On Digg today:
Five Reasons Google Will Invent Real AI
Turns out that it’s from the gossip blog Valleywag. Its author recently attended the Singularity Summit. Looks like the topic is staying in his mind a bit! Of course the recent George Dyson article (“Turing’s Cathedral”) likely contributed as well.
Is the idea plausible? For Google to have a chance at reaching real AI, they would have to make it a priority. As in, would they need to create a project specifically devoted to it, and nothing else. They would need to put a dozen or more supergeniuses to work full-time for years on end, at a likely cost of tens of millions of dollars with no substantial return in the forseeable future. Do any Singularity-watchers think this could happen? Not before some other group has made substantial progress already, is my guess.
What are the necessary conditions for any group having even a chance at AGI (artificial general intelligence) in the next couple decades, or before nanocomputing, whichever comes first? Here’s what I’m throwing out there:
Deep pockets. Enough funding that the project can proceed without having to worry about commerical spin-offs. So we’re talking about a government, corporation, startup, or non-profit with at least a couple million in the bank, perhaps more.
Exceptional brains. To get there first, core team members will need to be the best there is, close to the upper boundaries of what is possible with human intelligence. We’re talking people that are 1/100,000, not 1/10,000 or 1/1000.
Education in the right fields. Universities don’t offer degrees in Artificial General Intelligence. The knowledge set necessary to create AGI successfully is not known, but it will likely encompass all the following fields: cognitive science, math, programming, probability theory, traditional AI techniques, information theory, and maybe more. It is not a knowledge set that an employee at Google will just happen to be familiar with, even if they are well-educated.
Math and programming talent. It’s one thing to have a high IQ and be educated in the multiple necessary fields. However, if you’re going to do successful engineering, it’s likely that you’ll have to be specifically talented at implementing your ideas in code. There are plenty of really smart, really well-educated interdisciplinary scientists out there that are simply not engineers and can only program at an average level. They write papers and give lectures that are brilliant, but when it comes to actually building a huge program, the spark is just not there. And last but not least…
A correct theory. For every 10,000 theories of general intelligence that sound great, feel great, there may be only one (or zero) that can be implemented on available hardware without complexity overload or centuries of debugging. It doesn’t matter if you’re the smartest person in the world, if you settle on the wrong theory, and try to implement it, it just won’t work. Then you’ll need to tear everything down and start over, most likely from scratch. If only one person is in charge of the overall theory, then all it takes is that one person to be wrong for everyone’s time to be wasted.
As we can see from the above requirements, they’re not fulfilled yet. They could be before the decade is out, though, which would lay the groundwork for real progress.
Tangentially: a couple days ago, Bruce Klein, President of Novamente LLC, claimed here that “we estimate it will take 6 years w/ a full-time staff (about a dozen programmers) to reach human-level AI”. How can they be so sure about timeframes? Because they are convinced they already have a theory that will work, and they are just implementing it. Should we be skeptical of this claim? I certainly think we should. More often than not, when people think they have the right theory, it turns out they don’t. This doesn’t mean it will never happen… just that I doubt that anyone can be so sure so far in advance.
In case you’re wondering, a pdf describing Novamente’s theory of intelligence can be found here.
Also, here is a comprehensive list of projects working towards AGI today, in June 2006. I think this was also put together by Bruce Klein.
“Education in the right fields. Universities don’t offer degrees in Artificial General Intelligence.”
Sorry, this is more likely an impediment than an aid.
There is obviously *something* missing from all current approaches to AI.
Being well learned in the current pratices means you will just try to further the *known* solutions which have always been looking so “promising”, they do bring some results, otherwise they would not be there.
As a sad example, look at poor Doug Lenat’s CYC, still in the craddle after more than 20 years.
Focus should be on the “missing ingredient”.
There will not be any lack of skills to develop the “classical” parts of the solution, witness your link to the AGIRI projects list.
What could be “missing”, who knows…
Hi Michael,
For an explanation of the “6 year” number, please see: http://www.imminst.org/forum/index.php?s=&act=ST&f=11&t=10799&#entry112242
Interesting perspective. It is always interesting to read your thoughts on stuff, Michael.
