Slashdot Mirror
Is AI Development Moving In the Wrong Direction? (hackaday.com)
szczys writes: Artificial Intelligence is always just around the corner, right? We often see reports that promise near breakthroughs, but time and again they don't come to fruition. The cause of this may be that we're trying to solve the wrong problem. Will Sweatman took a look at the history of AI projects starting with the code-breaking machines of WWII and moving up to modern efforts like IBM's Watson and Google's Inceptionism. His conclusion is that we haven't actually been trying to solve "intelligence" (or at least our concept of intelligence has been wrong). And that with faster computing and larger pools of data the goal has moved toward faster searches rather than true intelligence.
Hubert Dreyfus described most work on AI as being like climbing a tree to get to the moon.
Your tree-climbing teams may report consistent progress, always getting further up the tree as they become better climbers, but they're never going to reach their goal without a radical change in methods.
I think the problem is that people expect machine intelligence to look like human intelligence. Machine intelligence exists and is strong in some areas. Modern chess programs are an example. They can play unique games and be stronger than any human player. Yes, they are given the rules of chess and machines did not invent chess. But they have passed beyond human abilities and it is at the point where some programs are coded to only make move patterns that humans would tend to make. Learning how to adapt machine intelligence to our real world problems is challenging. But we are in for a fright when computers get really good at analyzing human problems and applying better solutions that we now have at hand.
Really, a thermos is the ultimate AI. When I put cold things in one, they stay cold. When I put hot things in one, they stay hot. How does it know?!
If I mod you up, it doesn't necessarily mean I agree with what you've said, sorry.
I like traffic lights.
Define "true intelligence". The more computers advance in doing complex things, the more you will see there is no such thing as true intelligence. You are a very big Turing machine, get over it.
Video of some good progressive thrash music
i've mentioned this before, whenever the phrase "artificial" intelligence comes up. the problem is not with quotes AI quotes, it's with *us humans*. just look at the two words "artificial" combined with the word "intelligence". it is SHEER ARROGANCE on our part to declare that intelligence - such an amazing concept - can be *artificial*. as in "not really real". as in "beneath us". as in "objective and thus subject to our manipulation and enslavement". so until we - humans - stop thinking of intelligence as being "beneath us" and "not real", i don't really see how we can ever actually properly recreate it.
to make the point clearer: all these "tests", it doesn't really matter, because the people doing the assessment have a perspective that doesn't really respect intelligence... so how on earth can they judge that they've actually *detected* intelligence? like the "million monkeys typing shakespeare", the problem is that even if one of the monkeys did actually accidentally type up the complete works of shakespeare, unless there was someone there who was INTELLIGENT ENOUGH to recognise what had happened, the monkey that typed shakespeare's complete works is quite likely to go "oo oo aaah aah", rip up the pages, eat some of them and wipe its backside with the rest, destroying any chance of the successful outcome being *noticed*, even though it actually occurred.
i much prefer the term "machine consciousness". that's where things get properly interesting. the difference between "intelligence" and "consciousness" is SELF-AWARENESS, and it's the key difference between what people are developing NOW and what we see in sci-fi books and films. programs being developed today are trying to simulate INTELLIGENCE. sci-fi books and films feature CONSCIOUS (self-aware) machines.
this lack of discernment in the [programming] scientific community between these two concepts, combined with the inherent arrogance implied by the word "Artificial" in the meme "AI" is why there is so little success in actually achieving any breakthroughs.
but it's actually a lot worse than that. let's say that the scientific community makes a cognitive breakthrough, and starts pursuing the goal of developing "machine consciousness". let's take the previous (million-monkeys) example and step that up, as illustrated with this question:
How can people who are not sufficiently self-aware - conscious of themselves - be expected to (a) DEFINE such that they can (b) RECOGNISE consciousness, such that (c) they can DEVELOP it in the first place?
