Sunday, July 26, 2009

Will China Build AGI First?

One of the reasons I spent 4 weeks in China this summer, organizing the First AGI Summer School and collaborating on research with my friend Hugo de Garis's Artificial Brain Lab at Xiamen University, was to qualitatively investigate Hugo's stories about the great potential for AGI R&D allegedly existent in the country of China.

One thing I learned about China is: the answer to almost any nontrivial question is some complex, multidimensional form of "maybe" or "sort of."

(Eventually the maybes and sort-ofs must collapse: Chinese do make hiring and firing decisions, get married, publish papers, and take other definitive actions ... but on the whole I found that in China there is a much greater willingness to embrace uncertainty than in the US, and a much smaller desire to make things clear and definite.)

In this spirit, I can't claim I came to a definite conclusion about the potential of China to lead the world in advanced AGI R&D.

But I can say that it's a definite possibility ;-) ....

Though the situation is complex, my gut feel is that Hugo is probably right, at least in the following sense: If a moderate number of AGI researchers (from the West and China both) apply their energy to pursuing the R&D opportunities that China offers, there is a strong potential that AGI research could advance there much faster than in other parts of the world.

Here are some relevant facts, militating in favor of China's role in AGI technology:

  • Even in the current dismal world economy, China's economy is still growing
  • Chinese students and researchers are willing to work long hours (of course some Americans and Europeans are too ... but my impression is that this willingness is greater in China)
  • The Chinese education system is very good at teaching advanced mathematics and algorithm theory, which are important for AGI
  • China is interested in pursuing advanced technologies -- both with practical applications in mind, and with a nationalistic motive of displaying their technological strength relative to other nations
  • Unlike the US, the Chinese research funding establishment has no "chip on its shoulder" about AI or AGI -- it has the same status as any other advanced technology. There was never an "AI winter" in China, nor is there dramatically more skepticism about AI than about other computer technologies
  • Due to the centralized system of government, if the central administration decides they value a certain technology, there is the potential for a massive amount of resources to be directed to that technology in a relatively rapid time-frame. Things often move very slowly in China, but sometimes they can also move much faster than in less centrally organized economies.
  • The cost of highly educated labor is low in China, so that if funding from outside China is found to help support China-based projects, this funding can go a long way!

Here are some other relevant facts, which are challenges China has to overcome if it wants to lead the world in AGI:

  • The Chinese university education system is narrowly disciplinary, whereas AGI requires copious interdisciplinary knowledge
  • Because China lacks a large, robust, cutting-edge software industry, there isn't that much "cultural knowledge" about how to manage complex software projects using "agile" methodologies. Yet, AGI is a complex software project that really demands an agile methodology. (Note that China has a load of great software engineers; the issue I'm pointing out regards software project management, not software engineering.)
  • Due to the nature of Chinese culture, it is fairly common for work environments to arise in which participants don't feel free to share their innovative ideas, and to point out problems with the ideas others are pursuing (especially if these "others" have higher social status according to Chinese tradition). However, some Chinese work environments are very friendly to innovation and criticism; so IMO this is best considered as a problem that can be overcome with attention to the personalities involved and the management mechanisms
  • Due to various factors including the restrictions on travel and Internet sites that the Chinese government places, China often feels "cut off" from the rest of the world, which results in a less-than-optimal degree of interaction with the international research community

Given the above factors, my conclusion is that IF China is able to attract sufficient foreign AGI experts, it may well be able to leapfrog ahead of other nations in the race to create AGI.

What the foreign experts would bring is not just AGI expertise and ideas, but expertise in allied areas like

  • interdisciplinary education
  • agile software project management
  • the management of innovation-friendly, "flat-hierarchy" research groups

Further, the presence of foreign experts in China full-time would result in other foreign experts more frequently traveling to China to speak, and in Chinese students more often traveling outside China for conferences and research visitation -- all of which would decrease the "China isolation" factor, and increase the intellectual potency of Chinese AGI research labs.

It's also the case that in Chinese academia right now, foreigners can get away with "shaking things up" a little more than Chinese nationals can. So even if a Chinese national showed up with exactly the same expertise and personality as a foreigner, they would have different strengths and weaknesses in the context. They would be able to get some things done more easily due to their Chinese-ness, but would also not be able to "get away with" as many disruptive methodologies.

