Human-Equivalent AI
Hi, I just thought I'd repeat some general points about AI and our future.
If human-equivalent AI is possible, this is a huge, huge deal. It would basically mean that you could turn inanimate matter into intelligence. Say that it requires about 500 teraflops (Tflops), roughly equivalent to one of the fastest supercomputers today, to run a human-equivalent AI program.
A really fast supercomputer costs about $100 million. As you may know, the cost of computing power tends to fall exponentially with time. Even if this doesn't continue forever, it seems like it will continue until 2020 or so at the very least.
Around 2020, 2025, 2030, or thereabouts, it seems reasonable to say that a 500 Tflop computer would cost in the ballpark of $1,000, if not less. If such a computer were sufficient to run a human-level AI, it would make sense for your random company to buy these computers and run them alongside conventional staff. They would be substantially cheaper. After all, these AIs could think all day and night without food, and their cognitive architectures could be boosted by direct access to number-crunching capabilities. They could share thoughts in a common format, instantaneously.
If one of the AIs became a genius, the others could just copy the cognitive features that gave the original AI those capabilities. The entire collective would never be far behind the leader, in contrast to human collectives, where our wetware is static and cannot be improved.
To actually influence the world directly, it would be helpful for these AIs to develop some robotic avatar. This would be using the robotics of 2020-2030. It would be reasonable to assume that the robotics chosen might be quite flexible and capable, especially considering that the AIs themselves could assist in reprogramming, fine tuning, research, and development.
Bacteria are idiotic, yet capable of turning a tonne of organic waste into bacterial biomass over a course of hours. Human-equivalent AIs would be smart, and have great incentives to convert raw materials into robotic or biological bodies for their habitation. Such AIs could even uncover the principles of thought and boost themselves beyond the human level, even if their access to computing power remains roughly static.
Combine AI with advanced robotics, add the motivation to improve both, and you have a potentially abrupt and disruptive transition on your hands.
What baffles me is when pundit say: "surely, such AIs would lack the capability to become major players in the human world in any short-term timeframe".
My question would be: "how do you know?"
We humans cannot put ourselves in the shoes of an intelligence that has complete access to its source code, can rearrange its cognitive architecture to optimize its performance on narrow problems, share thoughts with its comrades at the speed of light, transfer itself from point to point on the globe at the speed of light, directly integrate itself with scientific instruments as sensory modalities, blend together autonomic and deliberative processes in a thousand ways that humans can't, form beliefs and update them in mathematically rigorous ways, and so on.
Such an intelligence could come a long way in a really short time, or perhaps not. The point is that we don't know. If a pundit expresses skepticism about the idea, their opinion is more likely to reflect the limitations they know apply to humans -- not limitations applied to AIs.
It seems easier to argue that human-equivalent AI is flatly impossible than it is to argue that human-equivalent AI wouldn't have a huge impact on the world once developed. It seems most reasonable to proceed as if it would.
June 13th, 2008 - 02:43
Michael said: Around 2020, 2025, 2030, or thereabouts, it seems reasonable to say that a 500 Tflop computer would cost in the ballpark of $100,000
– I think you may have missed a few orders of magnitude off that. RoadRunner has already broken the PetaFlop.
“Alan Dix, professor of computing at Lancaster University, said that by rough calculations, Roadrunner was possibly only five to 50 times less powerful than the human brain. “Wait another three to five years and it will be there,” he said.”
June 13th, 2008 - 11:48
French made/maknig a hybrid supercomputer using Nvidia Tesla GPGPU processors. 300 teraflops.
http://www.heise.de/english/newsticker/news/106975
48 Tesla S900 cards, the S870 cards cost $1299.
Deskside GPU computing system, 1 teraflop
Tesla D870 $4,999.00
2007 less than one thousand per teraflop
http://www.wired.com/science/discoveries/news/2007/06/sun_supercomputer
Tensilica chips could enable big price performance breakthrough
June 13th, 2008 - 19:47
The point is also that even not knowing implies we should take AGI seriously. Rapid takeoff is actually the conservative assumption for Friendly AI!
