Will artificial intelligence bring about a technological singularity in a soft take off? Ray Kurzweil at the 2006 Singularity Summit at Stanford gave an overview of smooth doubly exponential progressions that he believes could lead to such an outcome. While his projections are considered radical by some observers, it is often because they are thinking linearly and leave out the historically accurate exponential perspective.

The following transcript of Ray Kurzweil’s Singularity Summit at Stanford presentation entitled “The Singularity: A Hard or Soft Takeoff?” has not been approved by the author. Video and audio are available on the Singularity Institute website.

The Singularity: A Hard or Soft Takeoff

Thanks for all coming out. I’m looking forward to a spirited discussion of these events. Speaking of how AI solves a problem, we then tend to put down the problem, it’s not that important after all. AI has actually been defined as the set of problems that we haven’t solved yet. As soon as we solve it, it loses its mystery. Intelligence has a mystery, so if we know the solution of the problem, that can’t really be intelligence. It was very mysterious how a car could actually be driven by intelligence, only humans could do that.

That was actually challenged recently at the Gilder Conference, because I had predicted a car would do that and it was pointed out how pathetic all the DARPA cars were running off the road quickly. I said, “There will be a surge of progress at the end. Soon you’ll see this happen.” Thanks to Sebastian and his team, his car led five cars that completed that course. Of course, we say, “That’s not really intelligence, after all.” One by one, I think there’s no silver bullet, AI is getting stronger. I’m going to make that point again this morning.

I would like to start by reading a passage this morning from my book and actually demonstrating a new product. There are a couple hundred blind guys and gals who are running around, reading the labels on their clothing, signs on the wall, soup cans. It corrects for three different degrees of freedom of rotation and tilt. We developed this with the National Federation of the Blind. We did the research five years ago and anticipated what the hardware would be like. We took a large computer and simulated what we thought would be available on a PDA five years hence.

Those projections were accurate. A high-end camera that had four megapixels and the hardware was there. The software was not. We really could not manage the requisite algorithms to handle the vagaries of real-world print taken from a digital camera. Three different degrees of tilt and rotation, uneven illumination, curved images and so on was not feasible five years ago. It is one of many examples, and I will mention a number of others, there are hundreds in my book, of how software and AI has progressed. People say, “Where are all these AI applications?” It’s a little bit like people who go into the rainforest and say, “Where are all these species that are supposed to be here?” when there are fifty species of ants within fifty feet of them. But they are deeply integrated into the ecological infrastructure.

We have hundreds of examples of narrow AI deeply integrated into our econonomic infrastructure. Every time you send an email or connect a cell phone call, intelligent algorithms route the information. Intelligent algorithms design products, control just-in-time inventory, assemble products, land airplanes, guide intelligent weapons systems, make billions of dollars of financial transactions automatically everyday, and many other examples. These were research projects when we met in 1999 at the Spiritual Machines conference.

We are making progress. There is no silver bullet. But hardware and software are progressing at an exponential pace. In fact, seven years ago at this conference it was said that computers can’t even tell the difference between a cat and a dog. In fact, the progression in AI has been the opposite of the maturation of human intelligence. A little child will learn the difference between a cat and a dog. It’s not until we are adults that we can do things like solve mathematical theorems. Carnegie Mellon’s General Problem Solver in the 1950′s solved mathematical theorems that Russell and Whitehead had been unable to solve. So it was felt that it would not be long before computers could do anything.

In fact, the really challenging parts of human intelligence are the tasks that young children can do: Particularly our pattern recognition. That is the strength of human intelligence. That’s how Kasparov plays chess. He doesn’t build this tree of billions of move, counter-move positions. Deep Blue does 300 million board positions a second, in terms of move, counter-move sequences. He was asked, “How many can you do a second?” He said, “One. Maybe less.” How is it that he holds a candle to machines? It’s the power of pattern recognition. That is really the heart of human intelligence.

