Slashdot Mirror
DARPA Contracts For AI Technology
heptapod writes "USA Today is reporting that DARPA has contracted two professors from RPI to develop artificial intelligences that can learn by reading and understanding natural language. Interesting taking DARPA's Grand Challenge into account. Mentioned in the article is Cycorp, Inc. which has been pursuing this goal since 1994!"
"First passing of the Turing test!"
DARPA announced today the funding for Skynet.
Teaching a machine to read a text book and answer questions doesn't necessarily mean cognitive reasoning. It's just a new form of input/output. Ask it to write an essay with a definative argument and solid conclusion based on the material read would impress me, not regurgitating facts and figures found in a book.
"It's difficult to meditate on amphetamines." - Joe Walsh
I have enough dynamite to blow up 10 super-computers!!!
...artificial intelligences that can learn by reading and understanding natural language...
OK, but can it learn from mistakes?
OpenCYC.org project Sourceforge CVS repository has not beent updated since October 22nd 2003. I hope some of that DARPA money will go a little way towards completeting the 1.0 release.
as you know we non americans cannot access darpa.mil :)
:)
If something is kind enough to give us a mirror to the "Great Challenge", kudos to him
Or else I'll go through a US proxy. Not a big task, it's just annoying, I'll do that later.. grab an anonymous US proxy on www.proxy4free.com , enter the crap in your browser and enjoy the slowness. Maybe I'll use switch proxy this time
Can it read books on AI; and then design a better, smarter AI?
I am the unwilling control for my Origin.
For all of the thousands of times I've read the phrase "artificial intelligence," it's only recently occurred to me to wonder whether there's a point in using "artificial." Certainly the first flavors of this are at best insect-like, or sort of idiot-savant (like chess playing), but when we first experience a system that's as awake as we're all hoping for... then it's just "intelligence," isn't it?
I know - read four thousand sci fi novels and then come back to this conversation... but it seems that the "artificial" of this phrase is increasingly awkward. It makes some people dismissive about the potential, other people feaky about the same, and seems destined to always shortcut the philosophical payload. Not because I fret over the machine's eventual feelings (though if it's Linux-based, I'm sure it will have very warm, friendly, altruistic feelings), but because by boxing code-based intelligence into the "artificial" category, it props up the more mystical perception of our own native smarts.
The very word, from "artiface," suggests that whatever it will be, it won't really count as intelligence. But we're very comfortable (or at least I am) talking about, say, an intelligent dog or primate. So, if we can even approach that with a system that isn't any more fragile than walking, breathing meat... then surely that's not artiface? OK, smack me around now. Thanks.
Don't disappoint your bird dog. Go to the range.
I dunno, but to me it seems like there are two different ways to go about this.
One is the way that CycCorp is going which is to create a giant knowledgebase and feed the AI tons and tons of data. Eventually just by the fact that is has so much data, it can become semi intelligent.
Another way, would probablly be to actually have the AI interact with the enviroment and learn by doing. Even in this case though, it would still be preferable for the AI to have a knowledgebase it could look into to find general information. Just like how humans have the internet and books to read up on when they want to learn something new that might be prohibitive to actually do in real life.
For example, say you want to be intelligent about nuclear reactors. You have two choices
You could, build a nuclear reactor
You could also read up a lot of information on nuclear reactors
In most cases you will end up reading up information on the reactor, maybe if your lucky you can actually design one and work on it. But just because you haven't built one and only read up on them thoroughly does not make you any less "intelligent" in that field.
I guess what im trying to say is, why reinvent the wheel. Why have the AI try to learn everything the way humans have?
Each class should have its own championship title of $1 million, with each runner-up winning 1/2 the money of the next higher.
Each contestant must provide 2 systems -- a compressor and a decompressor. DARPA feeds the compressor the corpus and the compressor feeds DARPA the compressed corpus. DARPA then measures the ratio and feeds the decompressor system the compressed corpus, which then returns the original corpus, or is disqualified. Compression and decompression times must add up to no more than the time limit for the competition class.
The rationale for this approach to advancing the state of AI is given by a short paper by Matthew Mahoney titled "Text Compression as a Test for Artificial Intelligence" (1 page poster, compressed Postscript) published in the 1999 AAAI Proceedings. Matt Mahoney shows that text prediction or compression is a stricter test for AI than the Turing test.
So far there have been lots of promises and decades spent. Let's try something different with well-founded objetive metrics tied to serious near-term commercial incentives for evolutionary progress.
Seastead this.
It's a $400k grant with two optional extensions. The school will take half, the profs will take part of their own salaries out of it, and then it'll support a couple PhD and MS students. This is no big deal.
AI has always "failed" because every time it's succeeded, the problem it succeeded on has been retroactively defined to "not require intelligence". Cf. automated theorem proving, chess playing, control of chemical plants, and just about any other AI success of 1940s - present.
10 PRINT CHR$(205.5+RND(1)); : GOTO 10
It's not just canned questions and answers; it has an inference engine. It can do "if A is B and B is C, then A is C". But only if all the right predicates match perfectly.
Lenat was claming it would somehow become intelligent in a few more years. That was a decade ago. Today, Cyc is regarded as the definitive demonstration that that idea won't work.
Here's a critique of Cyc from 1994.
The parent is suggesting that artificial selection is proof against natural selection.
And you can't breed dogs or horses or humans or anything else to enhance a specific trait can you?
The fact of the matter is that we are fundamentally no different from the amoeboid life we evolved from, and the rest of the life that evolved from it, just more complicated. If simple insectoid neuro circuitry can be approximated with simple neural nets (read this for more info on this highly debated subject) it could easily be argued that it is not the distinction between artificial and "natural" intelligence that should be question/examined but the existence and definition intelligence itself, and quite possibly life for that matter. These are concepts as arbitrary and ill-defined as the spirituality that their nay-sayers flaunt so wantonly in protest.
For christ's sake (pun and capitalization intended), think before you flap your rot. (There's just no escaping them on this subject)
Of blankness, I know nothing.