So: the reason I think foreign AGI experts are critical is not that Chinese lack good ideas about AGI. (Yes, I think my own ideas about AGI are the best ones, but that's not my point right now!) It's that I don't think any one country has a monopoly on great AGI ideas and people, so the prize is likely to go to a country that can build an international AGI research community ... and also that there are certain organizational skills that are very useful for AGI, but not that well developed in China right now.

But there are serious challenges involved in recruiting foreign AGI experts to China:

  • The salaries are low by international standards. The salaries foreign faculty get at Chinese universities allow a very nice lifestyle in China -- but even so, they don't go that far in terms of international travel, purchase of electronics, or helping family members in the West.
  • Many AGI researchers have spouses and children who don't want to live in China. There are good international schools for researchers' children; but finding appropriate jobs for spouses can be difficult due to the language barrier and the different nature of the economy
  • The Chinese system of government is alien and off-putting to some foreigners. (As an example, this blog cannot be read by most Chinese Internet users right now, because is blocked by the Chinese government. This sort of thing really bothers some foreign researchers and/or their families)
  • Some researchers will fear "career damage" if they go to a university without international name recognition (though, this factor would disappear quickly after a critical mass of researchers went to China)

So, the sixty-four trillion dollar question is to what extent these latter factors can be overcome.

I believe they could be overcome if Chinese universities or research agencies made very clear, very clearly research-friendly offers to foreign AGI researchers -- and then followed through on these offers once the researchers arrived. AGI researchers are a dedicated bunch, and many would put up with the problems cited above in order to have a good chance to lead a team of brilliant, qualified students at implementing powerful AGI systems.

As you may have guessed, one reason I'm cataloguing these factors so systematically is that I'm debating trying to rearrange my life to either move to China or (more likely) spend one semester per year in China.

At the moment, in my own discussions with Chinese universities about AGI research funding, I am finding things mildly confusing. The discussions are going interestingly, but I feel much less clarity than I would in a comparable discussion with a US university. This is nobody's fault -- it's a natural consequence of "cultural differences" -- but it's a factor that will have to be smoothed-out, IMO, if China is going to recruit a sizeable number of foreign AGI researchers.

So we now reach the conclusion of the above chain of thought. IF Chinese universities manage to fine-tune the art of recruiting foreign AGI researchers, then I think that China has a real chance of leading the world in the development of AGI.

I predict that, if China doesn't adopt the world lead in AGI, it will be because it fails at the things I cited above. They will fail to dominate in AGI if it turns out that the Chinese way of recruiting and retaining foreigners is too alien, causing a failure to accumulate a critical mass of foreign R&D leaders. This could certainly happen. Time will tell.

What Are the Risks if China Pulls Ahead in the AGI Race?

So China might plausibly take the lead in the AGI race.

As a citizen of the USA and Brazil (not China), does this worry me?

Not really.

Like Buckminster Fuller, I consider myself a "passenger on Spaceship Earth" (and I won't hesitate much to board another vessel when one becomes available -- or better yet I'd like to send multiple copies of myself on multiple vessels! But, I digress ;-).

One thing I'm being insistent on in my collaboration with Hugo de Garis's Artificial Brain Lab at Xiamen University, is that all our work be released as open-source code. The university folks there have no problem with this. So, in the case of my and Hugo's work, it's not a situation where we're trying to develop an AGI that will be exclusively owned by the Chinese government.

And, I feel strongly that anyone else doing AGI work in China -- or anywhere else! -- should take the same approach. The main reason I decided to open-source my own AGI project (OpenCog) (while keeping some valuable AGI-related technologies proprietary within Novamente LLC), is the intuition that AGI is a sufficiently big and thoroughgoingly important thing that it should be developed by the human collective mind as a whole, not by a small group or even a single nation.

Of course, if more and more AGI research gets done in China, then more and more of the world's AGI expertise will exist in China -- which will give China a substantial leadership position in AGI, regardless of whether the AGI code is open-source. But this really doesn't worry me much either, partly because I've been so impressed with the character and spirit of the younger generation of Chinese, who (in these scenarios we're discussing) will be doing most of the AGI work.

