Bravo this is great. I like the part about Marx's labor theory of value. AGI economics - isn't that a world that lends itself to making a compelling financial pitch.
John From: Edward W. Porter [mailto:[EMAIL PROTECTED] Robin, I am an evangelist for the fact that the time for powerful AI could be here very rapidly if there were reasonable funding for the right people. There is a small, but increasing number of people who pretty much understand how to build artificial brains as powerful as that of humans, not 100% but probably at least 90% at an architectual level. What is needed is funding. It will come, but exactly how fast, and to which people, is the big question. The below paper is written with the assumption that someone -- some VC's, Governments, Google, Microsoft, Intel, some Chinese multi-billionaire -- makes a significant investment in the right people. I have cobbled this together rapidly from some similar prior writings, so please forgive the typos. I assume you will only pick through it for ideas, so exact language is not important. I you have any questions, please call or email me. Ed Porter ================================================================== The Time for Powerful General AI is Rapidly Approaching by Edward Porter The time for powerful general AI is rapidly approaching. Its beginnings could be here in two to ten years if the right people got the right funding. Starting in two years it could start providing the first in a series of ever-more-powerful, ever-more-valuable, market-dominating products. In five to ten years it could be delivering true superhuman intelligence. In that time frame, for example, this would enable software running on less than $3 million dollar hardware to write reliable code faster than a thousand human programmers - or, with a memory swap, to remember every word, every concept, every stated rational in a world-class law library and to reason from that knowledge hundreds to millions of times faster than a human lawyer, depending on the exact nature of the reasoning task. You should be skeptical. The AI field has been littered with false claims before. But for each of history's long-sought, but long-delayed, technical breakthroughs, there has always come a time when it finally happened. There is strong reason to believe that for powerful machine intelligence that time is now. What is the evidence? It has two major threads. The first is that for the first time in history we have hardware with the computational power to support near-human intelligence, and in five to seven years the cost of hardware powerful enough to support superhuman intelligence could be as low as $200,000 to $3,000,000, meaning that virtually every medium to mid-size organization will want many of them. The second is that, due to advances in brain science and in AI, itself, there are starting to be people, like those at Novamente LLC, who have developed reasonable and detailed architectures for how to use such powerful hardware efficiently to create near- or super-human intelligence. THE HARDWARE To do computation of the general type and sophistication of the human brain, you need something within at least several orders of magnitude of the capacity of the human brain, itself, in each of three dimensions: representational, computational, and intercommunication capacity. You can't have the common sense, intuition, and context appropriateness of a human mind unless you can represent and rapidly make generalizations from and inference between substantially all parts of world knowledge - where "world knowledge" is the name given to the extremely large body of experientially derived knowledge most humans have. Most past AI work has been done on machines that have less than one one millionth the capacity in one or more of these three dimensions. This is like trying to do what the human brain does with a brain roughly 2000 times smaller than that of a rat. No wonder most prior attempts at human-level AI have had so many false promises and failures. No wonder the correct, large-hardware approaches have been up until very recently impossible to properly demonstrate and, thus, get funding for. And, thus, no wonder, the AI establishment does not understand such correct approaches. But human-level hardware is coming soon. Systems are already available for under ten million dollars (with roughly 4.5K 2Ghz 4 core processors, 168 TeraFlops/sec, a nominal bandwidth of 4TBytes/sec, and massive hard disk storage) that are very roughly human level in two out of the above three dimensions. These machines are very roughly 1000 times slower than humans with regard to messaging interconnect, but they are also hundreds of millions of times faster than humans for many of the tasks at which machines already out perform us. Even machines with much less hardware could provide marketable powerful intelligences. AIs that were substantially sub-human at some tasks could combine that sub-human intelligence with the skill at which computers greatly out perform us to produce combined intelligences that could be extremely valuable for many tasks. Furthermore, not only is Moore's Law expected to keep on going for at least several more generations, but, perhaps even more importantly, there has been a growing trend toward more AI-capable hardware architectures. This is indicated by the trend toward putting more and more processor cores, all with high speed interconnect, on a chip. Intel and IBM both have R&D chips with 80 or so such mesh-networked processors. There are also plans to provide high bandwidth memory connections to each such networked processor with the memory being placed on multiple semiconductor layers above or below the processors and with a huge numbers of data transferring vias connected between layers. This will substantially break the Von Neumann bottleneck, a well known hardware limitation that has greatly restricted the usefulness of traditional computers for many tasks involving large amount of complexly interconnected data, such those involved in computing from world knowledge. With the highly redundant designs made possible by such grids of tiled networked processors, wafer scale and multiple level wafer scale manufacturing techniques (or equivalents of them provided by Sun Microsystems' capacitive-coupling interconnects) become extremely practical and can greatly decrease the cost of manufacturing massive amounts of memory and processing power all connected by very high internal bandwidths. When you combine this with the rapid increases in the bandwidth of optical interconnect being made by companies such as Luxera, it becomes possible to extend this extremely high bandwidth in a third dimension, making it possible to create computers not only with much more memory and computational power than the human brain, but also much greater interconnect. In fact, if the