The end of AI winter?
machineslikeus.com
machineslikeus.com
Do these things really directly descend from pure AI reasearch? Or are they really the result of a bunch of clever, yet extremely specialized algorithms, independently developed, combined with an incredible increase in hardware power? Not to say such things aren't incredible (Google clearly has changed the world), but is it clear that pure AI research has paved the way?
From an outside perspective it seems the "God of the gaps" argument - AI is what AI researchers haven't done yet - is used as a smokescreen to cover up that AI research hasn't really done much in the last 30 years. (Counterexamples without hand-waving?) And not only that, but it consistently, wildly, and systematically makes incredible predictions that don't come true. For example, Kurzweil is clearly a genius, but also wildly deluded and wrong about his time-frames.
Some of you guys will down-mod me for saying this, and you're the same people who also won't admit my side was correct when, in 20 years, automated translation tools still suck and you'll still be making the same old arguments about how AI is what we haven't done yet... (but in ten years you'll be able to upload your brain!)
Here is what you don't understand: There are very, very, very few people who do research on "AI". People research medical diagnosis, or search, or data mining, or speech recognition, or vision, or chess. If you go to a conference on AI, these are the people you will meet. If you look at who the NSF is funding with their "robust intelligence" area, these are the people you will find. You can only say that "AI research hasn't done much in the last 30 years" if you also say "no one has worked on AI in the last 30 years".
update: if you want to see what people in the mainstream (Kurzweil is not mainstream!) are actually working on, try browsing the IJCAI proceedings:
Surely not "Well, for instance one of the greatest achievements might be that we will work on chess algorithms, and we will see some incremental improvements resulting from tweaking certain heuristics and more intelligent pruning through more specialized algorithms and hardcoded chess knowledge. The programs still won't be able to learn in any interesting sense, but with the help of several orders of magnitude of hardware speed, they will be 200 ELO stronger than the best human!"
http://news.ycombinator.com/user?id=eyudkowsky
His research is in making AGI (artificial general intelligence) not go skynet by building in morals. That seems fairly pure to me.
The point is that people who call themselves AI researches, the vast majority of the time are not working on AGI but on weak AI.
Revolutionary or not, I don't think it was trivial to make an autonomous vehicle.
Uhm, Hans Moravec isn't exactly "anyone". :-)
http://en.wikipedia.org/wiki/Moravec%27s_paradox
Edit: To be clear, Moravec shows that you misrepresent the AI field of that time. But sure, he wasn't the mainstream.
For instance, Machine Translation (as we know it today) was originally inspired by the "noisy channel" for speech recognition that is based on Bayes Rule. In speech recognition, the probability of some words given some input waveform is proportional to the probability of someone saying those words times the probability of hearing some wave form given those rules. The same model led to statistical machine translation: if I speak to you in French, what I'm really doing is speaking to you in English, but the "channel" is so noisy that it comes out sounding like French.
Today, the parallels are less clear (for instance, German and English have substantially different word order, and in speech you don't usually have reordering in the channel--though there's actually some cool new research in MT to bring it closer in line with modern speech processing!).
Yes, there are huge amounts of specialization and hacks for each of these fields, but they are (mostly) based around core good statistical ideas.
In fact, some people are worried that the AI community is so focused on log-linear models and the like that we're in some kind of local minimum, and that we're unlikely to work our way out any time soon.
That said, Kurzweil is still wildly deluded, as you suggest.
Is there a meaningful difference?
There's a good case to be made that the human mind (AKA "natural intelligence") is "a bunch of clever, yet extremely specialized algorithms, independently developed, combined with incredible increase hardware power"
Perhaps the human brain's neural wiring that was required for throwing could have been, and was, subverted into something else. Better negamax alpha-beta algorithms for chess (that require the programmer to know more and more about chess to make any further research progress) will never be useful for anything but chess.
What seems to have died is the dream of any kind of useful generalized intelligence. Any kind!
This means you kind of have to start from scratch as it is not backwards compatible (the idea of a superuser is baked into most archs), which is an enormous amount of effort.
I'm thinking about trying to start a "reboot computing" campaign to get people to think about how we could improve computing if we didn't have backward compatibility to worry about (different security archs, self-maintenance etc).
Check out the fleet architecture while you're at it.
Ever read Steven Pinker's "How the Mind Works" and "The Language Instinct"? He makes good arguments that the human brain doesn't have "generalized intelligence", it has a lot of specific modules, and is less like a single organ for thinking, more like a system of organs that work together.
Ever considered that progress in "generalised" AI may come about when there are enough "specific" AI modules developed that can be joined up?
This looks strange, could you elaborate?
Right now, we seems to be just a few years away from a new age in robotics. They will have some self learning, but at first not be much smarter than insects.
For instance, there are cheap systems that can (roughly) understand what they see. And yes, the robot vision systems are specially built for that -- but the same functionality in animals has afaik also lots of specially built hardware.
Does it really matter if we have to specially build systems, if we can e.g. make system-building-systems as smart tools?
Edit: Some syntax and word choices, etc. Also, on consideration, I make the same point as the GP (StrawberryFrog), but he does it better.
Edit 2: Hmm... Another argument, then: Even if generalized learning will work in practice, it will probably be inferior to networked systems where problems are automatically found and then solved (and updated) from a central location -- like bugs in operating systems. Since everything will be on the net soon, all future generations of robots will probably work like this.