I met wonderful Chinese people of all ages during my visit to China. But there are huge generational differences among Chinese, and the Chinese who grew up with the Internet have a drastically different view of the world than the immediately prior generations. Most of the Chinese I met aged under 30 had a reasonably modern, international understanding of the world -- and some of the Chinese I met aged under 22 had such modern attitudes that they really could have been youth anywhere. The Internet is spreading international ideas and culture around the world, just like it's spreading the AGI meme around the world -- and may increasingly start spreading AGI researchers around the world ... we'll see.

Although I used a picture of Alfred E. Neuman above (like Hugo, I'm kind of a sucker for dramatic effect), I want to be clear that I don't have a cavalier attitude about the threat that could be posed if ANY government took control of the world's first AGI for their own parochial ends.

But I think this is a problem we need to work around, regardless of which country we do our research in.

By developing open-source code (made available on SourceForge, Launchpad, Google Code and so forth), and by carrying out our research in a way that emphasizes linkages with the international research community, we'll guarantee that AGI comes about as a product of the international collective mind of AGI researchers. This provides no grand guarantee of "AGI safety" (nothing can do that), but I strongly feel it's the best approach.

The First AGI Summer School in Xiamen, China (Retrospective and Random Related Musings)

I returned home 2 weeks ago from the First AGI Summer School, which was held in the Artificial Brain Lab at Xiamen University in Xiamen, China at the end of June and the beginning of July.

Ever since I got back I've been meaning to write a proper summary of the summer school -- how it went, what we learned, and so forth -- but I haven't found the time, and it doesn't look like I'm going to; so, this blog post will have to suffice, for the time being at any rate.

First of all, I need to express my gratitude to Hugo de Garis

and Xiamen University for helping set up the summer school. Coming to Xiamen to do the Summer School was a great experience for me and the others involved -- so, thanks much!

Some photos I took in Xiamen are here (mixed up with a few that YKY took on the same trip). (Viewer beware: some of these are summer school photos, some are just "Ben's Xiamen tourism photos"....)

To get a sense of what was taught at the summer school -- and who was on the faculty -- you can go to the summer school website; I won't repeat that information here.

The first two weeks of the summer school were lecture-based, and the last week was a practical, hands-on workshop focused on the OpenCog AI system. Unfortunately I missed most of the hands-on segment, as I wound up spending much of that week meeting with various Chinese university officials about future possibilities for Chinese AGI funding (but I'll write another blog post about that), and demo-ing the Artificial Brain Lab robot to said officials.

See here for some videos of the above-mentioned robot, along with some "OpenCog virtual pet" demo videos that were shown at the summer school. (And, the OpenCog virtual pet was also gotten up and running "live" in Xiamen, of course....)

The number of students wasn't as large as we'd hoped -- but on the plus side, we did have a group of VERY GOOD students who learned a lot about AGI, which was after all the point.

(In fact, most conferences have found their attendance figures down this year, due to people wanting to save money on travel costs: an obvious consequence of the faltering world economy.) The majority of students were Chinese from Xiamen University and other universities in Fujian province, but there were also some overseas students from Europe, the US, Korea and Hong Kong (OK, well, Hong Kong isn't quite "overseas" ;-).

All the lectures were videotaped by Raj Dye (thanks Raj!!)

and will be put online once Raj gets time to edit them. I think these will form an extremely valuable resource, and will reach a lot more people than the summer school itself did. (Long live the Internet!!). Raj's active camera work captured a bunch of the dialogues during and after the talks as well, and I think these will make quite interesting viewing. As you might expect, there was some pretty intense give-and-take (especially, for example, during Allan Combs' talks on cognition and the brain).

I'm definitely interested to help organize some future AGI summer schools ... though the next one will be in a different location, as we've already done a pretty good job of spreading the word about AGI to the AI geeks of Fujian Province! Maybe the next one will even be back here in the boring old US of A ....

Random Observations on Chinese-ness in the AGI Context

I learned a lot about China in the course of doing the summer school (though I'm still pathetically ignorant about the place of course ... there's a lot to know) ... I won't try to convey 1% of what I learned here, but will just write down a few hasty and random semi-relevant observations.