Computers can now tell the difference between a cat and a dog. We are going to be adding a feature to this product so that a blind person could snap a picture and it will tell them, there’s a cat in front of you, there’s a lamp to your left. Your ex-wife is over to your right. We are in fact adding face recognition.

Why is it that we can tell the difference between a cat and a dog now with a computer that we could not do that seven years ago? Two things: pattern recognition algorithms have gotten steadily more sophisticated, and we also have this tremendous data mining that we didn’t have seven years ago. You need a lot of data to train these pattern-recognition algorithms. If you want a million pictures of cats and dogs, you know exactly where to get them. There are in fact three million pictures of dogs on Google. It’s a nice game to try to guess how many pictures of different things are up there. (There are more pictures of cats – 3.6 million.) You can train these algorithms. You would be hard pressed to find a thousand pictures of cats and dogs seven years ago.

The National Federation of the Blind, with whom we worked with on this project, has two projects: this pocket-sized reader and their other project is a car that blind people can drive. When that was announced four years ago, it was considered ridiculous. It’s not considered ridiculous anymore. We’re not ready to turn the keys of cars over to blind people, but it is now seen to be a very reasonable project. Sebastian, they are going to be in touch with you soon, because I gave them your name.

Underlying this is an exponential theory, the law of accelerating returns, which does say that information technology progresses exponentially. It is important to understand first of all that this is specific to information technology, so people very often pose an exponential trend that did not necessarily just keep on going. This is not relevant to things like population, which is not an information technology. In fact, information technology thwarted it, because economic progress we made has been largely due to the economic cost effectiveness of information technology. As soon as societies become wealthy, the population growth stops.

Information technologies progress, but doesn’t the exponential hit a wall, like rabbits in Australia? That’s true of information technologies. A specific paradigm will run out of steam and hit a wall, but then it leads to another paradigm, and I’m going to show you that a little bit later. Moore’s law was not the first but the fifth paradigm for exponential growth in computing. Isn’t there an ultimate limit for the ability of information technology to progress? I discuss that in detail in the book. There are limits, based on what we know about physics, to the ability of information technology to progress, both in computation, communication, biological technologies, and so on, but the limits are not very limiting.

One cubic inch of nanotube circuitry would be millions of times more powerful than the amount of computation required to simulate all regions of the brain, based on the most conservative estimates of that amont of computation. Nanotubes are not speculative. Nanotube circuits are working. There is a very large, dense nanotube circuit working that has self-organizing features that are set to hit the market next year. This is not speculative types of technology.

Even if you were skeptical about the ability of molecular three-dimensional computing to be feasible, and I don’t think that is actually a reasonable skepticism given recent progress, for just conventional two-dimensional chips the ITRS roadmap projects four nanometer features. We are now at 65 with 35 nanometer features being developed. 4 nanometer features are being projected for 2018. If you use application-specific integrated circuits, that will provide enough computation, 10^16 calculations per second, to emulate all regions of the brain, based on the most conservative estimates, for $1000. So you don’t need molecular computing, though that is quite feasible.

This is a scientific theory. I actually got into this because I realized that predicting technology trends was key to being successful as an inventor. I still do that primarily for this reason. We started this project for the reader five years ago because we projected it would be feasible in five years. I have a team now of ten people that gathers data in different areas, and I’ve been developing this theory for 25 years. It’s part of a broader theory of evolution.

But Moore’s law, which the public is aware of, is really just one example of many of this exponential progression. People think intuitively in a linear manner. I think that’s hardwired. We see what is in front of us and we project out linearly.

One scientist I was debating with him the pace of progress in brain reverse engineering. He said, “It took me 18 months to model this one ion channel and there are four other ion channels in this particular type of dendrite, it will be a century before we finish.” As if the computers we have to simulate these processes would stay the same and the pace of progress will be the same for the next hundred years.

