There's a lather/rinse/repeat model with AI publication. I encountered it in configuration (systems designed to build systems), and it goes like this: 1. We've built a system that can make widgets out of a small set of parts, now we will build a system that can generally build artifacts! 2. (2-3 years later). We're building an ontology of parts! It turns out to be a bit more challenging! 3. (5-7 years later). Ontologies of parts turn out to be really hard! We've built a system that builds other widgets out of a small set of -different- parts! The models of thought in AI (and to a lesser extent cog psych) are still caught up in this very algorithmic rule-based world that can be traced almost lineally from Aristotle and without really much examination of how our thinking process actually works. The problem is that whenever we try to take these simple models and expand them out of a tiny field, they explode in complexity.
"Always listen to experts. They'll tell you what can't be done and why. Then do it" (from the Notebooks of Lazarus Long)
What happened to the spirit of "shut up and build it"? Google is offering him resources, support, and data to mine. We have to just admit that we don't know enough to predict exactly what this kind of thing will be able to do. I can bet it will disappoint us in some ways and impress us in others. If it works according to Kurzweil's expectations, it will be a huge win for Google. If not, they will allocate all that computing power to other uses and call it a lesson learned. They have enough wisdom to allocate resources to projects with a high chance of failure. This might be one of them, but that's a good sign for Google.
So what you are saying is the computer, like humans, will be boxed in by their own perception?
How is this metaphysically different from what we *do* know about our own intelligence?
---Up Up Down Down Left Right Left Right B A START
COMMON SENSE - so rare, it's a god-damned super power!
A technology editor at MIT Technology Review says Kurzweil's approach may be fatally flawed based on a conversation he had with an MIT AI researcher.
From the brief actual quotes in the article it sounds like the MIT researcher is suggesting Kurzweil's suggestion, in a book he wrote, for building a human level AI might have some issues. My impression is that the MIT researcher is suggesting you can't build an actual human level AI without more cause-and-effect type learning, as opposed to just feeding it stuff you can find on the Internet.
I think he's probably right... you can't have an AI that knows about things like cause and effect unless you give it that sort of data, which you probably can't get from strip mining the Internet. However, I doubt Google cares.
Learning without forgetting is possible if, for example, you reconstruct the network, preserving the old one (and this can be optimized so the entire network doesn't have to be duplicated.)
But I'm curious why you think a mind is necessarily a neural network. Are you saying there is no other possible way to construct a mind? As far as I can tell, there are lots of other designs, many of them far superior to neural networks, especially for such basic things as representing knowledge.
Yes, and actual Intelligent Machine would be boxed in by its own perceptions. Our reality is shaped by our experience though our senses. Lets say, for the sake of argument, that Watson is actually a Machine Intelligence/Strong AI, but the actual problem with it communicating with us is linked to its "Reality". When the Urban dictionary was put into it all it did was start swearing, and using curses incorrectly. What if that was just it having a complete lack of context for our reality. Its reality is just words and definitions after all. To it the Shadows on the wall is literally books and text based information. It cant move and experience the world in the way that we do. The problem of communication becomes a metaphysical one based in how each intelligence perceives reality. We get away with it because we assume that everyone has the same reality as context, but a machine AI does not necessarily have this same context to build communication off of.
Hogwash! The weightings you talked about are the memories. They may not be easily recognized as a coherent memory (or part of) by a casual observer, but that's not the same as not being a "memory". You are confusing observer recognition with existence. Confusion does not end existence (except for stunt-drivers :-)
As far as whether following the brain's exact model is the only road to AI, well it's too early to say. We tried to get flight by building wings that flap to mirror nature, but eventually found other ways (propellers and jets).
Table-ized A.I.
We've heard this before from the top-down AI crowd. I went through Stanford CS in the 1980s when that crowd was running things, so I got the full pitch. The Cyc project is, amazingly, still going on after 29 years. The classic disease of the academic AI community was acting like strong AI was just one good idea away. It's harder than that.
On the other hand, it's quite likely that Google can come up with something that answers a large fraction of the questions people want to ask Google. Especially if they don't actually have to answer them, just display reasonably relevant information. They'll probably get a usable Siri/Wolfram Alpha competitor.
