AGI from AI
November 12th, 2007 –
Will AGI emerge from a preexisting narrow AI field, incrementally improving?
In my opinion, the answer is likely no, but people working in narrow AI like to tell me that their work will eventually give rise to the Friendly AI I want to see.
Should the idea of AGI emerging from narrow AI be dismissed outright? Probably not. Let’s say AGI does indeed emerge from AI. If so, what are possible routes?
- Entertainment robots, like the upcoming Zeno
- Combat robots, like SWORDS
- Systems used in mathematical finance
- Agents in virtual worlds, like Novamente’s latest effort
Can you think of any others? Different paths have different advantages from both a FAI and an AGI perspective. Some of these “narrow” applications, such as Novamente’s, are in fact built on an AGI-oriented architecture. Could the first AGI blindside us, by superficially appearing like narrow AI?
I tend to agree with you, although I’d add a couple of other categories to your list:
- Industrial / home robots, including personal assistance robots, maintenance / inspection robots, and car-driving robots.
- Natural Language Processing systems such as document analysis and machine translation.
- Complex planning systems, perhaps including software-writing systems.
I think this is an excellent question and I think the answer is yes, this might happen. Although I would guess the first AGI will come from some form of reverse engineering effort like the current work with cortical column simulations, I’d suggest these possible locations for an AGI that develops without human intention:
Google’s server farm - the largest parallel processing environment in the world.
Yahoo’s server farm.
The Microsoft Neural Net that powers Live search.
Amazon’s server farm, which houses an increasing number of disparate programs.
I had a chance to ask Marissa Mayer about this and she noted that the output of the Google algorithm already “looks like” human thought output in ways that have surprised them. She was not saying they were close to consciousness with the Google algorithm, but my personal guess is that we overrate the distinction between narrow AI and general, and consciousness may simply spring up when the connections and the dialog meet some threshold.
If I had to *bet my life* I think I’d bet on Google for this even though it’s not a focus of theirs (yet), since they have some good AI folks, a huge environment, and massive money to do the job. I’m guessing they’ll probably wait until the cortical column research is bearing fruit and then take over those projects.
There is *NO SUCH THING* as “general intelligence”, this is known and *PROVEN* by professionals in the field:
“Although the human brain is sometimes cited as an existence proof of a general-purpose learning algorithm, appearances can be deceiving: the so-called no-free-lunch theorems [Wolpert, 1996], as well as Vapnik’s necessary and sufficient conditions for consistency [Vapnik, 1998, see], clearly show that there is no such thing as a completely general learning algorithm.
All practical learning algorithms are associated with some sort of explicit or implicit prior that favors some functions over others.”
From “Scaling Learning Algorithms towards AI” ( http://www.iro.umontreal.ca/~lisa/pointeurs/bengio+lecun_chapter2007.pdf ) to appear in “Large-Scale Kernel Machines”, L. Bottou, O. Chapelle, D. DeCoste, J. Weston (eds) MIT Press, 2007.
So….
What are you talking about?
It’s true that you can’t create an intelligent algorithm with unlimited (or infinite) generality. Even the most “general” modules of the human brain still require that the environment be learned (ie. “specialized” by the environment) - before any significant intelligence can be deployed. But the human brain *is* an existence proof that “general” intelligence can exist *at least* up to the human level. And I think that there is good evidence that the human level can be well exceeded.
“Will AGI emerge from a preexisting narrow AI field, incrementally improving?”
It might emerge from patching a lot of disparate narrow fields together, but I can’t see how the progeny of one single narrow field will end up being general, this is almost true by definition. I mean no matter how advanced a financial model you make, it won’t be able to understand natural language; no matter how good a handwriting recognition program you write it won’t be able to do commonsense reasoning, etc.
I actually like the idea of patching narrow AI programs together, and I’m looking into doing a PhD in this area. The key thing to notice is that “patching together” is not as straightforward as it sounds; it seems to me as if a general program is more than the sum of its narrow parts, and thus any effort to go from a collection of narrow algorithms to one general one is likely to be an, um “interesting” exercise.
“Could the first AGI blindside us, by superficially appearing like narrow AI?”
I’d say that it’s possible, but not the most likely outcome. After all, a good portion of the human brain consists of specialized, essentially “narrow” interacting modules (eg. visual cortex, parietal lobes, prefrontal cortex, etc.). But I think that the most likely projects to succeed first will be the ones that are focused on creating an AGI from the start, like Novamente and others.
