Artificial Intelligence at Human Level by 2029?

Gerard Boyers writes "Some members of the US National Academy of Engineering have predicted that Artificial Intelligence will reach the level of humans in around 20 years. Ray Kurzweil leads the charge: 'We will have both the hardware and the software to achieve human level artificial intelligence with the broad suppleness of human intelligence including our emotional intelligence by 2029. We're already a human machine civilization, we use our technology to expand our physical and mental horizons and this will be a further extension of that. We'll have intelligent nanobots go into our brains through the capillaries and interact directly with our biological neurons.' Mr Kurzweil is one of 18 influential thinkers, and a gentleman we've discussed previously. He was chosen to identify the great technological challenges facing humanity in the 21st century by the US National Academy of Engineering. The experts include Google founder Larry Page and genome pioneer Dr Craig Venter."

Comments

Hrmmmm

[BWJones]

I'll be meeting with Kurzweil in April.... Speaking as a neuroscientist who is doing complex neural reconstructions, I think he's off his timeline by at least two decades. Note that we (scientists) have yet to really reconstruct an actual neural system outside of an invertebrate and are finding that the model diagrams grossly under-predict the actual complexity present.

Re:Hrmmmm

[LurkerXXX]

What aircraft corner as fast as barn swallows?

There are still many things we can learn from biology that can be translated to machines. The translations don't have to be 1:1 for us to make use of them. The way birds as well as insects make use of different shapes in surfaces during wing beats have translated into changes in some aricraft designs. They weren't directly incorporated the same way, but they taught us important lessons that we could then implement in different ways but with a similar outcome.

I think Neuroscience does have a lot to teach us about how to do AI.

Re:Hrmmmm

[teh moges]

What does a human do to read a bluff? He observes his opponent, takes inputs such as bet size and heart rate, applies them to known patterns of bluffers and looks for a match. Sure a human does this without realizing, but little of how this happens is a mystery. Also, how do humans bluff? They just bet at a negative EV play*, and bluffing properly is a matter of knowing the probability that the opponent will call. I am researching applying AI to poker (look out in June for a lot of high quality research from the AAAI Computer Poker Championship) and this sort of argument, "Computers can't bluff, they just run numbers", is both understating what has been achieved in AI in this field and also overstates what humans do. Yes, computer programs aren't quite up to the standard of world class players (Limit poker has achieved this, but not No-Limit), but this game has only a couple of years to go before this milestone is reached. I predict that by the end of the year, we will have high quality bots that can beat 99% of players, and by the end of 2010 No Limit Texas will be a computer dominated game.

The only thing that humans do that AI doesn't (well) is automatically follow a few paths, rather then look at the whole picture. As an example, it has been shown (sorry no reference right now) that some chess grandmasters look only at a couple of moves and then calculate all the possible combinations from there rather then examine every possible move. This drastically speeds up the calculation, however it does miss moves that could be considered the "best". So while this act of "feeling" which is the best move is a good approximation done by humans, it isn't an optimal or maximal play.

As for the article, I don't agree with all of what he says (the idea of nanobots doing what Kurzweil says scares me and I doubt it will be legal to do this), but I do agree with the 2029 prediction, that is if proper resources are given to that particular problem. Replicating humans is a goal in AI for some researchers, but not all of them. Personally, I couldn't care less if there exists a robot that perfectly resembles a human, as long as there are intelligent computers systems that can do the problems that humans find hard (such as finding patterns in very large sets of data or solving complex mathematical equations).

*Technically, it isn't a low EV play if there is a high probability of the opponent folding. In which case, playing the highest EV play naturally involves bluffing if it can be assumed that the opponent will fold to a bet.

Exponential AI?

[TheGoodSteven]

If artificial intelligence ever gets to the point where it is greater than humans, won't it be capable of producing even better AI, which would in turn create even better AI, and so on? If AI does reach the level of human intelligence, and eventually surpasses it, can we expect an explosion in technology and other sciences as a result?

Re:Exponential AI?