The reality is that we are doubling the spacial resolution in 3D of brain scanners every year. It was only recently that we could see inside the brain with sufficient resolution to see individual interneural connections. The same skepticism was expressed about the genome project. When that project was started, we had only collected one ten-thousandth of the genome. Halfway through the project, skeptics said, “I told you this wasn’t going to work. And here you are, seven and a half years into a 15 year project and you’ve finished 1% of the project.” But you double 1% seven times and you get to 100%. That’s exactly what happened.

We exactly doubled every year the amount of genetic data we sequence. It took us 15 years to sequence HIV. We sequenced SARS in 31 days. There are many examples. If you can measure something in information, it inherently grows exponentially. But our intuition is linear. So lots of people think intuitively, and the kind of projections you get for an exponential trend just seem incredible. That is really how the technology, in every different manifestation of information technology, has progressed.

The paradigm shift rate is doubling every decade. The telephone took fifty years to be adopted. That was the first virtual reality technology. Before that, you really had to be in the same place to hold a conversation with somebody. These are logarithmic graphs, so a straight line on a logarithmic graph is exponential growth. These early communication technologies took decades to be adopted. More recent ones took a few years time. This acceleration has continued. The word “blogs” wasn’t used three years ago. People did not use search engines five or six years ago. Social networks are new. Three years ago people said you can’t make money in internet advertising. Now you have one company started here on this campus that is worth $100 billion and 99% of its revenue comes from internet advertising.

The pace of change is accelerating. Evolution creates a capability and then uses that capability to advance the next stage. That is why an evolutionary process accelerates. The first paradigm shift on this double logarithmic graph took billions of years, then evolution used DNA, an information backbone, to guide the next stage. The Cambrian explosion went a hundred times faster and biological evolution kept accelerating. Homo sapiens took only a few hundred thousand years to evolve.

There are only three simple genetic changes comprising tens of thousands of bytes of information that really distinguish us from other primates. A larger skull at the expense of a weaker jaw, so don’t get into a biting contest with another primate. More of the brain is devoted to the cerebral cortex, so we can do abstract reasoning. There are things like mirror neurons and particular neural structures that can deal with recursion, which are also important in humans’ ability to create hierarchical symbolic structures. A chimpanzee’s thumb actually work, but it is an inch down. They don’t have a power grip or fine motor coordination and cannot manipulate the environment.

These tens of thousands of bytes of genetic information comprising these changes are really the enabling factor that brought on the next evolutionary process, which was human-directed technological and cultural evolution. That was a little bit faster. It only took tens of thousands of years for stone tools, fire and the wheel. We always use the latest generation of technology to advance the next stage, and so technological evolution continued to accelerate in a straight line from the biological evolution that preceded it. If you look at this on a linear graph, it looks like everything has just happened.

People said Kurzweil only put points on this graph that fit on a straight line. I took fourteen different lists, like Carl Sagan’s Cosmic Calendar, the American Museum of Natural History, and Encyclopedia Britannica. These were not lists trying to make or break my point. This is just what they thought the key events were. You see there is some spreading of the points, but there is a clear trend line. This is a linear graph, and an exponential looks like a linear progression for a few years. You go out for a long period of time and the perspectives diverge quite radically. The further out you go, the more unrealistic the linear perspective becomes. The ongoing exponential is made of a series of S curves, each one representing a specific paradigm.

From personal experience, if we look at a computer that took up about half the size of this room when I came to MIT in ’65, it cost $11 million and was a thousand times less powerful than the computer in your cell phone today. Moore’s law is really one example of many. It was the fifth paradigm to bring exponential growth to computing. The lower left-hand corner, the first data processing equipment to automate an American census. They used these old punch card machines. I think they were subsequently shipped to the Florida election commission. I’m going to leave out the rest of my jokes just to make the time frame here.

Alan Turing created a relay-based computer that cracked the German Enigma code. Vacuum tubes were used in the 1950′s. They were then shrinking vacuum tubes to make them smaller and smaller and to keep the exponential growth going. That paradigm hit a wall, and that was the end of the shrinking of vacuum tubes. It was not the end of the exponential growth of computing. It went to another paradigm, transistors, which are not small tubes. Then it went to integrated circuits. Moore’s law, the shrinking of integrated circuits, was the fifth paradigm to bring exponential growth.