First, I learned to speak verrrry slowly and clearly since Chinese students are more accustomed to written than spoken English! ;-)

More interestingly, I learned that the Chinese educational system is more narrowly disciplinary than the US system, and also more focused on memorization of declarative knowledge rather than practical "know how." Compared to their US counterparts, computer science graduates in China know an AWFUL LOT of computer science, yet don't have much advanced knowledge of areas beyond computer science, nor all that much software engineering knowledge or hands-on coding experience. ("Software Engineering" is a separate department in the Chinese university system, and I didn't get to know the Software Engineering students, only the Computer Science ones.) So one role the summer school served was just to introduce a bunch of Chinese AI students to some allied disciplines -- neuroscience, cognitive psychology, philosophy of mind -- that they hadn't seen much during their formal education so far.

(Actually, separately from the Summer School, I did give a talk to some undergrads in the Software Engineering School, on AI and Gaming, which contained one funny bit (unfortunately that talk was not videotaped). I wasn't sure if the students understood what I was talking about, so as a test I showed them this picture as part of my powerpoint

Normally I use this lovely picture as an example of "conceptual blending" (a cognitive operation that OpenCog and other AGI systems must carry out), but this time I announced it differently; I said: "Furthermore, the Artificial Brain Lab here at XMU has an ambitious backup plan, in case our computer science approach to AGI fails. We've devised a machine that can remove the head from a graduate student, and attach it to the body of a Nao humanoid robot, and thus create a kind of synergetic cyborg intelligence." I was curious to see if these Chinese undergrads understood my English well enough to tell that I was joking -- but from their reaction, I was unable to tell. They laughed because the picture looked funny, but I still don't know if they understood what I was saying! Fortunately the Summer School students were less inscrutable, and more reactive and communicative! And overall the AI in Games lecture went well in spite of this perplexing crosscultural joke experience....)

(As an aside within an aside, I also learned during various conversations that typical Chinese high school students spend from 7AM till 10PM or so at school, 6 days a week. Damn.)

Another thing that surprised me was the strength of knowledge the Chinese students had in neural nets, fuzzy logic, computer vision and other "soft computing" and robotics related AI, as compared to logic-based AI. By and large, they had a very strong mathematics background, and a good knowledge of formal logic -- but fairly little exposure to the paradigm in which logic is applied to create AI systems. Quite different from the typical American AI education.

All in all the Chinese seemed to have a lot less skepticism about "strong AI" than Americans. It's not that they had a great faith in its immediacy -- more that they lacked the egomaniacal confidence in its extreme difficulty or implausibility, which one so often finds in Westerners. Chinese culturally seem much more comfortable with accepting situations of great unconfidence, in which the evidence just doesn't exist to make a confident estimate.

I came to the summer school from the Toward a Science of Consciousness conference in Hong Kong, where I led a Machine Consciousness workshop -- which I won't write about here, because I wrote a summary of it for H+ magazine, which will appear shortly. Issues of machine consciousness came up now and then at the summer school, but interestingly, they seem to hold a lot less fascination for Chinese than for Westerners. When I put forth my panpsychist perspective in China (that the universe as a whole is conscious in a useful sense, and different systems -- like human brains and digital computers -- manifest this consciousness in different ways ... and our "theater of reflective consciousness" is one of the ways universal consciousness can manifest itself in certain sorts of complex systems), no one really bats an eye (and not just because the Chinese lack a taste for eye-batting). Not that Chinese scientists consider this panpsychist perspective wholly obvious or necessarily correct; but nor do they consider it outrageous -- and, most critically, very FEW Chinese seem to feel like many Westerners do, that "reductionism" or "materialism" is obviously correct. Once you remove the tendency toward dogmatic materialism, the whole topic and dilemma of "machine consciousness" loses its bite....

