Infinite Memory and Bandwidth: Implications for Artificial Intelligence
February 21, 2001 by Raj Reddy
Not to worry about superintelligent machines taking over, says AI pioneer Dr. Raj Reddy. A more likely scenario: people who can think and act 1000 times faster, using personal intelligent agents.
Herbert A. Simon University Professor of Computer Science and Robotics, Carnegie Mellon University
Originally presented as a talk at the Newell-Simon Hall Dedication Symposium, October 19, 2000. Published on KurzweilAI.net February 22, 2001.
It is a pleasure for me to speak at this symposium honoring Allan Newell and Herb Simon. I came to CMU in 1969, mainly because this was the home of Newell and Simon. It has been an amazing 30 years. The next 50 years promise to be even more exciting, as the topics at this symposium indicate.
The main thesis of my talk is that none of the dire consequences of Bill Joy or the predictions of Kurzweil and Moravec about the possible emergence of a robot nation will come to pass. Not because they are incorrect, but because we live in a society in which progress depends on the investment of research dollars. In 1978, John McCarthy said, “human level AI might require 1.7 Einsteins, 2 Maxwells, 5 Faradays and .3 Manhattan Projects, the project coming after the conceptual breakthroughs”. Not only do we do not have the long term infusion of billions of dollars of funding to realize human level AI, we not have all the conceptual breakthroughs we need either.
Many of the early predictions in AI never had a chance of being realized because the pundits of research policy did not see the benefit to society of investing in tasks such as chess playing or theorem proving! I do not see that the society ready to invest billions of dollars needed to create a robot that exceeds human capabilities. Policy makers in Congress seem to be more comfortable with investing in healthcare research than in information technology. For example, this years increment to the NIH research budget is 2.3 billion dollars while the increment to IT research budget is only 125 million, in spite of the fact that 1/3 of all the jobs created in the last 50 years are in IT related fields and over 25% of all the taxes come from IT businesses, workers, and the capital gains taxes paid by the Dotcom millionaires! The other source of funds, namely, the venture capitalists are no help either. They seem to want to invest primarily in IT research with near term market potential.
A more likely scenario may be the emergence of “artificial intelligences with super human capabilities.” These will emerge in the guise of intelligent agents and productivity tools, providing some individuals with capabilities that will appear to be extraordinary to the rest of us. The consequences of such technologies are worthy of introspection, examination and debate.
Let us look at the technology trends that are fueling this debate. This year as expected, we are beginning to see the arrival of a giga-PC which delivers a billion operations per second, a billion bits of memory and a billion bits per second network bandwidth, all available for less than two thousand dollars. Barring the creation of a cartel or some unforeseen technological barrier, we should see a tera-PC by the year 2015 and a peta-PC by the year 2030.
The question is, what will we do with all this power? How will it affect the way we live and work? Many things will hardly change; our social systems, the food we eat, the clothes we wear and the mating rituals will hardly be affected. Others, such as the way we learn, the way we work, the way we interact with each other and the quality and delivery of health care will undergo profound changes. Some of the computing power will be used to create self healing computers and networks that never fail and self healing software that never needs rebooting
Advances in magnetic disk memory have been even more dramatic. Disk densities have been doubling every twelve months, leading to a thousand-fold improvement every ten years. Today, you can buy a thirty-gigabyte disk memory for about a hundred dollars. Thirty gigabytes can be used to store about 10 hours of video, 100 paintings, 1000 hours of MP3 music and 10000 books–larger than most of our personal collections at home. By the year 2010, we should be able to buy “30 terabytes” for about the same price. At that cost, each of us can have a personal library of several million books, a lifetime collection of music and movies–all on our home PC. What we don’t have on our PC will be available at the click of the mouse from the universal digital library containing all the authored works of the human race.
By 2020 you will be able to buy “30 petabytes” for a 100 dollars, essentially an infinite amount of memory for all practical purposes. What can you do with a petabyte? If you choose to, you will be able to capture everything you ever said from the time you are born to the time you die. That would take less than one percent of a petabyte! Everything you ever did and experienced can be captured in living color with only a few petabytes.
Most dramatic of all recent technological advances is the doubling of bandwidth every 8 months, propelled by the advances in fiber optic technology. Today you can buy commercial systems that permit transmission of 1.6 terabits per second on a single fiber using dense wavelength division multiplexing (DWDM) technology. This technology uses 160 different wavelengths each capable of transmitting 10 gigabits per second. Experimental systems are able to transmit as much as 25 terabits per second on a single fiber today!