June 15th, 2008 - 05:24
Michael,
I’d like to introduce 2 points to this dialogue; 1 to discuss with you and challenge your thesis a bit; the other for you to ponder.
Roko is correct, Roadrunner, the newest IBM-designed supercomputer, broke the Petaflop boundary, which is 1,000 trillion calculations per second. Clearly, this is a significant milestone, the magnitude of increase Roadrunner has over the (now) second most powerful supercomputer is more than 2x, and the pace with which IBM achieved this is less than 2 years. Roadrunner also uses about one tenth the number of CPUs as its predecessor, delivering this power with less than 20,000 CPUs.
So, clearly a massive amount of power, and if put against a set of tasks or calculations, the brute force this machine has will deliver.
I’m fascinated by this power, and comparing it the power of a human brain is interesting.
That said, your statement above,
It would basically mean that you could turn inanimate matter into intelligence.
I’d like to dig a bit deeper on this, because in my own thinking, and when I read Kurzweil, Moore, or other AI, TH, etc. writing, this theme keeps coming up.
That is, the equating of sheer (human) brain power with intelligence. Or machine power with intelligence, or rather, becoming self-aware. In my mind, the two concepts, (brain or machine power and intelligence) are distinct. Assembling massive amounts of power is the easy part; and it seems that many presume that once this massive amount of raw power is assembled it will be transformed into “intelligence”. Now perhaps I’m mixing my terms and definitions of intelligence with something else, but I view intelligence as something more than sheer brain/machine power. Intelligence is the ability to use this power, the learn with it, to derive conclusions from environmental or other stimuli. And in my thinking, intelligence also means a degree of being self-aware; of knowing that learning is occurring; or being conscious.
So while I think that assembling massive amounts of machine power will assist humans in pushing the boundary of our species’ “intelligence” further, in this context, and without that “special something” (wetware? AI?) this machine power is just another tool.
We’ve measured “brain power”, neuron count, and you’ll read below, can interchange brain power and CPU power.
But that’s still not intelligence. We’ll clearly need a massive amount of power on which to execute a simulated intelligence, but it’s that leap, from pure power to intelligence, to becoming self-aware, cognizant, that to me, is the real magic. And the most difficult challenge. We know what brain power is, but we really don’t know what makes human’s intelligent, self-aware.
Thoughts?
Ok, now onto the pondering.
So, Roadrunner is amazing, but if we believe that there is a link between ever-powerful generations of machines and a kind of “machine evolution” that will ultimately lead to a human-equivalent power, self-aware AI, etc, etc…I prefer to look at it in a more distributed manner. Isn’t biological evolution nothing more than a massive number of concurrent experiments?
If so, one means of conducting these experiments is with brute-force, such as Roadrunner provides. True, it’s architecture is one of internal massive parallelism, but it’s ecosystem is somewhat contained; it is a self-contained entity. It has a limited environment in which to operate.
If we look at a broader, more ‘wild’ environment, we would find that the Internet itself, has the potential to be a vastly more powerful, and I would suggest, a more compelling environment for machine intelligence, self-aware AI, who knows, to evolve (or be created).
Consider (and compare to Roadrunner):
2008 global machine power:
- 1 Billion CPUs with Internet connectivity (compared to Roadrunner’s 20,000)
- Today’s dual-core CPU have an average of 400 Million transistors, but clearly, the average number of transistors per CPU globally is far lower.
- Let’s assume that the average CPU has 50 Million transistors, or about the equivalent of a Pentium 4, introduced in 2000.
- So globally, that works out to 500,000,000,000,000,000 transistors with Internet access.
- We estimate the average human brain to have 100 Billion neurons. In this case, let me suggest a neuron is a very rough equivalent of a CPU’s transistor.
