Artificial-intelligence research revives its old ambitions
web.mit.edu
In my opinion, hard AI could be just around the corner. Wonder what kinds of startups would rise up to leverage such a technology that is so natively different.
web.mit.edu
The problem is we have no theory of intelligence, no theory of psychology. Research in the cognitive fields is fractured, all about tiny insignificant phenomena with little relation to anything else. Our best theory is "the brain is like a computer" which is, frankly, a terrible theory.
Here's something I find more promising: On Intelligence From First Principles: Guidelines for Inquiry Into the Hypothesis of Physical Intelligence [1]
In short, what we really need to understand is self-organization and non-equilibrium thermodynamics. Not image labeling.
[1] http://www.tandfonline.com/doi/pdf/10.1080/10407413.2012.645...
Goedel machines might actually be the closest thing to this in the computing literature, my reaction is less about the work itself than the rhetoric surrounding it, to be honest. JS should collaborate with a physicist on the thermodynamic side of the problem.
If you're intrigued, you could start with the article I linked above, or if you have journal access, anything from the same special issue. I chose that article just because it's the only one not behind a paywall.
I don't think we can have a theory of intelligence. At least in the public consciousness, intelligence is one of those "God-of-the-gaps" style concepts that continually evolves in order to maintain the illusion of human superiority.
But I do agree with your sentiment as far as the way intelligence is usually discussed, even among the science-literate.
But then again I'm more interested in the intelligence that differentiates a slime mold from a hurricane than the intelligence that differentiates a human from a chimpanzee.
For example: hurricanes are self-organized, constituted by a structured flow of energy and matter rather than specific pieces of matter. But a hurricane is a slave to the local potential. It will dissipate all the negentropy in its wake, and in doing so maintain its structure. But once there is no more energy differential to dissipate, the hurricane will itself dissipate as it is not able to break free of the local potential and use information to seek out non-local negentropy sources. The question for research is what is necessary to make that jump from self-organization to intelligence, given that operationalization.
Likewise for seeds, fish eggs (carried in the gut of birds), etc.
Also: my definition is meant to include fish and birds, even plants, as intelligent.
How does your comment about sparks address the alternatively stated requirement: "because it perceives more fuel in that direction"
* A spark flies out randomly and contacts a fuel source
* A blind person reaches out randomly and finds a glass of water
What is the essential difference between these events?
In any case my post above implied a definition for perception: responding to low-energy distributions with an asymmetric high-energy response.
Some rough and imperfect, but still useful, definitions of intelligence could be the ability to make good predictions based on past data, the ability to solve optimization problems well, and learning ability.
OTOH, looks like not very much was actually formalized.
What sort of test can show that a subject demonstrates an understanding of abstract concepts?
So far, from what I've seen, if a test can be written then software can be written to solve the test.
I don't understand why this is an issue though. Testing intelligence was never the hard part of AI. There are so many tasks that computers currently suck at that we would be happy if they were solved, regardless what label you gave the solution. And I don't think many people could see a computer doing tasks like having conversations or solving difficult problems and deny that it is intelligence. Even if there is no formal test to perform that is 100% certain.
How so? To me, it appears completely open-ended. You could sit there forever asking questions and never reach a definitive result.
Once we figure out all the issues of general intelligence, it will stop seeming so special. We may even begin to think that humans are really bad at it afterall.
Because of this there are a number of games that computers still can't beat because just stupidly trying every possible move doesn't work like it does for chess.
Watson actually does use a lot of natural language processing and machine learning so it is kind of intelligent. Though at it's core it's still just a glorified search engine. Jeopardy was always just a game of memorizing facts, not a demonstration of intelligence.
What people did predict wrong is that it would take general intelligence to solve chess. As in, if you solved chess, you could also pass the Turing test and everything else. Here is a quote from Douglas Hofstadter:
>There may be programs which can beat anyone at chess, but they will not be exclusively chess players. They will be programs of general intelligence, and they will be just as temperamental as people. "Do you want to play chess?" "No, I'm bored with chess. Let's talk about poetry." That may be the kind of dialogue you could have with a program that could beat everyone.
And they would have been right if computers hadn't become exponentially faster.