Also, Ben is the CEO… I’m just the lowly President ;)
Out of curiosity do you have any support for your requirements? They seem unusually forceful and declarative for you.
How do you know it requires genius? What leads you to believe there aren’t many viable approaches to machine intelligence?
Where does this all come from?
Hi Justin,
Here’s where they come from. I have my own fuzzy conception of the size of the problem, just like a lot of other people. I think it is relatively large and difficult – but not as large and difficult as many people think. There are already geniuses out there with decades of programming experience who are working towards AGI, and have been for several years now. If it were possible for a non-genius to make serious progress, then they would have done so. Meanwhile, many geniuses seem relatively stuck.
It’s also a matter of comparative ability to progress. If you have a team of 10 geniuses that makes serious progress on AGI, then I think it will surely make news, at least within the tech community. At that point, 10 *supergeniuses* could easily be inspired to drop what they’re doing and start pursuing the problem. They may lag a bit further behind at first, but I think the problem is big enough that the factors I point to above can be enough for a project with these resources to overtake another quite easily. What I think limits supergeniuses from pursuing AGI today is not so much their inability to make progress, but their belief that AGI is decades in the future.
Notice also that my requirement for “deep pockets” – a couple million dollars in the bank or more – is actually not that much for a group that people are convinced has a chance of success. Also note that the challenge seems to attract exceptional brains psuedo-naturally, and that these brains tend to have math and programming talent and have absorbed at least a portion of the “necessary” fields. So while the requirements sound stringent, they are actually less stringent than they appear at first glance.
In other words – I think it’s more likely that a small project will snowball into fulfilling the requirements I list here, then achieve success, than achieving success in the absence of these requirements.
I believe that there may be many viable approaches to machine intelligence, except 99.99% of these approaches only produce an intelligence that runs at glacial speed on available hardware. The number of designs that work in realtime in a minute space of configurations in a much larger space of failure. The previous sentence also applies to conventional engineering like building a new bridge or a new automobile – but we have little guiding metrics for AGI architectural requirements, compared to the thoroughly known requirements for building a new car or bridge.
Look at the space shuttle. You make a plan for what you think will work, then execute it. If the plan is fundamentally off, no amount of tweaking will give you a design that suddenly works. You have to redesign and build it again from the ground up. Software is slightly different, because it is rewriteable, but again – if the theory is fundamentally wrong, I think that an incremental-tweak approach will get people nowhere. A sophisticated AGI may not need to be as complex as the space shuttle, but it might approach it.
thanks for the reply. I guess I was just looking for definitions. I’m always interested in bold predictions like this.
The race will begin, as I see. Google with Goertzel are just one horse. Still, ten years or so to the Result, I think.
The game is on.
Michael wrote:
”
Tangentially: a couple days ago, Bruce Klein, President of Novamente LLC, claimed here that “we estimate it will take 6 years w/ a full-time staff (about a dozen programmers) to reach human-level AIâ€. How can they be so sure about timeframes? Because they are convinced they already have a theory that will work, and they are just implementing it.
”
Basically, this is right. Our theory is not worked out to every little detail though, so there is still plenty of “low level” computer science and software design work to be done, not just straightforward software code monkeying. But we have a comprehensive theory of intelligence which is tied in with a fairly detailed and comprehensive software design for an AGI. We need to finish implementing the design, test it and teach it.
Could something not work out as expected? Of course this could happen. But we’re not going to be wimpy just because others have failed at similar projects. My view is that others have failed because their AGI designs were based on an insufficiently thorough understanding of the nature of intelligence. There is no well-documented aspect of human intelligence that the Novamente design does not, in its own way, directly address. This cannot be said of prior AGI designs that have been tried and failed.
In terms of the list of “necessary conditions” for an AGI listed above, I would in all modesty say that Novamente currently has all of these except the “deep pockets.”