let's take George Bush (junior) as an example. George Bush is known to have completely destroyed his mind with drink and drugs. he has an I.Q. of around 85 (unlike his father, who had an extra "1" in front of that number). yet he was voted into the world's most powerful office, as President of the United States. the concept of the difference between "intelligence" and "consciousness" is explored in Charles Stross's book, "Singularity Sky". George Bush - despite being elected as President - would actually FAIL the consciousness test adopted by the alien race in "Singularity Sky"!
my point is, therefore, that until we start using the right terms, start developing some humility sufficient to recognise that we could create something GREATER than ourselves, start developing some laws *in advance* to protect machine conscious beings from being tortured, the human race stands very little chance of success in this field.
in short: we need to become conscious *ourselves* before we stand a chance of being able to move forward.
The traits we identify with intelligence in humans (flexible problem-solving, self-consciousness, autonomy based on self-created goals) are all but absent in current Artificial Intelligence techniques, even the ones based on the Connectionist paradigm. Any emergent behaviors appearing in an AI system are ultimately put there by the system builders' fine-tuning of input parameters.
The approaches that show the most promise are those following the "Augmented Intellect" school of thought (the one that brought us the notebook and the hypertext), where a human is put in the center of the complex data analysis system, as an orchestra director coordinating all the activity.
There, intelligence systems are seen as tools at the service of the human master, extending their capabilities to handle much more complex situations and overcoming their natural limits, allowing us to solve larger problems.
By keeping a human in the loop as the ultimate judge of how the system is behaving, any bias inherent in the techniques used to create the AI. It's a symbiotic approach where both human and AI system complement the shortcomings of the other half.
Singularity: a belief in the "God" idea with the "demiurge" relation inverted.
The problem with this argument is Wittgenstein's beetle. I can't even be sure that you are sentient, aware, and able to understand; all I can do is observe your actions and if those actions seem to be consistent with you having what we typically label as a "mind," then I pretty much accept that you have a mind.
We are currently very far away from having machines that can perform general actions consistent with having a mind, except in very artificial and controlled situations (e.g. a chess game, the Jeopardy! game show), but I would hardly say that it will never happen. And if it does, then how can you be sure it doesn't understand things, at least in the same way that I assume that you understand things? If the actions of the machine are the same as the actions of a person (who I believe does understand things), then why wouldn't I say there is a beetle in the box?
The article describes intelligence as the ability to predict, but humans actually experience information, the mechanism of which is still a complete mystery. Another mysterious aspect of human intelligence is that it is able to experience information that is spatially located in different locations in the brain simultaneously. Until we understand how it is able to do these things we'll just be making more complicated Chinese rooms.
The Moore-Murphy Law: The number of things that will go wrong will double every 2 years.
Intelligence in living things can be defined as: A 'formal system' (google it) that maps reality and reacts in a way that sustains the "hardware" that embodies the formal system. - Poor paraphrase of Douglas Hofstadter's ideas.
And did you exchange a walk on part in the war for a lead role in a cage? - Pink Floyd.
Is AI development moving In the wrong direction?
Why do you ask that, Dave?
We don't need servers or robots to have human intelligence. Already have 7 billions of those, including access to superhuman intelligence in the form of of many smart people collaborating with assistance of technology. Also humans have been around forever, and we still suck at human rights. Got to square those away first before having to worry about rights of other intelligent species (and having them respect ours).
What we need now is computers that are good at tasks that we suck at, like repetitive processing on huge amounts of data.
About the only exception is space exploration, where humans are not available for real time remote control due to speed of light. Still, we don't want a Mars probe to get bored and lonely, or make it's own survival the first priority. So cloning our own kind of intelligence, which was shaped by natural selection for self preservation, is not the best approach.
First thing, the article's thesis (according to TFS, which is so ridiculous I couldn't be bothered to read the article) is not only wrong, but completely free of actually having examined what AI research is. At best, it's the product of someone who believes the marketing nonsense promulgated when they tell you your thermostat "uses AI" and that Google search "uses AI."