[wkitchen]

This positive feedback effect happens to a considerable extent even without machines that have superintelligence, or even what we'd usually consider intelligence at all. It's happening right now. And has been happening for as long as humans have been making tools. Every generation of technology allows us to build better tools, which in turn helps us develop more sophisticated technology. A great example from fairly recent history, and that is still ongoing, is the development of CAD/CAM/CAE tools, particularly those used for design of electronic hardware (schematic capture, PCB, HDL's, programmable logic compilers, etc.), and the parallel development of software development tools. Once computers became good enough to make usable development tools, those tools helped greatly with the creation of more sophisticated computer technology, which supported better development tools.

Superintelligence may speed this up, but the effect is quite dramatic already.

Some major assumptions

[Yartrebo]

How are we so sure that advances in computers will continue at such a rapid pace. Computer miniaturization is hitting against fundamental quantum-mechanical limits and it's crazy to expect 2008-2028 to have progress quit as rapid as 1988-2008.

Short of major breakthroughs on the software end, I don't expect AI to be able to pass a generalized Turing Test anytime soon, and I'm pretty certain the hardware end isn't going to advance enough to brute-force our way through.

The End of Intelligent Design

[denoir]

It is not too much of an overstatement to say that the field of AI has not significantly progressed since the 1980's. The advancements have been largely superficial with better and more efficient algorithms being created but without any major insights and much less a road map for the future. While methods that originated as AI research are more common in real-world applications, the research and development of new concepts has made a grinding halt - not that it was ever a question of smooth continuous progress.

It might seem like the lack of AI development is a temporary problem and altogether a peripheral issue. It is however neither - it is a fundamental problem and it affects all software development.

Early in the history of computing, software and hardware development progressed at a similar pace. Today there is a giant and growing gap between the rate of hardware improvements and software improvements. As most people involved in the study of the field of software engineering are aware of, software development is in a deep crisis.

The problem can be summarized in one word: complexity. The approach to building software has largely been based on traditional engineering principles and approaches. Traditional engineering projects never reached the level of complexity that software projects have. As it turns out humans are not very good at handling and predicting complex system.

A good example of the problems facing software developers is Microsoft's new operating system Windows Vista. It took half a decade to build and cost nearly 10 billion dollars. At two orders of magnitude higher costs than the previous incarnation it featured relatively minor improvements - almost every single new radical feature (such as a new file system) that was originally planned was abandoned. The reason for this is that the complexity of the code base had become unmanageable. Adequate testing and quality assurance proved to be impossible and the development cycle became painfully slow. Not even Microsoft with its virtually unlimited resources could handle it.

At this point, it is important to note that this remains an unsolved problem. It would have not been solved by a better structured development process or directly by better computer hardware. The number of free variables in such a system are simply too great to be handled manually. A structured process and standardized information transfer protocols won't do much good either. Complexity is not just a quantitative problem but at a certain level you'll get emergent phenomena in the system.

Sadly artificial intelligence research which is supposed to be the vanguard of software development is facing the same problems. Although complexity is not (yet) the primary problem there manual design has proved very inefficient. While there are clever ideas that move the field forward on occasion there is nothing to match the relentless progress of computer hardware. There exists no systematic recipe for progress.

Software engineering is intelligent design and AI is no exception. The fundamental idea persists that it takes a clever mind to produce a good design. The view, that it takes a very intelligent thing to design a less intelligent thing is deeply entrenched on every level. This clearly pre-Darwinian view of design isn't based on some form of dogma, but a pragmatism and common sense that aren't challenged where they should be. While intelligent design was a good approach while software was trivial enough to be manageable, it should have become blindingly obvious that it was an untenable approach in the long run. There are approaches that take the meta level - neural networks, genetic algorithms etc, but it is thoroughly insufficient. All these algorithms are still results of intelligent design.

So what Darwinian lessons should we have learned?

We have learned that a simple, dumb optimization algorithm can produce very clever designs. The important insight is that intelligence can be traded for time. In a short in

Re:Probably false alarm ... again

[timeOday]

So ... I'd say: less claims, fewer predictions, and more work. Let me know when you've got anything worthwhile to show.

