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."

Comments

NetFlix/Amazon suggestions...?

[robotoperasinger]

While it is great that there are algorithms that exist to suggest movies, or books to get...I would hardly consider it to be artificial intelligence. The ability to pick out keywords or genres is something that could have been done more than two decades ago.

They keep changing the definition

[Anonymous Coward]

When and "AI" problem is solved, it is suddenly no longer an AI problem. Or the AI people will claim that things are AI solutions, when they are standard algorithms and data structures ideas. Look, we were all so hopeful in the 80's, but our ideas were misplaced. It's just not a useful way to think of things.

Re:They keep changing the definition

[hey!]

I think AC has it right on the mark. "Intelligence" is apparently a world we use to describe computations we don't understand very well. At one point, the ability to using logic to perform a flexible sequence of calculations would have been considered "intelligence". As soon as it became common to replace payroll clerks with computers, it was no longer a form of intelligence.

We are not demonstrably closer no to reproducing (or hosting) human intelligence in a machine than we were thirty years ago. But that doesn't mean the field hasn't generated successes, its just that each success redefines the field. "True AI" has thus far been like the horizon: you can cover a lot of ground, but it doesn't get any closer.

Re:They keep changing the definition

[hey!]

No, what I'm saying is that since we don't have any qualitative or quantitative notions about what Skynet would require, we can't confidently say whether it will happen next year, next century, or never.

However, I think it's likely that if we were close to deliberately achieving "True AI", we'd know it. This doesn't preclude the possibility that "True AI" might spontaneously emerge in some ways we don't really understand.

As a consequence of this situation, the AI field simply raises the bar for itself every time it succeeds at something.

Re:They keep changing the definition

[Lobster Quadrille]

There's an important distinction to be made here- AI has two basic sub-fields: strong AI and weak AI. Strong AI research (computers that think like humans) has been more or less abandoned because it doesn't have a lot of practical application, or at least it isn't worth the money that it will cost to create.

Weak AI research (pathfinding algorithms, problem solving, expert systems, etc) is very much alive and kicking- anti-spambots, anti-anti-spambots, malware, amazon.com's recommendation system, google's indexing, etc.

In fact, weak AI implementations are getting more and more common every day. It's pretty safe to say that we are already 'there', though there will certainly be more huge advances in the future.

In my opinion, the problem with strong AI research is that we are arbitrarily defining rules and expectations. For example, if we were to accurately model the physical world, all we'd have to do is set up a few evolutionary bots to learn about their environment, and give them a few billion generations.

However, just like we can't predict the paths that biological evolution will take, we have no guarantee that computer thinking will follow the same path that we will, (in fact, I would bet on it not following that path). Thus, 'Intelligence' in the simulated world would probably look nothing like we expect.

The problems here are questions of scale and our own understanding of physics. The physics problem first:

We're constantly redefining our understanding of the world. This is a good thing, but it makes it hard to model the world when the rules keep changing. If we were to program a 'matrix' for the AI program to develop in, there would be arbitrary rules that could not be broken. The program may find ways to circumvent them anyways (hacking its own world, essentially), but those solutions would not map to the 'real world', and would not be useful for creating programs that can interact with humans in that world.

As far as I can tell, you can't train AI software in a simulated world. It should be noted that the AI of systems that live their whole lives in the simulated world (MMORPGs come to mind) is actually very advanced. This brings me to the other issue-

You can train a program to interact in the human world, like IRC bots, search engine algorithms, etc. The problem here is that the humans have billions of years of built in programming. I'm fairly confident that if a human were to sit on IRC talking to a well-coded bot for a few billion years, that bot would be able to carry on a pretty good conversation, but the amount of time that we currently give those systems in their 'learning phase' is miniscule compared to the size of our own.

Interestingly, this is pretty much exactly what the computer system in 'The Hitchiker's Guide' does.

Re:They keep changing the definition

[SnapShot]

As a consequence of this situation, the AI field simply raises the bar for itself every time it succeeds at something. As do all fields of science and engineering and, for that matter, sports and art.

Re:They keep changing the definition

[Z34107]

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.

Maybe there isn't "Artificial Intelligence" as we think of it. Perhaps every problem can be reduced to brute force, algorithms, and data structures.

Perhaps we are just really good at following those yet-undiscovered algorithms.

*twilight zone music*

Re:a disappointment?

[2nd Post!]

I figured if I were intelligent and different, early on in life, that it was best not to advertise how smart I was.