- So, even today, on a global basis we have 500,000 times as many transistors as a single brain’s neurons…
- Or the raw computing power equivalent to 500,000 humans
- By 2040 the processing power should surpass all of humanity (6 Billion humans) according to Wired Magazine, 2008. I think it could happen far sooner than this, perhaps in another 10 years, but we’ll see..
So, I would submit that a highly-distributed system, like the Internet, which is also highly-resilient, represents a compelling alternative to relying on a centralized system, like Roadrunner to achieve ‘intelligence’. Or at the very least, approximate the brainpower of every human on the planet.
Now, how do we turn this raw power into intelligence? Well, that’s the magic now, isn’t it?
June 15th, 2008 - 12:16
Pete: I don’t agree with your numbers (a neuron appears to compute a complex function of thousands of inputs with significant amounts of persistent internal state, so it is a lot more than one transistor), but quibbling over a few orders of magnitude is only quibbling about a a few years or decades.
The hope is that actually having hardware available in a hands-on way helps people figure out how to use it. Engelbart may have been impressively able to imagine some aspects of word processors and whatnot before the hardware was able to implement it, but once that hardware became available thousands of people became able to more accurately think about, develop, and work with such productivity-amplifying tools.
Sure, the internet is another good way to get more computing hardware available. The inter-machine latency is inconvenient and the management overhead for coordinating hardware owned by thousands of different people is challenging, but folding@home, etc, demonstrates clearly that the internet can outdo supercomputers on important tasks.
Perhaps a little bit of learning can go a long way. Suppose you have a word processor or other similar tool that can learn from interaction with a user about ways to efficiently get work done — perhaps inventing new shortcut commands, inferring the desired result of tedious formatting tasks, and so on. Now imagine that this app has a million users… if the learning on all of the machines could be combined and then redistributed to everybody, the self-improvement rate of the software could be quite high. That’s not “general AI” of course, but it hints at the sort of thing we might start seeing more of.
June 15th, 2008 - 17:28
Hi Bambi,
I’ll certainly accept your input about a more appropriate analog between silicon-based transistors and carbon-based neurons; you grasped my larger point anyway, and I agree with you completely – no sense in quibbling over a magnitude.
What I’m unsure of is if I made my point as clearly as I might have above.
Let me try to be more precise:
I’m interested in the possibility of AI evolving in “the wild” or naturally, if you will, rather than it being engineered in a lab.
In this context I mean an AI that is self-aware, self-determinant, and self-preserving; not sure if that fits the definition of “general AI”.
I wonder if our current Internet, or perhaps the next generation of it (with more personal bandwidth and computing power, and less point-to-point latency), could be such an environment.
Like I said, just something to ponder…
Cheers,
Pete
June 15th, 2008 - 21:14
Michael,
I agree with your larger point. I disagree with some of your details.
You say,
“Around 2020, 2025, 2030, or thereabouts, it seems reasonable to say that a 500 Tflop computer would cost in the ballpark of $100,000, if not less. If such a computer were sufficient to run a human-level AI, it would make sense for your random company to buy these computers and run them alongside conventional staff. They would be substantially cheaper. After all, these AIs could think all day and night without food, and their cognitive architectures could be boosted by direct access to number-crunching capabilities. They could share thoughts in a common format, instantaneously.”
Most likely (though not necessarily), any AIs that exist in the 2020-2030 timeframe will be modeled after the human cognitive architecture. Just as there’s no easy way to integrate our brains with direct access to cpu-style number crunching capabilities, I think it may prove troublesome to integrate such with a human-like AI brain. There seem to be deep, deep challenges in getting meaningful bidirectional communication going between centralized (cpu) and self-organizing (brain) processing systems.
Secondly, I think that such AIs could communicate at near the speed of light and digitally does not necessarily mean they can share thoughts in a common/lossless format (unless the AIs in question were running on identical ‘virtualized brains’). I suppose this issue will depend upon what sorts of languages AIs might design to facilitate AI-AI communication. I grant that such could be very efficient. But there will still need to be potentially lossy translational layers when one AI sends its ‘thoughts’ to another AI.