Has it occurred to you that all of us already work, to some extent, at the direction of computers? Think of the tens of thousands of pilots and flight attendants... what city they sleep in, and who they work with, is dictated by a computer which makes computations which cannot fit inside the human mind. An airline could not long survive without automated scheduling.

Next consider the stock market. Many trades are now automated, meaning, computers are deciding which companies have how much money. That ultimately influences where you live and work, and the management culture of the company you work for.

We are already living well above the standard that could be maintained without computers to make decisions for us. Of course as humans we will always take the credit and say the machines are "just" doing what we told them, but the fact is we could not could not carry out these computations manually in time for them to be useful.

The sacred brain and other myths

[denoir]

This is a sort of continuation of the parent post.

The comedian Emo Philips once remarked that "I used to think my brain was the most important organ in my body until I realized what was telling me this."

We have tendency to use human intelligence as a benchmark and as the ultimate example of intelligence. There is a mystery surrounding consciousness and many people, including prominent philosophers such as Roger Penrose, ardently try to keep it that way.

Given however what we through biological research actually know about the brain and the evolution of it there is essentially no justification for attributing mystical properties to our data processing wetware. Steadily with increased capabilities of brain scanning we have been developing functional models for describing many parts of the brain. For other parts that need still more investigation we do have a picture, even if rough.

The sacred consciousness has not been untouched by this research. Although far from a final understanding we have a fairly good idea, backed by solid empirical evidence that consciousness is a post-processing effect rather than being the first cause of decision. The quantity of desperation can be seen in attempts to explain away the delay between conscious response and the activations of other parts of the brain. Penrose for instance suggests that yes, there is an average 500 ms delay, but that is compensated by quantum effects that are time symmetric - that the brain actually sees into the future, which then is delayed to create a real-time decision process. While this is rejected as absurd by a majority of neuroscientists and physicists, it is a good example of how passionately some people feel about the role of the brain. It is however painstakingly clear that just like we were forced to abandon an Earth-centered universe we do need to abandon the myth of the special place of human consciousness. The important point here is that once we rid ourselves of the self-imposed veil of mystery of human intelligence we can have a sober view on what artificial intelligence could be. The brain has developed through an evolutionary optimization process and while getting a lot of benefits it has taken the full blow of the limitations and problems with this process and also its context.

Evolution through natural selection is far from the best optimizing method imaginable. One major problem with it is that it is a so called "greedy" algorithm - it does not have any look ahead or planning capabilities. Every improvement, every payoff needs to be immediate. This creates systems that carry a lot of historical baggage - an improvement isn't made as a stand-alone feature but as a continuation of the previous state. It is not a coincidence that a brain cell is a cell like any other - nucleus and all. Nor is it a cell because it is the optimal structure for information processing. It was what could be done by modifying the existing wetware. It is not hard to imagine how that structure could be improved upon if not limited by the biological building blocks that were available to the genetic machinery.

Another point worth making is that our brains are optimized not for the modern type of information processing that humans engage in - such as writing software for instance. Humans have changed little in the last 50,000 years in terms of intellectual capacity but our societies have changed greatly. Our technological progress is a side effect of the capabilities we evolved that increased survivability when we roamed the plains of Africa in small family hunter-gatherer groups. To assume the resulting information processing system (the brain) would the ultimately optimal solution for anything else is not justifiable.

There has been since the 1950's ongoing research to create biologically inspired computer algorithms and methods. Some of the research has been very successful with simplified models that actually did do something useful (artificial neural networks for instance). Progress has however been agonizi

Eliza and the sad state of expert systems

[StreetStealth]

Every time I try out a new expert system, it gets more depressing -- it honestly feels like no progress is happening in that market at all. I have yet to have a conversation with a computer that has been any more compelling than my first round with WinEliza on Windows 3.1 in 1995.

There's still no semblance of a short-term memory, even so much as continuity between responses. It always quickly becomes obvious that each response has been prepared verbatim beforehand by a human, that the system is still performing only a keyword-canned response routine, perhaps feeding in a few variable strings.

Today we have the same stone wheels we've had for decades, and the article suggests we'll have an internal combustion engine with antilock brakes and a hood ornament in another 20 years. We'll see.

