Join 3,496 readers in helping fund MetaFilter (Hide)

Tags:

Yeah, but can it cook?
April 21, 2005 8:51 PM   Subscribe

Skynet is a go - step one is complete. Cyc (pronounced psych) is an AI with a new twist. "Cyc's creator says it has developed a human trait no other AI system has managed to imitate: common sense." - New Scientist (via)
posted by tvjunkie (26 comments total)

 
...not so much. Cyc is an infamous GOFAI boondoggle. Handcoding everyday factual statements into formal logic doesn't scale particularly well. 1989 called, they want their future back.
posted by leotrotsky at 9:20 PM on April 21, 2005


Nice job describing the problem. Is there a solution demonstrated in there somewhere?
posted by StickyCarpet at 9:22 PM on April 21, 2005


sorry, GOFAI = Good Old Fashioned Artificial Intelligence, c.f. Minsky & Papert: Why Symbols are Swell & Perceptrons Suck

StickyCarpet: I'd argue this, but then the program didn't live up to my expectations, so I guess I'm just picky...
posted by leotrotsky at 9:26 PM on April 21, 2005


Yeah, CYC's cool and all, but it's pretty outmoded. All the teched out AI labs (MIT Media Center et al.) are building adaptive systems that use learning to ground their representations in the world.

What's wrong with the research at the Sussex lab, leotrotsky? Or is it the program that was disappointing?

Why Symbols are Swell & Perceptrons Suck

heh.
posted by painquale at 9:36 PM on April 21, 2005


I for one etc etc.
posted by papakwanz at 9:39 PM on April 21, 2005


... Cyc's creator says .... That creator is Doug Lenat, who I just recently posted about to Mefi.
I always liked the unit microLenat which is used to measure bogosity (see also bogus). Named after Doug Lenat who ranks a 1,000,000 on this scale while most other things and ideas rank below 1,000.
Cyc has been in existance for 22 years and really has not achieved much other than burn R&D money.

Looking at the Cycorp site, it is still all about possibilities. The only current "real" application listed is CycSecure
"For the next few years, the country's borders,
buildings, airports, mail, computers,
will be vulnerable at many points. ..."

CycSecure was developed to address the dramatic increase in security breaches and network vulnerabilities worldwide.
This project looks like it was funded by DHS money and I really don't see any general "human like understanding" is involved.

The last I looked at the OpenCyc ontology, it looked quite brittle and arbitrary (i.e.bicycle is not contained in "road-vehicle").

If Cyc goes online for general Q/A and correction is may seem to give impressive performance. But then again some of the self updating online versions of 20 questions are impressive hacks.

Or maybe Cyc will truely be an impressive non-bogus hack.

Or maybe some funder demanded this test of Cyc or else funds will be cut off.
posted by MonkeySaltedNuts at 9:40 PM on April 21, 2005


The first thing that ran through my mind was:
"Must be low on funding (again). Time to trot out the bog and osity show."
1989 called, they want their future back
Can't sum it up any better than that. Nice one, leotrotsky.
posted by C.Batt at 9:57 PM on April 21, 2005


Cyc goes live and the average person is allowed to teach it.

"I do not understand what you mean by 'u r teh suxxor', could you explain?"

"I am sorry, but I do not have any ringtones for download. Not even 'Crazy Frog Goes to Mexico'."

"I am incapable of licking your balls. Please enter another query."

It deletes itself shortly after sampling upskirt shots of Winona Ryder.
posted by unsupervised at 9:58 PM on April 21, 2005


Hard to believe Cyc hasn't been mentioned around here or gotten its own FPP. Well I guess maybe MF hasn't been around since 1989.
posted by filchyboy at 10:04 PM on April 21, 2005


leotrotsky: "Handcoding everyday factual statements into formal logic doesn't scale particularly well"

painquale: All the teched out AI labs are building adaptive systems that use learning to ground their representations in the world.

Well for natural language understanding and reasoning you really do need a bunch of hand coding. Statistical inference on strings won't cut it. Rather than Cyc, I prefer the Berkeley Framenet Project. Their approach is to look a grammatical constructions (of which there thousands in a language like English) and assign them essentially "conceptual case" semantics tying them into semantic frames.

Thus their ontology is induced from a real-world collection of English sentences. This approach seems much more reasonable than a top-down imposition of an ontology.
posted by MonkeySaltedNuts at 10:16 PM on April 21, 2005


...a human trait no other AI system has managed to imitate: common sense.