(I don't know how correct it is, but Moravec made the predictions decades ago and they seem to follow the development curves quite well.)
>Do these things really directly descend from pure AI reasearch(sic)?
With the exception of Google's search engine [with whose internals I am not familiar], I can answer an emphatic "Yes". And so would any knowledgeable current or past researcher in those fields. The early AI researchers did a hell of a lot of good work and much of it remains relevant.
As you demonstrate, many if not most people have no idea of what was actually done back then, much less the lineage of their income-tax software or the control system for their digital camera.
Despite funding cuts AI continued to be an interesting and productive field, and remains so today.
There's more, though. Neuroscience continues piecing together how brains work. I've heard that the brain's embodied algorithms are recognizable from eg computer vision research. This seems to imply that there is more of a natural ramp-up from "narrow AI" into "humanlike AI" than at first it would appear.
This presumes a dichotomy between "pure AI research" (whatever that is) and "clever algorithms".
The term Intelligence sets the expectation of "universal learning", not just solving problems we previously thought to be hard. And the research necessary to accomplish that, probably isn't even in the same direction as these fraud AI algorithms. The Biological Computer Laboratory (which died because AI took all the funding) under Foerster probably had a better shot at solving these problems than the AI Lab under Minsky ever had.
This overselling soaked up the funding with empty promises and killed more basic longterm research. Lets hope serious researchers find a way to get their research funded again, despite the AI shills.
I think this is uncalled for. What makes those who worked on some of these AI problems any different from the founders of an unsuccessful startup? Both have a belief that a particular idea/plan will work and both seek to convince others to join/fund them.
Nobody really knew that many AI problems would be so tough. The people who worked on them expected success. Only through their failures did we know for sure that the problems where a lot harder than we thought.
(Sorry, couldn't resist.)
It is not a difference in degree, it is a difference in kind and seems to be unique to higher primates (you can learn a dog to do tricks, but it won't spend its time coming up with new tricks), one of the most interesting science videos I have ever seen is one where a bonobo learns to write a symbol that had been on a computer it used to express its emotions.
What spawns creative thinking and self-awareness is of course a mystery, but I think it's a consequence of creating a complex enough system and creating a system that's built on ordered chaos.
* http://en.wikipedia.org/wiki/Self-awareness#Self-awareness_i...
And it is a difference in degree. A dog frequently comes up with new, previously unseen behavior -- burying certain objects, shredding toilet paper, etc. The problem is that most behaviors are not seen as useful by humans, hence they go unrewarded as "tricks". But there are also examples of dogs learning useful behavior on their own, from learning to bark something that sounds like "I love you", to learning to knock down an owner who is about to have a seizure, to learning to fetch the leash when it wants a walk.
Defining "intelligence" as "doing something that humans can do that animals and computers can't" seems somewhat self-centered (although I admittedly cannot think of a better one off the top of my head).
From what I've gathered, the consensus, both in neuroscience and in philosophy of the mind, is that consciousness is totally emergent from the "simple" building blocks; there's no specific component unique to higher primates that explains the difference.
Universal AI: http://www.hutter1.net/ai/uaibook.htm
Anyone with a desktop computer (or pen and paper, for that matter) can do cutting-edge research in GAI. So it's not popular with the VC's and military-industrial-corporate-welfare-system.
The point is if you're interested, just do the research. You don't need $$$. If you want $$$, over-promise about crappy little web apps, and get funding that way.
This will be a good thing. Theoretical development thrived under the AI “winter”.
I am going to throw the rooster in the hen house and say it: It seems that a large part of AI is thinking up new functions with parameters to be tuned. These parameters are called something exotic and words “neural” and “network” is used liberally.
Doesn't that encompass all of Computer Science?
I'm not being facetious in bringing this up, just attempting to point out that a computerised version of intelligence may not look much like the activities it replaces.
I think Dijkstra said it best "The question of whether computers can think is no more interesting than the question of whether submarines can swim."
Think of an AI machine working with atmospheric data as one "sense" combined with seismic data and some others, with the directed goal of predicting certain types of disasters (tsunamis...?).
Free will and emotions are other assumption we would likely not give these machines, so the worry of self-interest may not exist either, which would aid in making it good at something useful for us.
That doesn't mean that it won't have certain goals, though it remains to be seen whether it will be possible to design a clean goal system with a top-level goal (see also "Friendliness"). Humans clearly do not have this kind of goal system.
(By the way, humans, when asked directly for random numbers, are terrible at the task.)
Free will and emotions are other assumption we would likely not give these machines, so the worry of self-interest may not exist either, which would aid in making it good at something useful for us.
The question is, is it even possible to mimic human intelligence without emotion or free will? Who is to say they aren't wholly dependent on one another?
And if any group of people finds out how to imbue a machine with free will, I'd bet my life they'll go through with it.
An AI would make predictions based on data just like a human would, but the mechanism it used to do it would certainly be different.
Perhaps the parent actually was referring to the qualia component to emotion--or indeed qualia in general--which is much more difficult to explain.
AI has a certain level of expectation associated with it and I do not see google matching that level in any of their released products.
Computers to date do what we tell them to do. As soon as that changes we call it a bug, not a manifestation of intelligence. As long as that view persists I would argue that we have not yet achieved 'AI'.
The Semantic Web is not a good use-case in terms of reputation for modern AI..