However, I don’t agree that these are necessary conditions. Only the last one, a correct theory, comes close to being a necessary condition. The others are very-nice-to-haves but not strictly necessary.
Michael also wrote:
”
Should we be skeptical of this claim? I certainly think we should.
”
Sure, of course you should be skeptical of our specific time estimate. After all, you have not seen the data and thinking that went into it.
And obviously, it’s hard enough for Microsoft to estimate the time-to-delivery of their next OS version, let alone to predict the completion of a project the likes of which has never been seen in human history….
We made the 6-years estimate using pretty boring, standard schedule estimation techniques in Microsoft Project — figuring out the different tasks required to get from here to the end goal and their dependencies, etc. But of course there is substantial uncertainty in the timing of many of the component tasks.
The key point is not whether the end goal is 2, 6 or 11 years away. The key point is that with a correct design, creating AGI is a large but tractable task of coding, testing, tuning and teaching … the time requirement for which is measured in years not decades.
Finally, as for Google … as I have commented before: they have the $$ to seriously attack the AGI problem, but I see no evidence they are doing so. The Google leadership seems to think AGI will emerge organically via their work on large-scale statistical linguistics. They have displayed no propensity to hire AGI researchers, but have rather hired coders and statistical linguistics. Of course, some of their research staff could shift focus to AGI if they wanted to; but I think that if Google was making a **serious** thrust directly toward AGI one would see this in their recruiting patterns.
So: Could some Google researchers be working on AGI as their own research thrust? Of course.
But: Is there an institutional committment on Google’s part to deploying considerable resources to working directly toward AGI, rather than achieving AGI indirectly via large-scale statistical computational linguistics? I see no evidence of this, and plenty of counterevidence.
This could change at any time, of course…
– Ben Goertzel
Google’s core business is searching. AGI will result (almost by definition) if you have a petabyte database. Lets take an example
http://www.newswiretoday.com/news/17866/ But can it find itself?! You need a program that will find matches and matchmaking software.
?Quires dormir con fosforo?
Google translate does not demonstrate any underlying meaning. Match = fosforo regardless of any context.
If I ask for something and Google can find it, it has AGI. It can in the last analysis find me a program to do whatever I want. It could string programs together to perform a defined task.
Ian Parker Says: “AGI will result (almost by definition) if you have a petabyte database”
Really??!! What definition are you referring to?
The more I read people’s Ideas about AI and participate in online discussions, the more I realise that human natural languages are simply incapable of even specifying precisely what intelligence is, let alone creating it or analyzing it in a systematic way.
Everyone has their own obscure terminology; “complex adaptive system”, “complex mental dynamics”, “nondeterministic interaction”, etc, etc, etc. Why are there so many different interpretations of what the word “intelligence” actually means? Why are there so many competing ideas about how to create an intelligence in software? When Ian Parker says “intelligence”, does he even mean the same thing that I do?
As far as I can see, the problem is that we are trying to base science (or worse still, engineering) on little snippets of english, rather than on mathematics. Imagine trying to do physics or chemistry without writing any equations, but by simply stating your preferred theories in natural language. It would be a disaster – researchers would waste their time talking cross-purposes. In fact, it would be a lot like doing alchemy or pre-copernicus physics.
There is a brutal lesson to be learned from these observaions, nobody is going to like it but here it is:
** Do not read or listen to anyone’s ideas about AI unless they contain more math than non-math **
AGI research will be fruitful. My own research has confirmed this; I have been researching the AGI problem for 10 years, and my IQ is 180. Trust me on this one.
AGI refinements coupled with desktop supercomputing advancements will see Moore’s law left in the dust and the Technological Singularity occurring within 10 years. Probably more like 5.
kHz
“If I ask for something and Google can find it, it has AGI. It can in the last analysis find me a program to do whatever I want. It could string programs together to perform a defined task.”
Uhmm, that’s not agi. That’s a search engine, and google derives most of its searching prowess from analyzing the constant input of humans(links).
Now as for agi, I don’t think it will take too long. We know ever more about the human brain, our hardware is getting ever faster, and people are getting good ideas on the software side.