Tuesday, April 27, 2010

HUMAN-LEVEL ARTIFICIAL INTELLIGENCE --- AND ITS CONSEQUENCES --- ARE NEAR: Why AI will be created in roughly a decade & what that means

THE TIME FOR POWERFUL ARTIFICIAL INTELLIGENCE IS RAPIDLY APPROACHING
====================================================================

There is a good chance human-level AI will be created within five to fifteen years --- and, almost certainly, within twenty-five.

In ten years, for example, a machine costing one million dollars may well be able to: --- write reliable, complex code faster than a hundred human programmers --- remember every word and concept in a world-class law library, and reason from them hundreds of times faster than a human lawyer --- or ---- contribute more rapidly to the advancement of mathematical physics than all of humanity combined.

A cloud of such systems could represent all the knowledge recorded in books and on the web --- stored in a highly indexed, inter-mapped, semantic deep structure that would allow extremely rapid reasoning from it. Such a cloud would have the power to rapidly search, match, infer, synthesize, and create --- using that world of data --- so as to provide humanity with a font of knowledge, reasoning, technology, and creativity few can currently imagine.

YOU SHOULD BE SKEPTICAL. The AI field has been littered with false promises. But for each of history's long-sought, but long-delayed technical breakthroughs, there has always come a time when that breakthrough --- finally --- DID happen. There are strong reasons to believe that --- for powerful machine intelligence --- that time is fast approaching.

What is the evidence?

It has two major threads.

First, within five to ten years, we are projected, for the first time, to have hardware with the computational power to roughly support human-level intelligence. Within that time, the price for such hardware could be as low as three million dollars, down, by the end, to, perhaps, as little as one hundred thousand. These prices are low enough that virtually every medium to large-size business, educational, and governmental organization would be able to afford them.

Second, due to advances in brain science and in AI, itself, there are starting to be people who have developed reasonable and relatively detailed architectures for how to use such powerful hardware to create near-human and, ultimately, super-human artificial intelligence.



THE HARDWARE
=============

To do computations of the type at which we humans currently out perform computers, you need something within at least several orders of magnitude of the capacity of the human brain, itself. You need such capacity in each of at least four dimensions. These include representational capacity, computational capacity, processor-to-memory bandwidth, and processor-to-processor bandwidth. You can't have the common sense, intuition, natural language capabilities, and context appropriateness of human thought --- unless you can represent, rapidly search, infer between, and make generalizations from, vast portions of human-level world knowledge --- where --- "world knowledge" is the name given to the extremely large body of experientially derived visual, auditory, olfactory, tactile, kinesthetic, emotional, linguistic, semantic, goal-oriented, and behavioral knowledge that most humans have.

Most past AI work has been done on machines that have less than one one millionth the capacity of the human brain in one or more of these four dimensions. That is like trying to do what the human brain does with a brain the size of a spider’s. Even many current supercomputers, that cost tens of millions of dollars, have processor-to-processor bandwidths that are three or more orders of magnitude smaller than that of the human brain.

No wonder so many prior attempts at human-level AI have hit a brick wall. Nor is it any surprise that most of the AI establishment does not understand the importance of the correct ---roughly brain-level-hardware --- approach to AI. Such an approach has been impossible to properly study, and experiment with, at prior hardware costs and funding levels --- and, thus, has been impossible to use for advancing one’s career in the AI field, or for raising venture capital.

But starting in three to five years it should be possible to make hardware that is much more suited for roughly human-like computing.

Moore’s Law is likely to keep going for some time. 22nm node prototypes have already been built. Intel claims it is confident it can take CMOS two generations further, to the 11nm node, by mid to late this decade. But, perhaps even more important, there has been a growing trend toward more AI-capable hardware architectures, and, in particular, toward addressing the bandwidth deficiencies of current computing systems.

This is indicated by the trend toward putting more processor cores, with high speed interconnect, on a chip. Tilera has recently demonstrated a 100 core processor with extremely high internal bandwidth. IBM and Intel both have R&D chips with roughly 64 to 80 mesh-networked processors, and they both plan to provide high bandwidth connections between such processors and memory placed on multiple semiconductor layers above or below them. High bandwidth to such memory will be provided by massive numbers of through-silicon metal vias connected between layers. Intel has said it hoped to have such multi-core, multi-layer modules on the market by 2012. And one of its researchers has said inferencing is one of the major tasks that could make such hardware commercially valuable.