Currently, we've got machine learning--I think there's a market for it, but it's early. I think the near term is about automating data science tasks--we won't have a shortage of data scientists because we can automate much of what they do. Even that's a hard problem. Personally, I think deep learning (specifically relatively recent neural net related research) will play an important part in this--as far as automation goes, it will help automate feature engineering (currently a huge cost).
In general, I agree with the sentiment in the comments that "we are a long way from AI". Objectively, we are. However, I think we were similarly a long way from flight when the Wright brothers figured out how to keep a plane in the air. Or towards the end of the human genome project--although at least in that case, we knew what we were doing, whereas AI still doesn't really have a good theory behind it. The first software that seems awfully "AI like" though--I'm thinking that it could very well appear by accident while working on more general machine learning tasks. I don't think we're as far off as most people think, but I certainly don't have a crystal ball (and neither does anyone else)
The genome project is another science to engineering transition problem, where the science check mark is done, but as far as I know there are no widespread engineering applications for the data at all, although someday there might be in the generic sense that all research is potentially useful.
Something like this could happen in AI. OK the science proves conclusively that it could work, its just human engineering might not be up to it for a few centuries (like the Babbage experience WRT computation) or even the engineering is possible but the economics and politics make it impossible (like building a massive power generation dam across the straits of Gibraltar) We could prove the science of AI works, and then nothing happens for the next 500 years. Or maybe not.
In fact, there have been a handful of folks thinking and working on Strong AI/Artificial General Intelligence steadily for years. Ben Goertzel even started the conference [1] and journal [2] in 2008.
One of the biggest things I see in every one of these conversations and threads is everyone assuming humans are actually good at things, as opposed to just marginally better than most other machines or humans. So people's standards for machine intelligence are way above what they would expect from human standards.
Google translate is a perfect example. People were complaining about it here on HN the other day because it wasn't perfect, however in comparison to any group of average professional linguists (Some of whom I work with daily) Google Translate and some of the other machine translation services' accuracy for it's speed is light years ahead.
People also forget that "Strong AI" and "AI complete" is always a moving target for benchmarking - and there is no benchmarking standard (No, a turing test isn't a robust enough test for an AGI). I posit that humans will never truly accept an AI as smarter than us until it dominates us. It will always be - well yes it can X, but is it really thinking? Does it have consciousness? Great questions philosophically, but practically it really doesn't matter.
This is good though, I think if there is anything the human race should be working on it's this. Everything, and I mean EVERYTHING pales in comparison in my opinion.
[1] http://www.agi-conference.org/2013/ [2] http://www.degruyter.com/view/j/jagi
A related HN post from a few weeks ago: https://news.ycombinator.com/item?id=6605015
(Ai?)
But I forgot the exact reference (http://en.wikipedia.org/wiki/Dartmouth_Conferences) so I will now edit my original post to include the right numbers...
On the other hand, I have a hard time thinking that anyone's going to treat virtual reality as all that cool after having seen The Matrix. Our culture now perceives "virtual" as something you're trapped inside, hiding you from the "real", rather than as a pleasant place to spend an afternoon shooting zombies.
It's going to be as usual: first hackers get it, then geeks, then it becomes a mass pop-phenomenon.
The interesting question is: where is that upper limit on intelligence-per-FLOPS?
EDIT: Fixed it to "FLOPS". We already have AI algorithms that Solve The Problem if given unlimited "thinking time" between steps of interaction with their environment (up to the point of needing a Halting Oracle), and we have extensions of those that are asymptotically speed-optimal modulo some additive or multiplicative constants (which are larger than our current universe). I strongly suspect that scaling down these Idealized AI Agents to real, useful programs isn't going to just involve feeding raw instances of the Ideal Agents small enough environments to be tractable, but actually finding or deriving some algorithm that can scale continuously up and down with available computing power per unit of real time, where the time units are the actual real-time length of an interaction cycle.
To have a white-box learner would imply that humans were smart enough to build it. Yet it's unlikely that humans can build something significantly smarter than a human, so this AI would have the same limitation: It's unable (not smart enough) to build a smarter AI.
For instance, if you have infinite computing power, you just have to dovetail over every possible Turing Machine until one of them "wakes up" and does AI-y things. Yes, there are in fact real approaches based on a more sophisticated version of such a Universal Search.
