OpenCog isn't worth any attention, imnsho. It basically takes the "if we glue together enough parts that intuitively seem relevant to AGI, maybe AGI will pop out!" approach. There's no convincing defense of its design as far as I can see, just unconvincing heuristic hand-waving. (I respect the guys behind it, but completely disagree with them on the approach.)
"if we glue together enough parts that intuitively seem relevant to AGI, maybe AGI will pop out!" approach.
That sounds like the way I approach most problems I don't know the answer to. I fiddle with various relevant parts, and see what pops out. My first version is rarely a good one, but it leads (always has so far) to version 2 which works well.
Research wouldn't be necessary if we knew beforehand what we needed to look for. Granted, it doesn't help to glue things together indiscriminately, but I doubt that is that OpenCog is doing.
Someone has to try this and if their idea doesn't work, I will bet good money that they'll stumble upon other interesting results (if they keep going).
Research is of course necessary and warranted. And while they aren't being entirely "indiscriminate", it's not nearly discriminating enough from what I can tell. AGI currently needs research at fundamentals, at the theoretical foundations of it. It hasn't been reduced to an engineering problem yet.
As one of the more recent, visible efforts at general AI, OpenCog is automatically worth some attention (in fact, you've given it some..).
I agree that Open cog will fail but I think it's failure should be looked at.
"if we glue together enough parts that intuitively seem relevant to AGI, maybe AGI will pop out!" may sound dumb but sooner or later, an "intelligence" program needs to reach a boot-strapping stage so it is actually credible that there is some algorithm with the properties that if you add enough processing power to it, intelligence will indeed "pop" out.
Many believe the Mammalian neocortex essentially involves such a generic algorithm. But the problem is there are fairly good arguments that the kinds of algorithms we've tried so-far could include not this algorithm. They are not algorithms that can reach the "critical mass" level.
The first failure of AI was based on the now clearly-false belief that intelligent behavior was essentially logical deduction - these efforts main appeared in the 1970's. Shredlu was the most successful.
The second failure of AI seems to be the belief that intelligent behavior involves primarily something like statistical inference. Narrow "machine learning", OpenCog, and Jeff Hawkins stuff all seem to fall out of this later approach.
My sense is that the world humans live in is too chaotic for pure logic yet much more orderly than phenemena suited to pure statistical inference ("machine learning" can work but its failure is obvious in excessive time it take make "obvious" (to us) inferences). The most uniquely human intelligence involves neither making complex deductions nor filtering scattered and chaotic masses of data but rather putting "semi-structured" data together in a "common sense" fashion. Humans follow patterns immediately, not after seeing them over-and-over again. Humans are "terrible" at pure calculation but also "terrible" at pure inference problems like gambling games (and the way human fail at gambling is very instructive - we tend to see too many patterns in random data. )
The challenge is how to create a "third way" beyond these two failed approaches (keeping in mind that both approaches can have successes too).
I'll leave my grandiose announcement of this algorithm for a different post.
On the contrary, there IS a detailed, well-thought-out defense of the OpenCog design, but it's sufficiently complicated that the casual observer may not be able to appreciate it. Wait for the book "Building Better Minds" (http://opencog.org/wiki/Building_Better_Minds).. to be available later this year... for a detailed treatment of the theoretical foundations of OpenCog. Or for a few clues see http://goertzel.org/dynapsyc/2009/general_ai.htm and the references therein. For a crisp write-up of many of the ideas, and a comparison to other AGI approaches, see the proposal http://www.goertzel.org/CogBot_public.pdf . -- Ben Goertzel, leader of OpenCog ...
Looking forward to it, Ben. And while I respectfully disagree on principles, I greatly appreciate the work you do to advance thought and discussion on these issues.
The original email (http://sluggish.uni.cc/david_hart_siai_rant.txt) makes it clear that they had good reason for firing him. He presents himself in a very immature and egotistical fashion, and the email to google was impulsive and potentially damaging to the organization's image. Characterizing the expulsion as an "excommunication," as if there was some kind of religious basis to it, is quite sensationalist. From the bottom of that email:
The funding request you sent to the Google guys was inappropriate for a
couple of reasons:
1) Google funded OpenCog on the basis of the Google Summer of Code
only. For-profit organizations like Google do not regularly fund
outside projects besides special programs like Summer of Code. If
they want that sort of association, they will buy out the entity.
2) You CCd a number of SIAI people in your email, making it very much
seem that we had made the decision to approach them as a group, when
no such thing ever happened. This ruins our brand by association it
with requests for handouts from for-profit companies. We value our
relationship with Google very much -- by jumping in with your
unplanned email, you make us look like beggars. .
The correct way of making your request would have been to send separate
requests to SIAI and Google as a personal request, not using SIAI's name
to make a request to Google without clearing it with us.
I used to work with Ben Goertzel (a lot of fun, BTW) but I don't know David Hart. Sounds like personal issues.
BTW, OpenCog is an interesting system. I have spent a little time playing with it on Ubuntu. I would be more interested if it was not mostly written in C++ (with some Guile Scheme and Lua scripting). If you like working with C++ and are interested in AI, then it is worth a look.
11 comments
[ 5.0 ms ] story [ 36.0 ms ] threadThat sounds like the way I approach most problems I don't know the answer to. I fiddle with various relevant parts, and see what pops out. My first version is rarely a good one, but it leads (always has so far) to version 2 which works well.
Someone has to try this and if their idea doesn't work, I will bet good money that they'll stumble upon other interesting results (if they keep going).
I agree that Open cog will fail but I think it's failure should be looked at.
"if we glue together enough parts that intuitively seem relevant to AGI, maybe AGI will pop out!" may sound dumb but sooner or later, an "intelligence" program needs to reach a boot-strapping stage so it is actually credible that there is some algorithm with the properties that if you add enough processing power to it, intelligence will indeed "pop" out.
Many believe the Mammalian neocortex essentially involves such a generic algorithm. But the problem is there are fairly good arguments that the kinds of algorithms we've tried so-far could include not this algorithm. They are not algorithms that can reach the "critical mass" level.
The first failure of AI was based on the now clearly-false belief that intelligent behavior was essentially logical deduction - these efforts main appeared in the 1970's. Shredlu was the most successful.
The second failure of AI seems to be the belief that intelligent behavior involves primarily something like statistical inference. Narrow "machine learning", OpenCog, and Jeff Hawkins stuff all seem to fall out of this later approach.
My sense is that the world humans live in is too chaotic for pure logic yet much more orderly than phenemena suited to pure statistical inference ("machine learning" can work but its failure is obvious in excessive time it take make "obvious" (to us) inferences). The most uniquely human intelligence involves neither making complex deductions nor filtering scattered and chaotic masses of data but rather putting "semi-structured" data together in a "common sense" fashion. Humans follow patterns immediately, not after seeing them over-and-over again. Humans are "terrible" at pure calculation but also "terrible" at pure inference problems like gambling games (and the way human fail at gambling is very instructive - we tend to see too many patterns in random data. )
The challenge is how to create a "third way" beyond these two failed approaches (keeping in mind that both approaches can have successes too).
I'll leave my grandiose announcement of this algorithm for a different post.
Either make a point in the article itself explaining what the situation is or don't bother posting this to HN.
I used to work with Ben Goertzel (a lot of fun, BTW) but I don't know David Hart. Sounds like personal issues.
BTW, OpenCog is an interesting system. I have spent a little time playing with it on Ubuntu. I would be more interested if it was not mostly written in C++ (with some Guile Scheme and Lua scripting). If you like working with C++ and are interested in AI, then it is worth a look.