China versus California (A Semi-Digression)

(This section contains some ramblings on Oriental versus California culture, and the Singularity -- which are only semi-relevant to the summer school, but I'll put them here anyway, because I find them amusing! Hey, this is a blog, anything goes ;-)

In mid-July I voyaged from the Xiamen AGI summer school to California where I gave the keynote speech at the IJCAI workshop on Neural-Symbolic computing (a really interesting gathering, which I'll discuss some other time), and then gave a lecture on AGI at the Singularity University (at NASA Ames Lab, in Silicon Valley).

The contrast between the SU students and the Chinese AGI Summer School students couldn't have been more acute.

For one thing, there was a huge contrast of ego ... to phrase things dramatically: The SU students emanated an attitude that seemed to say "We know more than anyone on the planet!! We already knew almost everything we need to know to dominate the world as part of the techno-elite!"

The Chinese students were not actually more ignorant (though their knowledge bases had different strengths and weaknesses than those of the SU students), but they were dramatically more humble about their state of knowledge!

The SU students also seemed extremely eager to project everything I said about AGI into the world they knew best: Silicon Valley style Internet software. So, most of the questions during and after my talk centered around the theme: "Isn't it unnecessary to work on AGI explicitly ... won't AGI just emerge from the Internet after Silicon Valley startup firms create enough cool narrow-AI online widgets?" When I said I thought this was unlikely, then the questions turned to: "OK, but rather than writing an AGI that actually thinks on its own, shouldn't you just write a narrow-AI that figures out the best way to combine existing online widgets, and achieves general intelligence that way?" And so forth.

But I don't want to make it sound like the SU student body is "all of one mind" -- it's certainly a heterogeneous bunch. At the lunch following my talk at SU, one SU student surprised me with the following statement (paraphrased): "One reason I think AI systems may not achieve the same kind of ethical understandings or states of mind as humans, is that they lack one of the most important human characteristics: our humbleness. We humans have a lot of limitations in our bodies and minds, and these limitations have made us humble, and this humbleness is part of what makes us ethical and part of what makes us profoundly intelligent in a way that a mere calculating machine could ever be."

I laughed out loud and immediately said to the student: "OK, I'm onto you. You're not American." (The student did look Asian ... but I was guessing he was not Asian-American.)

He admitted to being from Korea ... and I noted that few Americans -- and especially no Silicon Valley techno-geek -- would ever identify humbleness as a central characteristic of humans or a key to human intelligence!

Then I couldn't help thinking of the saying "Pride comes before a fall" ... and Vinge's (correct) characterization of the Singularity as a point after which HUMANS WILL HAVE NO IDEA WHAT'S GOING ON ... i.e. no real ability to predict what happens next, as superhuman nonhuman intelligences will be dominating the scene.

Philosopher Walter Kauffmann coined the dorky but evocative term "humbition" to denote the combination of humility and ambition. There's not much humbition in Silicon Valley ... nor for that matter in the public trumpetings of the Chinese government ... but there was a LOT of humbition in the Chinese students at the AGI summer school and the Artificial Brain Lab. Perhaps this quality will serve them well as the world advances, and our knowledge and intuitions prove decreasingly adequate to comprehend our situation...

If you believe that AGI will be created from piecing together narrow-AI internet widgets, then yeah, mostly likely AGI will be created by the Silicon Valley techno-elite. But if (as I suspect) it requires fundamentally different ideas from the ones now underlying the world's technological infrastructure ... maybe it will be created by people who are more open to fundamentally new and different ideas.

But this leads into the next blog post I'm going to write, exploring the question of whether Hugo de Garis is right that AGI is going to get created in China rather than the West!

Musings on the Concept of a Systematic AGI Curriculum, and Lessons for Future AGI Summer Schools

Next, what did I learn this summer about the notion of an AGI summer school, and about teaching AGI altogether?

One big lesson that got reinforced in my mind is: Teaching AGI is very different than teaching Narrow AI!

There is basically no systematic AGI education in universities anywhere on the planet, and this fact certainly helps to perpetuate the current AGI research situation (in which there is very little AGI research going on). By and large, everywhere in the world, students graduate with PhD degrees in AI, without really knowing what "AGI research" means.

Another conclusion I came to is that a carefully crafted "AGI Summer School" curriculum could play a major role -- not only in providing AGI education, but in demonstrating how AGI material should be structured and taught.