What can you do with 1.6 terabits per second bandwidth? In one second, you can transmit 10 HDTV movies, or 40 regular full-length feature films, or 20000 hours of MP3 music on single fiber. It would take about 50 seconds to transmit all the books in the Library of Congress. All the phone calls in the world can be carried on single fiber with room to spare. The main bottleneck today is not the bandwidth but rather the speed of computers capable of accepting and switching data packets arriving at terabit data rates! Only recently have there been routers that can accept 10gbps data rates (from start ups like Juniper Networks which was founded by a CMU alumnus)! The speed of light is also proving to be a problem. The maximum sustainable bandwidth using tcp/ip protocols is governed by the round-trip delay times! At terabit rates, with round trip times of about 30 ms across the US, 30 billion bits would have been transmitted before an acknowledgment can be received!
It is expected that the exponential doubling of memory and bandwidth will continue for the next 10 to 20 years, leading to the availability of petabyte disks and petabytes per second bandwidth at a cost of pennies per day! This qualitatively changes the way we think of computation and algorithm design, because often we can compensate for scarcity of one resource by using another. This is nothing new. Ever since the 1950s, computer scientists have been using such compensation techniques. For example, to minimize computation needed to calculate sine and cosine values, pre-computed tables were used. To minimize memory usage, iterative and recursive algorithms were developed. To minimize bandwidth, registers and cacheing architectures were conceived.
These concepts are equally valid today and can be stated broadly as the following laws of computation:
- Memory can compensate for lack of computation
- Bandwidth can compensate for lack of memory, and
- Computation can compensate for lack of bandwidth and vice-versa!
As we head toward computational resources, which are “for-all-practical-purposes-infinite”, we can expect that, permitting exchange of one scarce resource by another leading to revolutionary consequences.
Perspectives on the Future of AI
In the rest of this talk I will try to derive consequences of this phenomenon for the future of artificial intelligence. In particular, I would like to reexamine the concepts such as Teleportation, Time Travel, and Immortality and explore alternative realities than those proposed by science fiction writers.
First, lets examine the prospects for Teleportation. With increased bandwidth and computational capabilities, it will become possible to perform 3-D visualization, remote control of micro-robotic surgery and other sophisticated procedures. It’s not quite teleportation in the classical sense of StarTrek, but consider the following: If you can watch the Super Bowl from the vantage point of a quarterback in the midfield, or repair a robot that has fallen down on the surface of Mars or perform tele-surgery three thousand miles away, then you have the functional equivalent of teleportation–bringing the world to us, and bringing us to the world, atoms to bits.
This form of teleportation promises to become a reality given the recent advances in 3-D modeling and multi-baseline-stereo theory developed by Professor Takeo Kanade and other researchers at Carnegie Mellon University. The virtualized reality studio dome is fully covered by many cameras from all directions. The range or depth of every point in an image is computed using a multi-baseline-stereo algorithm. The scene can be reconstructed with the depth and intensity of information by placing a virtual, or soft camera from the front, from the left, from the right or from the top, or moving the soft camera as the user moves freely. Currently this system requires about a teraflop per second for the 3-D reconstruction of the basketball scene at the video rate. Instrumenting a football field with a dome consisting of ten thousand high-definition cameras will require twenty petaflops of computation and a terabit per second bandwidth to transmit the 3D Model.
Toward Time Travel
This brings us to the prospect of using time travel. In the future, it will no longer be necessary or essential for the teacher and the student to be at the same place at the same time. For example, if we had captured Einstein in living color and 3-D when he was alive, it would be technically possible today to have an imaginary conversation with him. Scott Stevens and Don Marinelli have indeed created such a synthetic interview with Einstein. They are also hoping to create a service that will permit you to converse with your great, great, great grandchildren in the same way. This is not quite the time travel that you’ve grown to expect from Star Trek, but it’s another example of substituting bits for atoms to achieve an equivalent experience. Actually one of the episodes of the “Star Trek: The Next Generation” featured a “holodeck” where people can converse with Einstein or anyone else in the same way we are describing.
With some pre-planning and appropriate data capture, future generations will be able to experience virtual time travel into the past and interact with the past generations. What about time travel into the future? One could invent alternative futures by developing simulated environments using virtual world technologies such as the ones pioneered by Randy Pausch in his popular CMU course. While it would not be time travel into the real future, it would be an instructive planning and training tool!
What are the prospects for immortality? Hans Moravec outlines a vision for immortality in his first book “Mind Children” which involves scanning the brain at high enough resolution to create a faithful silicon model that can then have an independent existence outside of the biological limitations of human tissue.
I would like to propose an alternative future. Immortality should not be thought of as some mystical transfer of atoms from one brain to the other as in the Star Trek movies. It could be viewed from an information-technology perspective whereby you provide your clone with all the important extragenetic experiences of everything you ever said and did.