ITRS roadmap projections continue to be met through to the 22nm node as expected, and if hardware were specifically designed to support general purpose AI, it is highly likely roughly brain-level AI hardware could -- if the Intels and Samsungs of the world focused on it -- be sold with a mark-up over marginal cost of 80% for between $200,000 to $3,000,000 dollars in just five to seven years. As one of the former head's of DARPA's AI funding said, "The hardware being there is a given, it's the software that is needed." SOFTWARE ARCHITECTURES Tremendous advances have been made in Artificial Intelligence in the recent past, in part due to the ever increasing rate of progress in brain science and the increasing power of the computers that brain scientists and AI researchers have to experiment with. For example, the paper "Learning a Dictionary of Shape-Components in Visual Cortex:...", by Thomas Serre of Tomasa Poggio's group at MIT, provides a some-what limited, but still amazingly powerful simulation of human visual perception ( <http://cbcl.mit.edu/projects/cbcl/publications/ps/MIT-CSAIL-TR-2006-028.pdf > http://cbcl.mit.edu/projects/cbcl/publications/ps/MIT-CSAIL-TR-2006-028.pdf ). It gives, just one, of many possible examples of how much our understanding the brain and its functions has grown. It learns and uses patterns in a generalizational and compositional hierarchy, that allows for efficient reuse of representational components and matching computation, and which allows a system to learn in a compositional increments. Similar amazing advances are being made in understanding other brain system, including those that control and coordinate the behavior of multiple areas in the brain, enough so that, that for the first time, we really have enough understanding from which to design artificial minds. The most impressive current brain architecture of which I am aware, is the Novamente architecture from Novamente LLC, a start-up headed by Ben Goertzel, the former CTO of the $20 Million startup, IntelliGenesis, that showed great promise in the dot.com boom until its financial plug was pulled, with less than a day's notice, during the dot.com crash. There may be other impressive brain architectures, but since I don't know of them, let me give a brief -- but hopefully revealing - description of the Novamente architecture as a good example of the state of the art, since it has a rather detailed blueprint for how to build powerful, even superhuman, artificial minds. Novamente starts with a focus on "General Intelligence", which it defines as "the ability to achieve complex goals in complex environments." It is focused on automatic, interactive learning, experiential grounding, self understanding, and both conscious (focus-of-attention) and unconscious (currently less attended) thought. It records experience, finds repeated patterns in it, makes generalizations and compositions out of such patterns -- all through multiple generalizational and compositional levels -- based on spatial, temporal, and learned-pattern-derived relationships. It uses a novel form of inference, firmly grounded in Bayesian mathematics, for deriving inferences from many millions of activated patterns at once. This provides probabilistic reasoning much more powerful and flexible than any prior Bayesian techniques. Patterns -- which can include behaviors (including mental behaviors) -- are formed, modified, generalized, and deleted all the time. They have to complete for their computational resources and continued existence. This results in a self-organizing network of similarity, generalizational, and compositional patterns and relationships, that all must continue to prove their worth in a survival-of-the-fittest, goal-oriented, experiential-knowledge ecology. Re-enforcement learning is used to weight patterns, both for general long term and context specific importance, based on the direct or indirect roles they have played in achieving the system's goals in the past. These indications of importance -- along with a deep memory for past similar experiences, goals, contexts, and similar inferencing and learning patterns -- significantly narrow and focus attention, avoiding the pitfalls of combinatorial explosion, and resulting in context-appropriate decision making. Genetic learning algorithms, made more efficient by the system's experience and probabilistic inferencing, give the system the ability to learn new behaviors, classifiers, and creative ideas. Taken together all these features, and many more, will allow the system to automatically learn, reason, plan, imagine, and create with a sophistication and power never before possible -- not even for the most intelligent of humans. Of course, it will take some time for the first such systems to learn the important aspects of world knowledge. Most of the valuable patterns in the minds of such machines will come from machine learning and not human programming. Such learning can be greatly speed if such machines are taught, at least partially, the way human children are. But once knowledge has been learned by such systems much of it can be quickly replicated into, or shared between, other machines. So will it work? The answer is yes - because that is substantially how the human brain works. It is hard to overstate the economic value and transformative power of such machines. That $3 million dollar system that I described at the start of this paper that could do the work of thousands of programmers or lawyers, that could be rented on the web for under $400 dollars an hour. And if nano-electronics delivers on its promise within twenty-five years such a machine might not cost any more than a PC, and so the work of people like programmers, lawyers, doctors, teachers, psychologists, and investment counselors, could all be largly replaced for roughly a penny an hour. All of a sudden the price of human labor, even in places like China, India, and Haiti becomes uncompetitive for most current employment. Marx's labor theory of value gets thrown on its ass, and is almost entirely replaced by the machine theory of value. Commercially, politically, and geo-politically it's a whole new ball game. It could go either way. It could greatly enlighten and improve human existence, or it could greatly darken it, or it could do some of both. One of the biggest challenges is making our social and political institutions intelligent enough to deal with it well We are truly taking about a "singularity," a technology so powerful that -- when combined with the massive acceleration it will cause in the web, in electronics, robotics, and nano- and biotechnology -- will warp the very fabric of human economy and society in somewhat the same way the singularity of a black hole warps the fabric of space-time. ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244&id_secret=63872877-8353e5