However, creating a thorough, systematic AGI curriculum would be a lot of work ... and we didn't really attempt it for the First AGI Summer School. I think the lectures mostly went very well this time (well, you can judge when the videos come online!!), and the sequencing of the lectures made good didactic sense -- but, for the next AGI summer school, we'll put a little more thought into framing the curriculum in a systematic way. Now, having done the summer school once, it's more clear to me (and probably the other participants as well) what an AGI curriculum should be like.

First of all, it's obvious that to make a systematic AGI curriculum, one would need some systematic background curriculum in areas like
  • Neuroscience
  • Linguistics
  • Philosophy of Mind
  • Psychology (of Cognition, Perception, Emotion, etc.)
We didn't do enough of that this time. Allan Combs' lectures at the Xiamen summer school formed a nice start toward a Neuroscience background curriculum for AGI, but due to lack of time he couldn't do everything needed.

In this vein, one thing that became clear to me at the Xiamen summer school is: The standard "cognitive science" curriculum would certainly fill this need for background, but it's not exactly right, because it's not specifically focused on AGI ... AGI students really only need to digest a certain subset of the cognitive science curriculum, selected specifically with AGI-relevance in mind. But judiciously making this selection would be a nontrivial task in itself.

Next, as part of a thorough AGI curriculum, one would need a systematic review of different conceptions of what "general intelligence" is -- we did such a review at the Xiamen summer school, but not all that systematically. Pei Wang gave a nice talk on this theme, and then Joscha Bach and I presented our own conceptions of GI, and I also briefly reviewed the Hutter/Schmidhuber "universal intelligence" perspective.

Then there's the matter of reviewing the various AGI architectures out there. I think the Xiamen summer school did a fairly good job of that, with in-depth treatments of OpenCog, Pei Wang's NARS architecture, and Joscha Bach's MicroPsi ... and a briefer discussion of Hugo de Garis's neural net based Artificial Brain approach ... and then very quick reviews of other AGI architectures like SOAR and LIDA. Of course there are many, many architectures one could discuss, but in a limited time-frame one has to pick just a few and focus on them. (It would be nice if there were some more systematic way to review the various AGI architectures out there than taking a "laundry list" approach, but this isn't an education problem, it's a fundamental theory problem -- no such systematization exists, even in the research literature.)

There were a lot of OpenCog-related lectures at the Xiamen summer school, and one thing I felt was that it was both too much and too little! Too much OpenCog for a generic AGI Summer School, but too little for a real in-depth OpenCog education. At future summer schools we may split OpenCog stuff off to a greater extent: give a briefer OpenCog treatment in the main summer school lectures, and then do a separate one-week OpenCog lecture series after that, for students who want to dig deep into OpenCog.

Another educational issue is that each AGI architecture involves different narrow-AI algorithms, so that to really follow the architecture lectures fully, students needed to know all about forward and backward inference, attractor, feedforward and recurrent neural nets, genetic algorithms and genetic programming, and so forth. (Most of them did have this knowledge, so it wasn't a problem; actually this might be more of a problem in the US than in China, as China's education system is very strong on comprehensively teaching factual knowledge.) That is: even though AGI is quite distinct from narrow AI, existing AGI architectures make ample use of narrow-AI tools, so students need a good grounding in narrow AI to grok current AGI systems. It would be good to make a systematic list of tutorials on the most AGI-relevant areas of narrow AI, for students whose narrow-AI background is spotty. Again, we did some of this for the Xiamen summer school, but probably not enough.

Finally there's the terminology issue. There is no good "AGI glossary", and every researcher uses terms in slightly different ways. Updating and enlarging an online AGI glossary would be a great project for students at an AGI summer school to participate in!

Undramatic Non-Conclusion

So, the first AGI Summer School went pretty interestingly, and I'm really glad it happened. It was interesting to get to know China a little bit, and to get some experience teaching AGI in an intensive-course context. I learned a lot, and I guess the other faculty and the students did too.... I also made a number of excellent new friends, both among the Chinese and the foreign students. As with many complex real-world experiences, I don't really have any single dramatic summary or conclusion to draw ... but I'm looking forward both to future AGI summer schools, and to future experiences with "AGI in China"....