There have been lots of projections that that will hit a wall. The first projections were for 2002. Intel and the ITRS roadmap now say 2022. For that time, the key features on conventional chips using photolithography just like today, with four nanometer features we will be able to do strong AI based on the most conservative estimates for the amount of computation required for about $100. That is with no molecular computing.

But molecular computing is working already. There are early versions due to hit the market. If you speak to Intel scientists, they are very confident of that particular paradigm. So we are not talking about any other exotic technologies. We can speculate what might happen after the Singularity. We can go into things like quantum computing and pico computing, but that is not really at issue here. We can even do it without molecular computing, which was somewhat speculative when we met seven years ago at the Spiritual Machines conference. It’s really very much a mainstream view today among the scientists working in that field.

This is Hans Moravec‘s chart. Same progression. Generally my projections, though they are considered in some quarters radical, end up very often being conservative. My book, which came out half a year ago, projected 2013 to hit 10^13 calculations per second. Japan recently announced two supercomputer projects will hit that level by 2010.

This is the price of a transistor. You could buy a whole transistor for a dollar in 1968. If you backup to when I was hanging around the surplus electronics shops on Canal Street in New York in the early 1960′s, I bought something this big for $40 equivalent to one transistor. It was a telephone relay with support circuitry. You then could buy a whole fast transistor for a dollar in 1968. 10 million in 2002. 100 million today.

But look at how smooth this progression is. It looks like the output of some tabletop experiment, but this is the measure of worldwide activity involving millions of people. You might wonder how could that be. How can you get such smooth and predictable curves from the chaotic activity of lots of people when any specific project is very unpredictable? We see a similar phenomenon in other areas of science, like thermodynamics. The path of one particle in a gas is completely unpredictable, yet the overall gas, made up of a large number of chaotically interacting particles has very predictable properties according to the laws of thermodynamics to a very high degree of precision. If you have a dynamic, chaotic system, where each element is unpredictable, the overall results will have certain predictable properties.

Any process that is truly an evolutionary process is such a chaotic, dynamic system. I say this not just back-fitting past data. I’ve been making these forward-looking predictions for a long time. As transistors have gotten smaller and cheaper, they have also gotten faster. We have been doubling the price performance of electronics about every year. That is fifty percent deflation. That is also true of other types of information technology, like genetic technology and brain data. If you point to a lot of different types of information technology, software has a 50% deflation factor.

Economists said there was no way we would actually double our consumption of information technology every year. Actually, what we find is that we more than double it. We have had 18% growth in constant dollars in information technology for the last 50 years, more than keeping pace with this doubling of price performance. People did not buy iPods for $10,000 ten years ago, which is what it would have cost. As the price performance reaches certain levels, new applications open up.

Magnetic data storage, I just put this up here because this is not Moore’s law. This is not shrinking transistors but shrinking magnetic spots. Different scientists, same progression.

Biotechnology is very revolutionary. We will be able to reprogram our own biology. We have tools now to actually reprogram our genes. We can turn genes off with RNA interference. We can add new genetic information with reliable new techniques that overcome some of the earlier problems. With gene therapy, we can turn on and off enzymes.

It cost $10 per base pair in 1990. It’s a penny today. The amount of genetic data, this slope on this chart represents doubling every year. That has continued past the collection of the genome.

Communication, there are fifty different ways to describe it – wired, wireless, the number of bits, the amount of capacity, bandwidth – all basically doubling every year. If we have enough data we can see the ripples of S curves. I had a little bit of this chart of the internet, called the ARPANET in the 1980′s, and predicted in The Age of Intelligent Machines, which I wrote 20 years ago, that there would be a worldwide communication network comprising tens of millions of people emerging in the mid-1990s.