AGI won’t _emerge_ from narrow AI; but these will put valuable tools in place for those researchers who make the AGI breakthrough.
A couple of questions.
One. How about setting up a networked computer grid for AGI researchers ala SETI@home, Folding@home and cosmology@home. Folding@home has reached one petaflop of processing speed, btw.
Two. What impact will quantum computing have on AGI research and theories? D-Wave aside, researchers *are* making advances in this area.
I, too, have longed for a distributed computing - type of thing for AGI research projects, but have been frustrated that none exists. I’m quite convinced that protein folding projects like Folding@Home (nearly 1.5 PetaFLOPS!) and Rosetta@Home have benefitted tremendously from such computing power, even if computer simulations are (always) approximations.
If AGI (or even narrow AI) research projects would benefit from some kind of distributed computing, I would gladly donate CPU time.
“Two. What impact will quantum computing have on AGI research and theories? D-Wave aside, researchers *are* making advances in this area.”
I expect, not much direct impact. Quantum computing is fit only for highly specialized processing. Eg. Decryption and Factoring large numbers. QC isn’t fit for programming “stable/rigid” algorithms that will be necessary for an AGI. AFAIUI, in QC, until the algorithms “decohere”, the algorithms can basically be considered to be random. (ie. not useful for constructing an AGI). QC might indirectly have a modest impact by furthering our understanding of science or engineering in a different domain. But, I’m no QC expert.
I’d have to agree with Jeff. QC is not the cure for all ills, and besides, there are no usefully large quantum computers around. Also, why is everyone so excited about getting more computing power? I’m on-and-off thinking about a potential PhD project in knowledge representation, and computing power is the last thing on my mind; I think that getting the right ideas is much more important. Well, perhaps I’ll be singing a different tune when I’m actually getting down to writing some programs and then trying to run them ;-0
“I’m on-and-off thinking about a potential PhD project in knowledge representation, and computing power is the last thing on my mind; I think that getting the right ideas is much more important.”
I agree; algorithmic knowledge representation is probably the proverbial “vital essence” of AGI (in a way not terribly dissimilar from how humans learn about their environment). I think that general intelligence is the possession of a (variably useful) internal model (algorithmic knowledge matrix) of “how the environment works”. The relevant forms of throughput are “filtered” through this model. Or put another way, the diverse throughputs are always perceived in-relation-to (or by-reference-to) this algorithmic knowledge matrix. That’s why our human intelligence is flexible - diverse forms of throughput are always filtered through our internal model of “how things work”.
Eh… scratch that. I think a better description would be an internal model (algorithmic knowledge matrix) of “how things relate”. I think that understanding a concept is being able to relate it to other things/concepts. “How things work” and “how things relate” are probably roughly synonymous, but “how things relate” I think is a better description. By “perceived” in the above, I meant: calculated. The diverse throughputs are always perceived (ie. calculated) in-relation-to the internal model. As a simple example: The procedure is “add 2″. When the input is 3, the output is 5. But when the input is 6, the output is 8. However, in both cases the output/throughput is always calculated “in-relation-to” the algorithmic procedure.
Based on my (currently very elementary) understanding of Novamente’s structure, I think that it may be a sufficiently ideal platform for directly implementing a version of Friendliness. Implemented as repeated throughput. It has the structure to incorporate (learned) abstract concepts, (eg. “what is compassion”) and “common sense”. Now if we can just get some large investments into Novamente R&D.
What is missing is an understanding of how knowledge is worked within the individual, in groups, and how these interact with each other and the social knowledge base. Until this entire knowledge working process is fully understood, no algorithm can enhance it or accurately represent it.
So neither AI or AGI are correct approaches. When knowledge working processes are appropriately understood, knowledge creation will finally be accurately differentiated from intelligence and we’ll no longer be trying to create any kind of artificial ‘intelligence’ (it already exists), but will understand that the real problem, and opportunity, on the table is artificial ‘knowledge creation.’
There’s a competition every year to see if somebody can write a program to pass the Turing Test. Once somebody wins that contest, we’ll have AGI.
If someone wins it , we are likely to have AGI. But there is also a chance that the Turing Test could be one by “fooling” the judges with human-like responses: Something like ELIZA but much better.
Also, we may achieve AGI without the Turing Test. What if someone creates an intelligence that can figure things out, be tricky and clever, but is fundamentally alien to the human way of thinking. Human thought is very specific — an AGI need not follow its model at all.