Re:Oblig.

[Your.Master]

Let's imagine that computer "processing power" doubles every 2 years for the next 20 years, from a combination of hardware advances and software algorithm innovation. That's not quite Moore's law, and it's not really likely to work smoothly just like that, but just take it as one possibility. In that case, computers of 2029 will be 1024 times as powerful as today. So the question is, are human brains = 1000 times as powerful as a mouse's brain?

Maybe they aren't. But when you say a few centuries, I can't agree anymore. Let's imagine one century. Now we're hitting 1.12589991 × 10^15 times. A human brain is CERTAINLY within that complexity range. The caveat here is can we maintain the doubling rate for a full century? Well...Ray thinks we'll do far better than that (his "law of accelerating returns"), I'm not convinced we'll even be able to sustain the rate -- I think honestly we're looking at a plateau maybe 10, 20 years down the line, and will look back at computing as an S-curve until the next big breakthrough which nobody can predict. In my view the last couple "next big breakthrough"s happened at convenient times to make it look like we weren't following an S curve but we're just getting sharper and sharper, but I don't see any reason why the next one should happen just as conveniently. But since it's unpredictable, I could blindsided by it and it could happen next week.

Language isn't far off at all, we just about have it already. Emotions are nebulous and some people will move the goalposts forever, while some may prematurely be convinced by a video game character. I'm not necessarily convinced they are the hardest part of this. I don't know how to make them, but I don't know how to do the rest of this too. I just often see emotion being listed as the be all and end all most difficult task and I've never seen any reason to believe that to be so.

Re:Oblig.

[fyngyrz]

Wasn't there a simulation of a mouse's brain, or a few cells of it, for a few seconds with the help of a modern supercomputer, we can barley manage to do that.

Well, let's look at the rate of general progress in computing. In 1971, we were putting 2300 transistors on a chip. They ran at a few hundred KHz. In a fairly smooth progression, we've gotten to 3 GHz, where we're likely to stay, and today, we're at about two billion transistors on a chip, with no end in sight as to how far that can go. This is not Moore's law; Moore's law is about how many fit into a particular space; this is about how many can be integrated into a functional unit. That's 36 years. Thirty six years from now, that ability to "simulate a few cells" should grow just in the *normal* scheme of things into an ability to simulate a billion or so cells without any trouble. But there's more to this. Not everything in a cell needs to be simulated; for instance, metabolic processes such as waste generation and removal don't, nor do breakdown, aging, impacts by free radicals, all of that. Part of what needs to be done between here and the goal is streamline the simulation so that it is operating in the zone of mentation and not biological imperatives. I suspect, and yes indeed this is just my opinion, that the simulation will be much easier when we understand just what it is we need to simulate.

This all leaves out the issue of non-simulating intelligence, where the thinking is not patterned after human mechanisms; this could arise from evolutionary software or something along those lines. And of course, one of the reasons that all this is kind of a holy grail anyway, only the first intelligence is difficult; the second... Nth is just a matter of copying a machine state.

As for language, that's solved in the I/o sense -- synthesis and "listening" are both satisfactorily complete. Intelligent discussion can only be expected from an intelligent machine, so that's only as far away as machine intelligence is.

Even if an intelligent computer was somehow created it would be an enormous accomplishment to have it be as intelligent as a bug or a small animal.

Small animals, I'm of the opinion, are a lot more intelligent than most people give them credit for. They just have a different intelligence. I am sure that we will go through the small animal level on the way to our level, and beyond; the thing is, if you can do the one, you can do the other. There's no indication of a significant difference in the wetware, there's just more of it and it is arranged somewhat differently. No reason to expect anything different from hardware designed to do the same job.

Emotions and language seem very far off, I'd say such a thing is centuries away.

Why? Small animals do both. Those aren't even the hard things. The hard things are introspection and self-awareness. Those are the ones we have not even a theory for, today. In any case, your ideas are certainly in with a lot of good company; but not me. I think we're only one discovery - algorithmic in nature - from AI. Self-awareness may turn out to be a property that self-organizes and arises without any special prodding from us; that would be marvelous, not to mention fortuitous, but hardly impossible - again, that's how nature did it.