It seemed absurd at that time, because it was 10,000 nodes doubling to 20,000 and used by a few thousand scientists. If you looked at it on a linear graph, this is what that same data looks like. And then in the early ’90s, people said, “Your projection is only a few years from now and still nothing is happening.” But when you are doubling these little numbers, it ultimately becomes explosive. By the mid 1990′s, the internet did take off.

We are shrinking technology at an exponential rate, according to my models at a factor of 100 per 3D volume per decade both electronic and mechanical.

These are some projections from Eric Drexler‘s seminal book, Nanosystems from 1986, which I just got a signed copy of and it has a prominent place in my office. I give examples in my book of hundreds of working nanosystems.

This is not speculative anymore. One scientist has created a little robot that walks with a convincing gait at the molecular level. Another scientist has built a device of which you put thousands of them in the bloodstream of rats and it cures type 1 diabetes, has 7 nanometer pores, lets out insulin in a controlled fashion. The idea of blood cell-sized devices for performing sophisticated therapeutic functions might sound very futuristic. There are four major conferences on BioMEMS to do exactly that. At MIT and at the University of Rochester there is a tiny nanodevice that can detect the specific antigens on the surface of a type of cancer cell, latch onto the cell, burrow into the cell, detect that it’s inside the cell, release toxins and destroy the cell. The steps of these nanoengineered systems has been demonstrated in vitro and it is going into animal trials. That’s today.

If you then contemplate, multiplying the power of these technologies by a billion in 25 years, which is what this doubling every year amounts to, while we shrink the size of these technologies by a factor of 100,000 in 25 years, you get some idea of what will be feasible by the late 2020′s.

This is an animation of a robotic red blood cell. The red blood cell is a system we have reverse engineered. We understand them; they’re not that complex. It does bring up an issue about biology that it’s very intricate but it’s also very suboptimal. When we can reengineer these systems, we can make them much more capable. A conservative analysis of these robotic respirocytes that Rob Freitas has designed indicates you could do an Olympic sprint in 15 minutes without taking a breath if you use these, or sit at the bottom of your pool for four hours. “Honey, I’m in the pool” will take on a whole new significance.

If you look at what is already happening, there are some very sophisticated nanosystems that are being demonstrated. Really understanding the principles of operation of the human brain has been difficult up until recently. You can’t do this with fMRI, getting fuzzy pictures seeing where there’s mental activity when you do certain types of thoughts. But we are doubling the spacial resolution of brain scanning every year. I describe a number of new scanning technologies in the book. One at the University of Pennsylvania allows you to see for the first time interneural connections signaling to each other in real time in a living brain. We are getting the data to actually begin to understand how the brain creates our thoughts.

My proposal is not just to reverse engineer the human brain and thoughtlessly put those algorithms on a suitable computational substrate. We have an existing AI toolkit. It’s increasingly powerful. We are doing things today that were impossible even seven years ago. We are just beginning to get hints from brain reverse engineering. When we got data on the front end from the auditory cortex, we put that into our speech recognition and got dramatically better performance. It was actually counter-intuitive. These transformations are not what we would have predicted, but after we got them, we saw why those transformations are being made. So we got that set of principles from having reverse engineered the auditory cortex. That was very helpful in this particular narrow AI application.

We can expand the AI toolkit from ongoing AI research. We will have hints, ideas and models coming from reverse engineering the human brain. We will study the performance of the human brain itself. One good laboratory for that is studying human language, which does manifest all the hierarchical symbolic types of structures that human thought is able to deal with. In fact, Turing based his Turing test on human language. Brain scanning is growing exponentially in resolution. The amount of data is doubling every year on the brain. Doug Hofstadter has posed the question, “Are we smart enough to understand our own intelligence?” Are we above or below that threshold? Clearly a giraffe’s brain, Doug points out, is below that threshold. Maybe we are also below that threshold.