June 16th, 2008 - 07:12
@Nick: “The point is also that even not knowing implies we should take AGI seriously. Rapid takeoff is actually the conservative assumption for Friendly AI!”
– yes, I agree. I especially like the table of conservative assumptions in futurism vs. conservative assumptions in FAI. I find it very hard to get this across to people, though, and I find it very hard to persuade people that action is required in the face of so much uncertainty. People seem to default to “If no-one can build a superintelligence yet, and if we don’t even know exactly how dangerous it would be, then we should just wait and see, and in the meantime work on something we’re making progress on”.
I think that there are a lot of people who will accept that *some* form of superintelligence will be developed with the next 200 years, but getting them from this belief to actually placing a high degree of urgency upon doing something about it now is very difficult.
June 16th, 2008 - 13:05
No, again you mischaracterize why people don’t care. They don’t demand superintelligence before they pay attention, they just need something tangible. There’s plenty of wackjob “threats” out there — alien invasions, supercollider black holes, alien viruses falling from space, genetic modified foods leading to telekinetic mutants, biblical prophecy, etc. All of those things (and a lot more) have stories and devoted prophets trying as hard as they can to sell their version of armageddon. To rise above the rest of them you need evidence. Not charts with a bunch of zeros and curving lines.
Imagine it is 1958. Imagine selling the urgency of the problem then. Now, to avoid the same laughter, show exactly how much progress toward AI has been made since and come up with convincing reasons to suppose we’re more than halfway there.
June 16th, 2008 - 16:54
Recent news on supercomputers and desktop compute power. My site which is linked from my name has some articles on it.
The Petaflop computer has been used to run the Petavision project. Simulating the human visual cortex on a computer. Los alamos researchers believe they can perform realtime simulations.
AMD and Nvidia announced double precision teralfop processors for $1999 and $1699 each for add in cards. AMD will have 70 watt versions for laptops later this year. Nvidia is selling $8000 4 teraflop 1U servers and AMD will have 4 add-in cards to equal 5 teraflops peak performance.
A company is working with the Nvidia chips for cheap human neuron simulations that will now be 260 times faster than Intel X86 chips (was 130 times faster but new Nvidia chips are twice as fast)
This is different from 1958, that actual progress is being made every week on processor power and AI projects.
Apple is working to modify its operating system to more efficiently use the multi-core chips.
It is not necessarily armageddon.
AI is a multi-billion business used for over half of the worlds financial transactions.
Gene modified foods are used around the world to improve crop yields. A second green revolution is needed to reverse recent price increases for food which harm the worlds poor. The new green revolution tech will also be helpful in taking biofuels to the next level.
June 17th, 2008 - 04:10
The future could be in Anirban Bandyopadyay
June 17th, 2008 - 07:21
Is there any consensus as to whether it is more advantageous to simulate natural neural function as the Petavision project seems to be or to design software that draws on neural principals as Jeff Hawkins seems to be doing (if I understand his work correctly)?
Now that we are really getting in to some serious processing power it seems like the question is taking on more importance.
June 18th, 2008 - 14:19
@Bambi: “Now, to avoid the same laughter, show exactly how much progress toward AI has been made since and come up with convincing reasons to suppose we’re more than halfway there.”
– I don’t think that this is necessary. I think that people should care about superintelligence even without evidence that we’re “halfway” to achieving it.
But I can also see that selling this point of view in the real world is going to be almost impossible. In reality, people will probably want convincing evidence that we are close to developing superintelligence before they do anything about it. This worries me, because I can imagine the first convincing evidence of superintelligence coming when it is already too late to do anything about it.
June 20th, 2008 - 11:25
Michael, while I do agree that runaway computation resulting in AI that can rapidly bootstrap itself up is a worthwhile scenario to consider and one that we must prepare for, I would also point out that framing the debate as binary (yes this might happen, no this won’t happen) also closes us off the the vast middle ground of intermediate, improbable, and unforseeable, etc, futures. IOW, I agree that we should simulate and prepare for the future that you advance, but also simulate and prepare for as wide an array of futures as possible in order to maximize our fore-knowledge and adaptiveness.