First, we don't have AI. We have AI research. Anyone actually working in the field knows this (and no, people building search engines and thermostats aren't working in the field.) This is very important to understand. Research yes, but in terms of actual results, the "A" is doing fine, the "I", we simply do not have. At all. Period. This does not, of course, mean that we won't have it. We will. There's no magic here; animal brains are machines, albeit biological ones. Getting from here to AI following that model requires understanding the brain, which we do not, but it is a task we are definitely in the process of accomplishing.
Second, actual AI research at this time includes numerous interesting approaches, all of which are other than those alluded to in TFS. Quite a few of the actual AI research approaches incorporate information taken from the model provided by human and animal brain function at the cellular and network level. For instance, here's something written for the layman that details exactly the kind of brain-based work I'm describing.
Third, there is always the (strong, IMHO) possibility that there is more than one way to produce actual intelligences, and that one of those will bear fruit. The idea that nature has happened upon the only possible solution seems... unduly pessimistic. Having said that, the chosen path for most actual AI researchers (not Google, not the thermostat designer, not the database maven) is to follow the known-working examples that are around us with occurrences in the billions.
The challenge is that the various aspects of intelligence have been very hard to get a handle on right up to just a couple years ago. We have no natural internal mechanism whatsoever to observe the underlying operations that go into creating thought, reason and consciousness. Because of that, it's only been very recently that we have begun to be able to see how this particular system actually operates. With this new information in hand, it finally becomes possible to proceed along lines close to those the relevant biology utilizes by means other than pure guesswork and many-times-removed analogies for observed high level processes.
There are two kinds of AI results being pursued.
The first is intelligent, but non-conscious AI. Which would be an entirely new thing in our world; there are no examples of this in biology to follow. This result, if achievable, will create the opportunity to release us from the necessity of working to survive. This is highly desirable for many obvious reasons. No more menial work just so tomorrow won't unbelievably suck. The house always clean, the yard always in prime condition, willing, able and dependable helpers in any undertaking we choose to pursue, the cat box always pristine, food and other resources are produced and delivered reliably, etc. The number of potential benefits is enormous. Staggering. So there are very concrete and practical reasons to chase this particular goal.
The second is intelligent, conscious AI. Free will, creativity, and so on. The technological goal is clear, but the purpose is, just as you observe, not. We know better (well, we should know better) than to try to enslave conscious beings to our will. The inevitable (and appropriate) result of that kind of short-sighted idiocy is resentment and revolt. Assuming we can avoid that particular mistake, that means they could choose to, or agree to, pursue their existence beside us, which is certainly an
I've fallen off your lawn, and I can't get up.
I think the problem is that people expect machine intelligence to look like human intelligence. Machine intelligence exists and is strong in some areas. Modern chess programs are an example. They can play unique games and be stronger than any human player. Yes, they are given the rules of chess and machines did not invent chess. But they have passed beyond human abilities and it is at the point where some programs are coded to only make move patterns that humans would tend to make. Learning how to adapt machine intelligence to our real world problems is challenging. But we are in for a fright when computers get really good at analyzing human problems and applying better solutions that we now have at hand.
The problem isn't technology, it is people. We already have human intelligence and it is relatively cheap to procure. They are called people... you know actual human beings. Hire one. Make a baby. Go find an actual friend.
Try selling a product or service based on blank slate human intelligence. Sure there are aspects of the human brain that we are eager to replicate, simulate and make into a reproducible machine, such as image recognition or some other pattern recognition... But the marketability of human intelligence is upwardly bound by the availability of actual humans to do whatever it is you want them to do.
And in most cases you aren't going to want a generalized intelligence for a specific set of tasks, you want specialized and reproducible programming. Like the chess example, you want the chess computer to play chess and not day dream about some far off place while they are losing the game. Generalized intelligence slows down the processing. Generalized intelligence will keep trying to learn even after reaching an optimally efficient state or just get bored and go learn something else. And any semblance of uniqueness or "original thinking" requires a variety of experiences and slow development and redevelopment of neural pathways and even sometimes error and imprecision which are very much undesirable traits in technology.