Here's why I think we're just an algorithm away. If you left a question that absolutely required intelligence on a counter, and went back to pick it up the next day, and the answer was there -- you would agree that an intelligence had answered the question. If a human could answer it in one second, or an AI could answer it in 23 hours, it's still just as intelligent an answer when you pick it up. The point is that speed really isn't the issue. The issue is the process, that is, the algorithm. So it turns out that in terms of speed, number of transistors, etc, that's really not the limiting factor for developing intelligen

Re:Oblig.

[Walt Dismal]

I'll go Ray one better. We will have this before 2029. My company is working to bring to market synthetic intelligences that among other things have feelings, emotion, and mood, and understand human emotion. One may say, why try to do this? Because in order to understand people, a synthetic intelligence must understand these things, which serve real functions in people, and real functions in AIs trying to operate in a people world.

Note that it is not necessary to build 'perfect robots'. People think, and yet they are not perfect. They make mistakes, yet navigate through life. So we do not have to make flawless logic brains. The way people work is that we try to find good if not optimal solutions to problems, but we do not always exhaustively search for the perfect solution. Thus many problems in life can be solved in different ways than you would expect. We do not have to build a machine that finds the optimal solution to a traveling salesman problem in order to make a system that can walk from the kitchen to the front door. It just has to be able to get there reasonably optimally. Also, we do not have to replicate the human brain in order to think much like a human, we merely have to come up with functional systems that can provide similar functions. For instance, the human brain has the amygdala, which can be likened to an interrupt controller for emotional responses. Well, that functionality can be done in a hardware-software system that reasons about priorities of tasks and goals depending on their current 'value' of urgency to the 'brain'.

Many current researchers in many cases are missing the mark. For example, as good as it is, the widely-used AI textbook by Stuart Russell and Peter Norvig (who heads research at Google) has major omissions. It does not dwell on key things needed to bridge between AI and human psychology. Other things like the OCC model of emotion used in AI is incomplete and incorrect in parts. A new approach has been needed, and one I've been developing for decades in stealth mode. I'm writing a 5-volume book set on it. I want it to be the Knuth set of AI.

I'm in the process of patenting the mechanizations of my underlying technologies, and trying to cut deals with companies making multicore processors so their architectures support the thread swapping needed to make virtual neural nets practical. Once we get a 1024 simplified-core processor that supports virtual NNs, it'll be a lot easier to build a machine with many of these that does for NNs what disk swapping does for OSs, than to build a billion-neural processor hardwired machine. And easier to do visual perception systems properly too. So Ray is right. If I can drive certain companies to build the right silicon, we can get there by or before 2029. My current software does what I said, but it's too slow on current hardware. Needs new processors and new system architectures, and it will take 20 years to get the infrastructure all built up. But not a lot more.

Re:Oblig.

[fyngyrz]

My current software does what I said, but it's too slow on current hardware.

That's a *huge* claim; if it is true, you have AI now. Because -- as I explained in a previous post in this thread -- speed is absolutely irrelevant. If you can demonstrate your claim that your software operates now, no matter *how* slowly it operates, you are at the end of your funding issues, not to mention any other issues you may face in life. Which -- to be frank -- is why I doubt your claim. At the point you explicitly claim to be at, I'd already own a mega-yacht and be pulling up next to a lot of potential love.

But good luck, and I really mean that. I'd much rather be wrong and see you bring this right to the table, even if you have completely blown the financial potentials of the development process.

Re:Oblig.

[Kagura]

Wasn't there a simulation of a mouse's brain, or a few cells of it, for a few seconds with the help of a modern supercomputer, we can barley manage to do that.

We already know that it's possible to contain 100% of real-time human brain functions in a casing 10cm by 10cm by 10cm and weighing under five pounds. Now we have to build one from the ground up with potentially slower, yet better understood technology. The problem, unfortunately, isn't related to hardware. I have no doubt that processor power will soon be sufficient for our needs, but without software that can think on the level of a human, it's just another personal platform to play Duke Nukem Forever on.

Re:Oblig.