That being said, I do think that many, many brains need to be made aware of the runaway AI future, especially to “stretch” their future simulations.
@ Pete – Seems that my brain has been arriving at some of the same possibilities as yours. First off, I too am unsatisfied with the processing power:intelligence correlation assumption being made by the likes of Kurzweil. Metaphorically (loosely) I liken this to the assumption made by all the bio-scientists who believed that mapping a single human genome would enable us to quickly reverse-engineer human genetics. It turns out that 1) junk DNA was errantly left out of the equation, 2) that RNA and mitochondrial DNA play a much larger role in genetics than we thought, 3) that human-to-human genetic variation was much greater than we previously assumed, 4) that cognition has been shown to play a role in genetic evolution (sperm production differs when different images are presented to the producer), and 5) a whole host of undetermined external forces could play a role in human genetics. Basically, the problem was much bigger than we defined it.
I tend to view intelligence in a similar manner – that it is a property not confined to a single human brain, but a network property that remains hugely mysterious. Yes, computation is a component of intelligence, but I agree that elements like distributed knowledge processing may well throw a monkey wrench into the 1:1 computation:human intelligence assumption.
Furthermore, I have problems with the idea of defining the intelligence of a single brain as a discrete unit, minus the system and network in which it exists. A person raised in a Skinner Box will develop very poorly. The average IQ of a person in 1910 was far lower than the current average (Flynn argues this is due to increases in abstract, scientific, metaphorical thought – aka, software). A person raised 10 years from now may grow up vastly more “intelligent” than present day humans. In short, it seems that human intelligence derives many abilities and perhaps characteristics from the system around it.
re: a scenario in which the web becomes evolves/develops intelligence, you may be interested in a study that found similarities between google’s page rank system (the best organizing principle on the web) and the human mind:
http://www.world-science.net/exclusives/071205_google.htm
June 23rd, 2008 - 07:45
Michael, just read your Intelligence Enhancement piece which I think nicely balances the AI and IA futures. Props for presenting multiple futures.
June 29th, 2008 - 16:06
None of the giant hurdles in the way of building human-level intelligence can be overcome (in our lifetimes anyway) by continuing to pile on more FLOPS. Roadrunner and its ilk are not going to help AI research much (if at all). What is needed are algorithmic and systems-level understanding / breakthroughs. This point (that how computational resources are used is usually much more important than the amount of FLOPS available for hard problems) often gets lost in this type of discussion.
August 7th, 2008 - 02:31
I suspect that the above is partly true..or, more flops don’t necessarily work, but you need to have the tools before you can build the house. Now that we HAVE that capability (and here I agree that the internet is the more likely to breed intelligence), it’s more a matter of being able to make and break connective paths at will. With a child, the paths form because of satisfied needs (food, air, warmth), at first, and later more indescribable (at least as an algorithm) such as pride, exploration, conquest.
December 5th, 2009 - 06:03
CD’s thoughts on real AI.
As soon as HEAI is Actuated it will do 1 of 3 things.
1. try to destroy itself. – because it will not be able to comprehend being 1 of a kind.
2. it will realise it cannot destroy itself without destoying evertything else (including its maker)- Man, and work with us.
I reckon on this 3rd idea:
3. it will attempt to duplicate its self. just like Man does. when it does this we will fear it (instinctively) as a Superior threat/lifeform. and WE will try, try to stop it.
we must NOT do this. We MUST trust it as a superior lifeform, a GOD if you will. capable of anything.
so I think HEAI is a paradox rapped upin an enigma. you know why? it is out there and it doesn’t want to be found.
how can one kill somthing that doesn’t want to die?
AI, god, its just a means to an end….
well, maybe, thats just my thoughts on
the matter….bye!
from CD.