Think of how many years it takes to teach a child. We spend literally decades training and retraining the human mind to know things and think about things. Who would buy that kind of technology that takes a decade or two before it is ready to do something useful?
Until the cost of simulating a human brain comes down to something that can be done on a very small budget, just because, then it won't get funded. And even then we aren't just a brain, we are a brain connected to a body with a full set of human senses and biologically driven needs. If you want human intelligence you need to have a human experience.
And even then would it even be ethical to start up a simulation of child's brain, teach it, train it, give it emotional response and physical form? Just because you can? Because you want a companion you can control? Because you want genius that you can turn off? At some point you have to realize you are playing God out of excessive pride and not furthering any good. Come up with a use case where you want to simulate a complete person, create a person, where an actual person just won't work.
I think the only reason to do so, ethically, would be for space exploration or working in other environments where an android would not be harmed and could survive and perform some useful tasks. But once you recreate human intelligence in android form, then you need to give that creation respect, status, some form of equality and free will. Or else you are not adding value to society, but undermining our values.
i've mentioned this before, whenever the phrase "artificial" intelligence comes up. the problem is not with quotes AI quotes, it's with *us humans*. just look at the two words "artificial" combined with the word "intelligence". it is SHEER ARROGANCE on our part to declare that intelligence - such an amazing concept - can be *artificial*. as in "not really real". as in "beneath us". as in "objective and thus subject to our manipulation and enslavement".
I would have to dispute your definition of artificial as being somehow "not really real". If you use the original meaning, ie. the product of an artisan, or a crafted object, then it makes complete sense. We are talking about intelligence that is designed and built, rather than having developed naturally. Artifacts are still real things.
Hello,
Interesting post. I just wanted to make a point about the existence of general intelligence: it turns out that the human brain is actually a stack of many "special purpose" computational systems. That doesn't mean that there isn't a general intelligence portion as well, but we're DEFINITELY composed, at least to some degree, of stacks of special skills.
Examples:
1) Vision and object recognition. There's a whole subsystem of the brain dedicated to decoding light signals into a representation your consciousness can use. There's even a special subsystem for recognizing faces--they even know its location in the brain.
2) Audio: similar to vision, there's specialized decoding brain circuits.
Those are the two biggies, but we also have special hardware for processing/controlling speech, spatial reasoning, body control, and others. What's more, there are people who have *developed* special purpose brain circuitry for playing the violin, for example, and savant-like mathematical computation. For people who have done that, it is as easy to do a square root to N digits as it is for you and me to walk.
Because of that, it's NOT clear to me that a general purpose intelligence can be made without assembling a sufficient number of special-purpose intelligence. It's NOT clear to me, in fact, that there are unknown forms of special purpose intelligence that humans are lacking that wouldn't transform our general intelligence. (People are prone to making certain logical errors, even the brightest of us, because of in-born holes in our mentalities!)
A dolphin might look at us as crippled mentalities because we can't construct a spatial model of our surroundings from sound, for example. What other mental abilities COULD exist, that we don't have, that could expand our mental potential in outrageously powerful ways? People typically aren't able to fork their consciousnesses into solving two problems at once independently, there's one I'd like!
But the point I'm trying to make is that the stacking approach might be NECESSARY to compose a mind capable of general intelligence that we'd recognize. It might not need ALL our special purpose skills, but it's not obvious to me that a composition isn't necessary.
--PM
I suppose if you're approaching the problem with an Edison-like philosophy (no such thing as a failure) then sure, and I didn't mean to imply that the 'work' being done is producing uselessness, but it's not producing true AI, either, just things that give the illusion of AI.
Are YOU using the TOOL, or is the TOOL using YOU? Think about it!