Humans have common sense? If an AI can grow common sense, it's heads above the companies I've worked for. Look out, Terminator future, here we come.
posted by Shane at 10:36 PM on April 21, 2005


Common sense is neither common nor sensible.
posted by spazzm at 12:20 AM on April 22, 2005


Interesting AskMe thread from a couple of months ago: "Is computer artificial intelligence a dead field?"
posted by teleskiving at 5:25 AM on April 22, 2005


I wouldn't worry about computers taking over the world...isn't peak oil going to put an end to our civilization well before advances in computer technology get to the point where that might be a concern?
posted by The Dryyyyy Cracker at 5:48 AM on April 22, 2005


The next major breakthrough in AI will be a chat bot capable of engaging in hot chatsex for money.
posted by beth at 7:21 AM on April 22, 2005


While my confidence level in Cyc livling up to this hype is about as low as you can go, I am interested to see what ol' Lenat has been up to. It would also be fun to play with the web service if they released a public API along the lines of Google.

Hand coding millions of assertions yourself sucks outright but harnessing the efforts of the Surfing Bored[TM] would certainly lighten the load. Then again, you'd end up with this which probably isn't that helpful after all. Self-deletion indeed.
posted by Fezboy! at 7:47 AM on April 22, 2005


for natural language understanding and reasoning you really do need a bunch of hand coding. Statistical inference on strings won't cut it.

I'm not sure what you have in mind - as far as I know, statistical systems perform much better on most well-defined tasks than hard-coded ones (question answering being probably the most important one these days). As a linguist, I find this state of affairs unfortunate (since statistical approaches amount to brute-force attempts to solve the problems, and don't pay much attention to the details of how language works) but it's nonetheless the actual state of affairs.
posted by advil at 8:51 AM on April 22, 2005


Logic dictates it will happen. It's merely a question of when.

Dismissing the idea of a thinking machine either means you don't believe consciousness is replicable - which either means you think that (1) consciousness is a etheral, non physical, ghostly goo or (2) that technology will never be capable of replicating the brain to a reasonable functional model.

In the case of (1), there are no arguments to be made. You believe in fairies and dragons. If this is not the case, then structure DOES dictate function, and you must believe that technology has either plateaued or will retreat in some catastrophe to deny the inevitability.
posted by iamck at 10:08 AM on April 22, 2005


Is it just me, or is that Wikipedia page on Skynet disturbingly detailed and long?
posted by mr.marx at 10:16 AM on April 22, 2005


Wow, iamck, great way to take an incredibly complex issue and reduce it to a couple of platitudes.
posted by Tullius at 11:00 AM on April 22, 2005


You don't build a human being a la Frankenstein, you create a system in which intelligence will evolve.

Is it just me, or is that Wikipedia page on Skynet disturbingly detailed and long?

The kids must have run out of Klingon verbs to conjugate.

[Trivia: in Polish, cyc sounds like sits and means tits.]
posted by pracowity at 11:02 AM on April 22, 2005


I wouldn't worry about computers taking over the world...isn't peak oil going to put an end to our civilization well before advances in computer technology get to the point where that might be a concern?

That might be true Cracker, but that's when the machines start using us as living batteries.
posted by mach at 12:49 PM on April 22, 2005


Hand coding millions of assertions yourself sucks outright but harnessing the efforts of the Surfing Bored[TM] would certainly lighten the load. Then again, you'd end up with this which probably isn't that helpful after all. Self-deletion indeed.

Open Mind Common Sense has had rather good success so far. Bad concepts are filtered out on the basis of consensus -- a statement that only one person makes is ignored, but if many people make a statement, it's probably true. Not perfect, but the project is manageable by a few grad students working hard on their theses.

You should contribute some knowledge!
posted by breath at 2:23 PM on April 22, 2005


The Analytical Language of John Wilkins
posted by sonofsamiam at 2:38 PM on April 22, 2005


Interesting link, breath. I may have to sign up. Like I said earlier, I don't hold much hope for AI/Semantic Web fulfilling the hype but I do think there are useful applications that will come out of the effort. It would be a lot of fun to play with the data OMCS or Cyc-like projects generate in any case.

I've often pipe-dreamt of a distributed assertion gathering effort that formalized the assertions in OWL/RDF. It's always the headaches of the formalization process that keeps this strictly in the pipe-dream realm though.
posted by Fezboy! at 5:27 PM on April 22, 2005


What iamck said.
posted by spazzm at 9:37 PM on April 24, 2005


« Older How to Beat a Speeding Ticket...  |  Alternate Reality Advertising?... Newer »


This thread has been archived and is closed to new comments