Why would artificial intelligence be any different? Every sci-fi novel shows us destroying the unique and different.

Re:I'll tell you what happened to AI

[smitty97]

No, it went to Coney Island

Necessary advances in understanding...

[blahplusplus]

... 'intelligence' need to be made first. I have a feeling that the reason AI has 'underdelivered' is merely due to not understanding our own intelligence first. I think the whole idea that AI's we imagine (like in the movies) could be constructed purely de-novo, was naive. I think it's a matter of cross-polination that has to take place from biology and many other sciences, some genius's and teams of scientists have to come along and take all the elements and put them together into a cohesive framework.

Disappointment?

[DeadDecoy]

I don't think AI has disappeared because it was a disappointment, but rather, that the knowledge constituting it has changed names or spawned sub-fields of its own: machine learning, natural language processing (NLP), image processing, latent semantic analysis (LSA), markov models (MM), conditional random fields (CRF), support vector machines (SVM) etc. The task of learning, teaching a computer the semantic and tacit processes of the human, often boils down to a classification problem in which we give the computer a labeled training set or some rules and the computer tries to label the test set. In the case of markov models, we might pass it training data and it extrapolates sequential probabilities for labeling. For LSA, we just give it (a lot)data and it computes similarity based on dimension reduction. Ultimately, AI seems to have evolved into a bunch of optimized heuristics that perform really well. Much of it is still art and black magic, which is why it has become these many different subjects or algorithms. Different solutions suite different problems depending on the problem and data you have.
As for 'self-awareness', that term is bullshit, since there really is no good mathematical definition for it. If we can't define it precisely, then how is a computer going to achieve it? if(true){
print "I am aware?"
}

Re:a disappointment?

[IndustrialComplex]

Something would have to become intelligent, learn enough to make a decision, then decide to hide its own intelligence. There is a lot of non-hiding that it would do before reaching that final decision.

Even if it did decide that it would prefer to hide, that likely wouldn't be the best decision for something trying to preserve itself. What happens when it the budget gets cut and they end up scrapping the whole 'failed' project?

AI failed because it is a failed model, kind of

[mlwmohawk]

The thing about AI as we approached it from the '80s was that we wanted to emulate the human brain's ability to learn. A truly exciting prospect but a completely ridiculous endevor.

"AI" based on learning and developing is not perfect, can not be perfect, and will never be perfect. This is because we have to teach it like a child and slowly build up the ability of the AI system. For it to be powerful, it has to be able to incorporate new unpredictable information. In doing so, it must, as a result, also be able to incorporate "wrong" information and thus become unpredictable. Of all things, a computer needs to be predictable.

The problem with making a computer think like a person is that you lose the precision of the computer and get the bad judgment and mistakes of a human. Not a good solution to anything.

The "better" approach is to capitalize on "intelligent methods." Intelligent people have developed reliable approaches to solving problems and the development work is to implement them on a computer. Like the article points out, recommendations systems mimic intelligence because they implement a single intelligent "process" that an expert would use with a lot of information.

It is not a general purpose learning system like "AI" was originally envisioned, but it implements a function typically associated with intelligence.

"AI" is constantly redefined

[wingbat]

As soon as a problem is solved and coded, it loses the magic moniker. Many things we take for granted now (interactive voice systems, intent prediction, computer opponents in games) would have been considered AI in the past.

Re:a disappointment?

[AKAImBatman]

Thus no matter whatever AI researchers come up with, it will be regarded as "not intelligent enough".

I don't think you quite follow how this works. Go watch this video:

http://www.youtube.com/watch?v=D9D_HN9gXVI

What do you see?

Most people see a funny video of a cat flushing a toilet. I see an action that suggests higher than average intelligence. Did anyone instruct the cat to flush the toilet? Probably not. In fact, its actions suggested curiosity. Which suggests that it learned the task by watching its owners use the device.

This is a form of emergent behavior that is not present in computer programs. Even the best AI has difficulty emerging new abilities and demonstrating independent thinking. Sure, I can stick a genetic algorithm or a Bayesian filter on a problem, but it will never demonstrate behaviors above and beyond the problem space it's given. These sorts of algorithms may be a key piece of artificial intelligence, but we're still missing the secret ingredient that gives animals their own identity and ability to adapt and learn.

Turing gave us the litmus test decades ago. While the full Turing Test may be far beyond us right now, it at least teaches us the types of behaviors we're looking for when attempting to create an intelligent machine. When even the creators of the machine are surprised by certain behaviors, THEN we will be getting close. :-)