There is work underway in areas such as geriatric robotics that will help senior citizens with simple disabilities lead normal lives well past their prime. And you may ask, can this go on forever? Transplant surgeries are one way of extending life expectancy beyond a hundred years or so, and given advances in cloning, we may be getting closer to achieving the dream of immortality. But as Nathan Myhrvold points out, you need to download extragenetic experiences–the software in your brain, not just the DNA-based system.
One possibility would be to bring you back to life in the fourth millennium using a frozen embryo of your clone and then infusing the clone with all the experiences you’ve undergone in this lifetime. You create a rapid, simulated learning environment in which the new clone, with a new brain gets all of your experiences, and can live on for another generation–bits to atoms! This process does not lead to immortality in the classical sense, but close enough, especially given that the cloning process can go on every millennium. That way you will live forever, except you will be learning the cumulative experiences of all the generations.
Emergence of “Artificial Intelligences with Super-Human Capabilities”
Lastly, I would like to discuss the possible emergence of a super human race. First, what do I mean by “Artificial Intelligences with super human capabilities”? I basically mean people who are able to think and act a 1000 time faster than other mere mortals, but without requiring any special carbon or silicon-based enhancements to their genetic makeup. They would achieve this super human capability thru the use of thousands of personal intelligent agents both on their body and in the cyberspace. These agents will exhibit and realize intelligent behavior mostly thru recognition rather than by recall, enabled by availability of infinite memory and bandwidth.
Intelligent behavior based on recognition rather than “recall and reasoning” is also the hallmark of human intelligence. Two examples of the power of recognition come to mind. First is the number 1729, made famous by the mathematical prodigy Srinvasa Ramanujan. Responding to a casual comment by Professor Hardy, who was visiting him in the hospital, who said that the taxi he came in had an uninteresting number 1729, Ramanujan is said to have responded by saying: “On the contrary Professor Hardy, it is a very interesting number: It is the smallest number which can be expressed as the sum of two cubes in two different ways”! 1729 happens to be 10^3 plus 9^3 and also 12^3 plus 1^3! It is also the smallest such number. No wonder that it was said, that every number was a friend of Ramanujan! He had obviously discovered this relationship at some earlier time and filed it away in his memory. This anecdote provides us with an interesting example how superhuman capabilities can be demonstrated by utilizing recognition memory instead of complex reasoning processes. Another example of the power of recognition memory is the famous experiment of Chase and Simon in reconstructing chessboard positions. Chess masters were able to reconstruct most of the chessboard by recognizing the classical patterns whereas amateurs failed miserably given the same task!
At present, human intelligence is limited by the knowledge that can be acquired and mastered in one’s lifetime. We have no mechanisms for instantaneously tapping into the collective wisdom of the human race as in the case of Borgs in StarTrek! Infinite bandwidth coupled with infinite memory promises to create such a capability. Individual human capability will be augmented in two ways. First, given a problem, thousands of intelligent agents can search and harness the personal knowledge contained in the terabytes of memory on the body and petabytes of memory on the desktop. Second, agents can query and retrieve relevant knowledge accumulated by other members of the human race (and their agents), who may choose to share their knowledge for free or for a fee. Robust and scalable mechanisms for peer-to-peer information sharing are beginning to emerge in systems such as Napster and Gnutella. Thus the future capabilities of an individual will not merely be dependant upon what he or she knows, but what is known to the agents and what is knowable by the agents by a rapid scan of the collective knowledge of the connected-humans. This is somewhat similar to a scene in the movie Matrix where one of the characters “learns” to fly a helicopter in a few seconds. Biological limitations of the human would make it impossible to learn to fly in few seconds. But the human and the personal agents together would have no such limitation!
Some of us will have superhuman capabilities, like getting a month’s worth of work done in a day, by harnessing and utilizing the power of thousands of intelligent agents. This super human race will not have horns or look like a robot race, but rather just like any of us. However, as in the movie “Gods must be crazy” where the coke bottle dropped from a plane becomes “a gift from Gods” and an object of worship to the natives of Kalahari Desert, actions of these super humans will look magical to the rest of us, not unlike Ramanujan’s feat with the number 1729. They will derive their power through the use of infinite personal recognition memory and access to the collective knowledge of the connected-humans by exploiting the power of infinite bandwidth.
Should we be afraid of the possible emergence of super human race? On planet Earth, we see millions of species coexisting with each other! I expect the same will happen with this new more powerful version of us! They will become an virtual nation of the techno-elite who will mainly interact with each other and to a large extent coexist peacefully with the other species on the planet. On the rare occasions when there is a conflict, as it happened when the Native Americans confronted the Settlers, it will be an uneven contest! We can see glimpses of this future in the sub-culture of Silicon Valley!