The long slog to AI up from the bottom is going reasonably well. We're through the "AI Winter". Optical character recognition works quite well. Face recognition works. Automatic driving works. (DARPA Grand Challenge) Legged locomotion works. (BigDog). This is real progress over a decade ago.
Scene understanding and manipulation in uncontrolled environments, not so much. Willow Garage has towel-folding working, and can now match and fold socks. The DARPA ARM program is making progress very slowly. Watch their videos to see really good robot hardware struggling to slowly perform very simple manipulation tasks. DARPA is funding the DARPA Humanoid Challenge to kick some academic ass on this. (The DARPA challenges have a carrot and a stick component. The prizes get the attention, but what motivates major schools to devote massive efforts to these projects are threats of a funding cutoff if they can't get results. Since DARPA started doing this under Tony Tether, there's been a lot more progress.)
Slowly, the list of tasks robots can do increases. More rapidly, the cost of the hardware decreases, which means more commercial applications. The Age of Robots isn't here yet, but it's coming. Not all that fast. Robots haven't reached the level of even the original Apple II in utility and acceptance. Right now, I think we're at the level of the early military computer systems, approaching the SAGE prototype stage. (SAGE was an 1950s air defense system. It had real time computers, data communication links, interactive graphics, light guns, and control of remote hardware. The SAGE prototype was the first system to have all that. Now, everybody has all that on their phone. It took half a century to get here from there.)
The human brain doesn't "store" information at all (and thus never processes it).
This sounds like mere semantics to me. Yes, there isn't a little television screen playing that one time when you broke your arm, with a post-it note attatched saying "memory #4 April, 3, 1956". But there is a deeply encoded structure of chemical potentials, and neural connections which represents this memory. It is stored, and it is, obviously, processed. If it wasn't so, then how could this memory be subject to action and further processing?
Yes, it isn't stored like a video file is stored on your computer, or a photo in your album; but this doesn't mean it isn't stored. If it is an object of thought, it is in the brain, and if it is re-callable, it is stored.
We know from fMRI that "free will" does not exist and that "thoughts" are the brain's mechanism for justifying past actions whilst modifying the logic to reduce errors in future - a variant on back-propagation. Real-time intelligence (thinking before acting) doesn't exist in humans or any other known creature, so you won't build it by mimicking humans.
Huh? I'm not going to get into the agency (free will) debate... But if it did exist, I don't think our understanding of the brain is really up to snuff enough to allow some fMRIs to show it. If it does exist (again, I'm not getting into it), I doubt very much that it will be a little glowing ball located in the middle of your brain (again with a post-it saying "free will"), it would be live pretty much everything else, distributed across large areas of the brain, and sharing functions with other processes of the brain (like memory, limbic functions, sensory processing, etc...).
This system creates the illusion of intelligence.
This sort of statement is why I generally laugh at the whole field of cogsci and AI. Look up p-zombies. At what point is an illusion not, and if you can't actually tell the difference with any test, how can you ever saying, meaningfully, that it IS actually a mere illusion? I make an AI, a very strong AI, and it acts exactly like a human. 100% indistinguishable from a human mind, to an outside observer. Is this an illusion? How do you find out? Given a Turing test like environment, where you can't judge on surface features, how could you ever tell? Ask it, and it will say it is intelligent (just like you or me), input stimulous, and you get the same output you or me would give.
At this point illusion becomes a meaningless statement, since it is completely unprovable.
I'm not a fan of Strong AI, and doubt it is possible, but these arguments have been pretty much beaten into the ground by now. I hate to say it, but with intelligence all that matters in inputs and output, the rest is a black box. This also ignores the fact that intelligence is a dumb term, completely meaningless when applied to anything non-human. In this case, by using "intelligence" we only mean "human-like", which pretty much means it gives an expected output to a given input.
A patriot must always be ready to defend his country against his government. -edward abbey
The crappy little superficial one-page MIT Technology Review article has a link to another, similarly crappy article on the same site, but if you click through one more layer you actually get to this much more substantial piece in the New Yorker.
Find free books.