Our ability to reverse engineer and model already twenty regions of the brain out of the several hundred regions that exist and the fact that we can create technology I believe puts us over that threshold. This is a detailed mathematical model and computer simulation of twelve different regions of the auditory cortex. Applying sophisticated psychoacoustic tests to the simulation gets very similar results to applying those results to human auditory perception. It does not prove that the model is totally accurate, but it does show that we are capable of understanding the principles of operation, expressing them in the language of mathematics. If we can do that, we can simulate them on suitable computational platforms. The number of regions that is being simulated is expanding. We have a simulation of the cerebellum. Applying skill formation tasks to the simulation gets very similar results to applying those tests to human skill formation. The cerebellum is an important region that comprises more than half the neurons of the brain. It does give us some good insight into what the complexity of the brain is.

There are a number of ways in which the brain is really quite different from computers. People often cite that computers cannot simulate the human brain because it is massively parallel, it uses analog computing, it uses holographic forms of storage, there are very few cycles available to make decisions, it’s self-organizing… but these are actually very good principles which we are beginning to utilize. As we begin to learn more about how these self-organize principles work, we are incorporating them into modern AI programs, particularly pattern recognition. We use these self-organizing paradigms, and as we learn more about how the brain does its own self-organization it gives us powerful insights to apply to those pattern recognition methods.

A key point about the complexity of the brain. One of the objections is the brain is too complex. There are 100 billion neurons, 1000 connections per neuron, 1000 ion channels per connection. It’s all these many nonlinearities. This is a level of complexity that is just way beyond today’s technology. That is the apparent complexity if you look at a mature brain. But the design of the brain is a billion times simpler than that. How do we know that? Well, the design of the brain is in the genome. The genome has the design of the human body and brain. It’s 800 million bytes, but it’s replete with redundancies. One sequence, ALU, is repeated 300,000 times. I show in the book how you can compress using lossless compression the genome down to 30 to 100 million bytes.

So the genome, which describes the design of the human body and brain, has 30 to 100 million bytes, which is less than Microsoft Word. My point is not that it’s a simple system, but it’s a level of complexity that we can handle. How do you get a brain that is much more complex than the design it starts with? Basically, it’s a probabilistic fractal. It expands itself through a lot of stochastic processes.

The cerebellum is a good example. If you look at the cerebellum, you see this vast amount of complexity. How could we ever hope to understand this? You see trillions of these incredibly tangled bundles of connections. It looks like this vast system is beyond our intelligence to understand. We understand it. Because there are actually only tens of thousands of bytes of design information. The genome says the following about the cerebellum: there are four different types of neurons, they are organized kind of like this in one cell, now repeat this ten billion times and add a little bit of random variation within the following constraints with each repetition. That is what the genome says about the cerebellum. We actually understand that. We’ve modeled it mathematically, and that’s behind this simulation of the cerebellum.

An example of a fractal, and Doug has written extensively about these, it looks very complex. How much information is in the Mandelbrot set? You’ve probably seen pictures of it. If you take the image, it’s a vast amount of information. Depending on the resolution, it can just be an indefinitely vast amount of information. The design of this is six bytes long. As it is iteratively applied, a fractal expands its information. The brain is actually a probabilistic fractal. There are a lot of random processes as it expands its information, and it’s self-organizing. This largely randomly wired cerebellum that interacts with a complex environment, and a child learns to walk, talk, and catch a fly ball, it gets filled up with meaningful information.

Models often get simpler at higher levels. Modeling one pancreatic islet cell, it’s very complex. Modeling the whole pancreas is much simpler. There is an artificial pancreas in FDA tests. The same thing is applied to these brain models. You have to apply the abstractions at the right level. Biology is theoretically based on chemistry, which is based on physics, but you don’t explain what a biological system does on the level of quarks or protons. Maybe at proteins.

You have to model things at the right level. Neurobiology has its own levels of abstraction. In fact, there are several levels of abstraction in the brain, with ideas at the highest level. We are beginning to reverse engineer this. We are at the early stages, but this will be an exponential process.