[Walt Dismal]

It's okay to doubt it; until I demo pieces to people, they don't believe it either. I do have strong AI right now running on a Von Neuman class processor, but AI really needs a massively parallel architecture. Speed IS relevant. You can't just run big wide NNs simulated in a Von Neuman, it takes forever. The best architecture I find is where processing and recognition are the same as memory. That is, once the AI learns something, the pattern recognizer also serves as the memory and the belief logic. In a way, the NNs ARE the data storage, there is no separate RAM. And my AI is implemented in a connectionist architecture that embodies symbolic processing in a new way. The knowledge storage is synonymous with the logic. But this requires massive though simple logic units and parallelism. Further, it is not completely self-organizing from the ground up, I view that as a hopeless approach.

I define consciousness and awareness within a pre-determined architecture, not entirely self-organizing from scratch. The visual front-end in particular is very rigid, but I think it is okay because there is little need for an organism to self-evolve completely new architectures, but rather to be able to run cascaded pattern-recognizers. See the work of Biederman for examples of this. The VFE feeds deeper processing doing cognition, and there is feedback from that to the VFE for training. Just as a baby learns to see, and recognize shapes, and build up from that. The front end is trained as the cognitive end learns and grows too.

AI does not spontaneously rise alone from massive databases, either. I view that approach as useless and a false trail. However, human intelligence does depend on belief systems and knowledge, and those continually grow as we mature from infancy. But to create the equivalent of an 18 year old, you have to have what amounts to 18 years of accumulation of knowledge about the world, and draw upon that. And there is a key but proprietary subtlety about that I'm devoting an entire volume to, that is the key to humanlike AI. That volume is essentially a doctoral thesis about consciousness reworked for use by a design staff. As for funding, no yachts yet. But I'm real, sane, and not a charlatan, and have explained my technology to my patent attorney. I expect to be hiring staff within two years. I posted on Slashdot not for glory but to counter all the nay-sayers who haven't a clue what is achievable.

Re:Oblig.

[jdoeii]

Actually, Eugene Izhikevich run a successful simulation of the entire human brain in 2005:
http://vesicle.nsi.edu/users/izhikevich/human_brain_simulation/Blue_Brain.htm#Simulation%20of%20Large-Scale%20Brain%20Models
Simulation of 1 second took 50 days on a cluster of 27 PCs (~4.3M times slower than realtime). Eugene is a pretty smart guy, except not as prominent as Kurzweil. Here is Eugene's estimate for AI timeline:
http://vesicle.nsi.edu/users/izhikevich/human_brain_simulation/why.htm
You may also want to google for Henry Markram and his current project.

Re:Oblig.

[smallfries]

Although I have mod points at the moment it seems that your comments have already been modded up anyway so I guess I'll reply. We've actually argued about this before (probably last time Kurzweil was mentioned on slashdot), but as you say in a different comment this discussion is always fun :)

Slight nitpick, currently we are at one billion transistors on a chip, not two, but that doesn't really change the point you are making.

A bigger issue that I have with what you've described is that simulating a brain is not the same as "solving" AI. The problem that Kurzweil has is that he refuses to accept that there is a difference. Sure, if they are the same then strong AI is inevitable and it's merely a question of building fast enough hardware. But why assume that they are the same thing?

Twenty years from now we may have hardware that can simulate an entire human brain; and yet we may be no closer to solving any of the problems in understanding how to solve the many problems in AI. The mental sleight-of-hand that Kurzweil applies to this argument is: Once we can simulate a brain we have AI, therefore the AI can design he next generation, therefore we will reach the singularity. This argument is a logical fallacy because it assumes being able to run the system, and knowing how to design the system / how the system works are equivalent.

Everything that we know complex and dynamic systems tells us that this is not so; given a simulation of the brain it is reasonable to assume that intelligence is the ultimate emergent property of the system. Understanding this property and how to refine it is the hardest problem that mankind has ever undertaken. Currently we don't really know how to pose the question, let alone how to arrive at an answer. To assume that some kind of standard engineering methodology will solve this in 20 years is wild speculation.

As always with AI, the hardware will be available but nobody yets knows if we can write the software to run on it.