In fact, we routinely make use of machine-learning algorithms that are "smarter than a human" in particular narrow domains. Are they smarter than a human at general intelligence, of the kind necessary to start looking more like a scifi "AI" than a cheap trick? No, but they are an indication that such a thing could be possible.
Overall, I would say that with Machine Learning maturing and Universal AI being a formal science at all, we are only just now beginning to find out where the actual tradeoffs lie in the design and construction of intelligent agents.
The speed of intelligence scales. But not the quality of it.
The quality of intelligence is something very hard to quantify. But as simply as I can state it, it's the quality of coming up with an idea that is not just an extension of an existing one, but an idea that came seemingly from nowhere. To someone else the idea appears magic, and no amount of thinking by that other person can lead him in a path that will end up there.
> but they are an indication that such a thing could be possible.
Real genius is not just "more" intelligence, it's intelligence of a different type that a non-genius simply can not do, no matter how they try. Analogously, just because a computer can do something fast does not give it that hard to define quality of true intelligence.
To make (rather than randomly search for) a computer with real intelligence we need to understand what it is, and where it comes from, and we don't know either one of those things. Our best (smartest) computers are all variations of "search randomly" (for example chess engines). Those can do some amazing things, but their limitations will not be overcome by doing more of the same.
> infinite computing power .. every possible
Talking about infinity is a total red herring and you know that - especially because if your method of increasing intelligence is "search randomly" then an AI will not be able to make a better AI - all it can do is continue the random search humans already started.
Well the good approach that requires a halting oracle is AIXI, which uses semi-approximable (IIRC on the terminology) Solomonoff Induction hooked up to an expectimax tree.
So it does have to examine every possible "universe" as a Turing Machine, but it uses proven-convergent probabilistic reasoning to get rid of the bad ones and raise good hypotheses to the top much faster than a universal search.
>The speed of intelligence scales. But not the quality of it. > >The quality of intelligence is something very hard to quantify. But as simply as I can state it, it's the quality of coming up with an idea that is not just an extension of an existing one, but an idea that came seemingly from nowhere. To someone else the idea appears magic, and no amount of thinking by that other person can lead him in a path that will end up there.
To be frank, before I'll believe this I want to see studies done on either humans or a mathematical examination of formal models of intelligent agents.
I think "never" is much too strong a word, considering that we seem to have a huge abundance of people capable of learning calculus even though few can rediscover it from scratch. There appear to exist very many people who would require such great efforts to learn certain things that doing so would be unpractical, but very few who simply cannot learn things no matter how much effort they put in.
You can find one smart person. But how do you find another smart person for them to work with? With AI-level smarts, you reduce the coordination problem to zero since it can spawn multiple copies of its own brain state (maybe?) and work concurrently and intelligently on the same problem. Just ignore the problem of killing off divergent brain states once the task is complete (nobody cares about the life of a kage bunshin).
I sure could get a lot more done if I had five more of me (unpaid, of course) to tackle a problem all at once.
The difference is that it can use them not for sending low-return spam, but for multiplying its power - even if it doesn't give a qualitative difference, there should be a quantitative difference (how many different tasks it can do at once) when a strong AI gains many orders of magnitude more computing power than any laboratory/datacenter where it was initially created.
There's an assumption in most AI discussions that AI "must" be embarrassingly parallelizable. For example, much as my brain is parallelized, surely an AI would have to be the same way, much as all heavier than air human flight required was a faithful reproduction of bird anatomy (LOL).
There's also a certain assumption that intelligence implies a level of self awareness that most humans don't have therefore AI would be highly self aware...
There's also a whole industry of sophistry devoted to proving no humans are intelligent and IQ as a concept or any other numerical measure of intelligence does not exist and there is no way to compare intelligence, and right or wrong those folks will surely make life difficult for people improving / testing / upgrading an AI, either because they're right or they'll be making political protests.
But is it the right kind of computing power? The human brain is a very different sort of machine, architecturally speaking, from any computer currently in existence. Current computers have a very difficult time dealing with massive, parallel, random data access. Perhaps strong AI requires billions of threads randomly accessing a database which is terabytes in size? That is going to make the memory bus very, very sad.