This is all driving economic gains. Productivity is increasing. In fact, it is that increase in the consumption of information technology which is increasingly pervasive in all of our economic institutions that is driving all of the economic growth we are seeing. The adoption of these technologies is exponential.

This is e-commerce, now a trillion dollars. You might say, “Wasn’t there a boom and a bust?” which was experienced in this geographic area quite acutely. That was a Wall Street phenomenon. Wall Street looked at the internet and said, “Wow. This is revolutionary. It’s going to change every business model.” And the values went like this. Three years later they came back and said, “It hasn’t changed all the business models. I guess we were wrong.” And all the values went like that. The actual adoption was exponential. But remember, an exponential looks like nothing is happening until you get some real traction. We are getting real traction now with e-commerce.

Let me show you an example of another technology we put together. We created the first large vocabulary speech recognition that was commercially marketed and the first text synthesis. We put those together with language translation and created a translating telephone. The language translation has actually made a lot of progress. It is another example of how we have made exponential progress in the quality of software using more sophisticated pattern recognition techniques.

I was at Google recently, and they applied it to this opportunity we have now for vast data mining. They took these very large Rosetta Stone texts and created a English to Arabic and Arabic to English translator that not only won the DARPA competition, but actually compared equally to human professional translators (and nobody on the team spoke a word of Arabic) basically, using pattern recognition applied to very large databases.

This is my proposal for how we can do language. Not with symbolic rules, which are very hard to do. When we did speech recognition, we knew there were 44 phonemes, we knew there were various phonetic rules and so on, but we didn’t tell the system what they were. We let it discover it by itself from vast databases using pattern recognition. This will be a routine feature of your cell phone early in the next decade.

This is an image I had in The Age of Spiritual Machines. In The Singularity is Near, I put dotted lines around those capabilities of computers that were soon to fall off. At the Gilder Conference I was criticized for “only humans can drive cars.” That was going to stay on the wall, but I did actually feel that that would come down.

Another progression that is of interest, we are already extending beyond our horizons. This is not a new phenomenon. If it was not for technology, half the audience would not be here. The other half would be senior citizens. What I principally tried to do in The Singularity is Near is deal with the objections that have come up to The Age of Spiritual Machines. One criticism is just from incredulity. The results of exponential growth seem just too incredible. That was true twenty years ago and is equally true today. I do think people think linearly. That is our intuitive view, but the intuitive view is not historically correct.

People say exponential trends cannot go on forever. Information technology transcends one paradigm to another, but even information technology has a limit. There are ultimate limits. I talk about what they are.

Information technology will hit a wall, but not before it reaches pretty extraordinary levels. Just based on the technologies we can touch and feel today that are already working, without going into any speculative realms of quantum computing or other types of computing ( you can even ignore molecular three-dimensional computing, which is working), conventional chips will achieve strong, cheap AI in the 2020s.

People say software AI is stuck in the mud. Computers still cannot tell the difference between a dog and a cat, for example. But we are making steady progress in AI. There are hundreds of applications, as I mentioned earlier, deeply embedded in our economic infrastructure. This was not true seven years ago when we had this last conference. These were all research projects at that time. People tend to discount each new accomplishment of AI. In the book I deal with this perspective from many different vantage points: software complexity on log scales is progressing exponentially. Chess was mentioned earlier. But the latest systems, their performance is not accounted for by the brute force increase in power. They actually have better software, better pattern recognition.

Genetic algorithms are progressing. We use them a lot. The early genetic algorithms had fixed genomes, unlike real biological evolution where it can add genetic information, reassign the meaning of genes, have other non-coding genes that control the expression of coding genes. We’ve added these types of innovations and are getting much more dramatic results. The proposal is not to have one big GA create strong AI. It is one self-organizing paradigm among many that we can apply to this problem. There are many different ways of looking at this. I’m on the Army Science Advisory Board, and the sophistication of the autonomous systems that are now being developed are far more complex and capable in software, not just hardware, than they were five years ago.

The criticism from reliability: software is just too brittle, too crash-prone. We can and do create reliable system. A majority of airplane landings are controlled by software. The number of times a software bug has caused a crash is zero. The same is not true for human landings. If we create technology that is decentralized and self-organizing, it is inherently very stable. The quintessential example is the internet. The amount of time that the internet has been down over the last decade is zero seconds.

People say the brain is too complex. I address this. Thomas Ray said it would take billions of lines of code to describe the brain. That is looking at this apparent complexity, not the complexity of the design, which is contained in the genome. It’s not a simple system, and we are not there, but the amount of complexity is a level that we can manage.

Let me say a few words about promise versus peril. The ethical guidelines do work for inadvertent guidelines. The Asilomar guidelines in biotech have kept us safe for thrity years. The Foresight Institute, which Christine heads up, has similar guidelines for nanotech. I believe they will work effectively. The big problem is advertent problems – that is actually a word. Designer pathogens, self-replicating nanotech, and Unfriendly AI, which Eliezer has written extensively and insightfully about, are really the key concerns. There is a movement to relinquish these technologies because they are just too dangerous. There are three problems with that. It would require a totalitarian system to implement, it would deprive us of profound benefits, and it wouldn’t work. It would just drive these technologies underground where they would be less stable. The responsible scientists would not have access to the tools needed to defend us.

The way we can defend ourselves is through narrow relinquishment of dangerous information and to invest explicitly in the defenses. This is something we have done with software viruses and which has worked actually quite well. Even though Bill Joy and I are often considered on opposite sides of this, we have actually worked together. Bill Joy got his ideas about the downsides from The Age of Spiritual Machines. We recently wrote this op-ed piece in the New York Times criticizing the dissemination of dangerous information, specifically the 1918 flu genome on the web, which we thought should be given to the scientists who needed it in confidence. And also calling for a Manhattan project to combat biological viruses. Majority leader Bill Frist recently cited this op-ed and endorsed both of those proposals.

Bill McKibben will speak later. He writes, “Is it possible that our technological reach is very nearly sufficient now that our lives, at least in the West, are sufficiently comfortable?” My view is no, not until we can meet our energy needs (which we will be able to do with nano-engineered solar panels and fuel cells, which can capture the only 1% of 1% of sunlight that falls on the earth to meet 100% of our energy needs in a clean, renewable fashion), overcome disease and death, overcome poverty, and a few other problems. Only technology – advanced, nanoscale, distributed, decentralized, self-organizing, increasingly intelligent technology – has the scale to solve these problems that the human species has struggled with for eons.

Let’s say overcoming disease and disabilities, those are good things, but we should not go beyond normal human capabilities. As I say, most of the audience would not be here if we had not already gone beyond normal human abilities. I really have two responses to that. One is, what is “normal”? I mean, the guy who was whistling some music in the lobby of my hotel this morning, is that “normal” musical ability? I hope not. Or Beethoven and the Beatles, is that normal musical ability? There is a wide diversity of abilities. Going beyond normal is not a new story. We are the species that seeks to go beyond our limitations.

People say “Death gives meaning to life.” But we really get meaning from what is unique about the human species. We create knowledge. That knowledge base is expanding exponentially. I’ve been at lots of conferences with scientists who are very fond of saying “We’re not unique. We used to think we were. The universe doesn’t revolve around us. We’re not descended from the Gods. We’re descended from primates and worms.” But we are unique, after all. Those few tens of thousands of bytes of genetic information, which were the enabling factors for us to create technology has enabled us to create an expanding knowledge base and expand our horizons.

In summary, I believe the takeoff is soft. I am a conservative in some quarters. Exponential growth, which will continue based on known physics of information technology, known paradigms and known technologies that are working, without any speculative technologies, is gradual, incremental and smooth. There is no silver bullet, but ultimately it will be profoundly transformative. Thank you very much.