Photonics will enable hundreds of gigabits per second to be communicated on photolithographically produced waveguides at relatively low energy and thermal costs. This, and the through-silicon vias, will substantially break the processor-to-RAM and processor-to-processor bandwidth bottlenecks that are currently the major barriers preventing current clustered systems from being used efficiently for human-like reasoning. These bottlenecks need to be broken because many types of human-like reasoning involve --- massively parallel, highly-non-local, out-of-order, memory accessing --- in huge, sparsely-interconnected, networks of world knowledge. With the rapid advances in integrated photonics --- and in low-cost interconnect between such integrated photonics and optical fibers --- being made by organizations like HP, IBM, Luxtera, and Cornell University, it will become possible to extend massive numbers of extremely high bandwidth optical links across chips, wafers, boards, and multi-board systems --- enabling us to create computers --- and clouds of computers --- having not only more effective representational and computational power than the human brain, but also greater processor-to-memory and processor-to-processor interconnect.

With the highly redundant designs made possible by tiled processors, and their associated memory and network hardware --- wafer-scale and multi-level wafer-scale manufacturing techniques can become practical. Such highly uniform, replicated designs make it easier to provide fault-tolerance and self-test. The conventional wisdom is that wafer-scale integration was proved futile in the 1980s. But that was when the large size of most circuit components made it inefficient to provide redundancy in anything other than highly replicated circuits, such as memories. In the coming decade, however, entire cores will be small enough to be fused out with relatively little functional loss. In addition, redundant vertical and horizontal pathways can be provided in 3D circuitry, so that a defect in part of one layer will not prevent functional access to components above, below, and beside it.

Combined, all these technologies can greatly decrease the cost of manufacturing the massive amounts of memory, processing power, and connectivity demanded for extremely powerful --- roughly brain-level --- artificial intelligence.


For example, if --- 11nm semiconductor lithography --- multilevel circuitry --- and --- integrated-photonic interconnect --- are all in mainstream production in ten years --- as many predict --- then one million dollars should be able to purchase a system with: --- roughly 4 million small processor cores, allowing a theoretical max of 4 thousand trillion instructions per second --- 32 TBytes of 2ns EDRAM, allowing roughly 400 trillion read-modify-writes to EDRAM per second --- over 200 TBytes of sub-20ns-read-access phase-change-memory (PCM), allowing roughly 160 trillion random reads per second --- and a global, sustainable, inter-processor bandwidth of over 20 trillion 64Byte payload messages per second.

The AI community does not know exactly how much representation, computation, processor-to-memory, and processor-to-processor capacity is needed for human-level computing. The estimates vary by four or five orders of magnitude. Some think we will have to match the complexity of the brain almost exactly to get brain-level performance, causing them to think we will have to wait until approximately 2040 to achieve human-level AI. But this fails to take into account the many superiorities electronic hardware has relative to wetware. From my research and calculations, I am relatively confident that the above computational resources --- that could be available for one million dollars by 2020 --- would have more than enough capability to provide something approaching --- or, very possibly, substantially surpassing --- all the useful talents at which a human mind can currently out perform computers.

In addition, a machine with this power could also execute tasks at which computers already substantially out perform humans at speeds, and with exact memory, that exceed that of humans by millions or trillions of times. Combining the types of computing at which humans and machines each currently excel will greatly amplify the power of artificial intelligence. Such a system could have a high-bandwidth, fine-grained interface between these two different types of computation. And it could have the ability to rapidly vary the degree of mixture between them in each of many different concurrent processes or sub-processes --- all under the dynamic control of powerful hierarchical mental behaviors that have been honed by automatic reinforcement learning. This mixture will enable artificial intelligences that are substantially sub-human in some ways, to be hundreds to millions of times more powerful than humans at tasks that now can only be performed by us --- such as --- trying to use on-line tools to find the set of legal cases that are most relevant to a new, complex legal problem --- or trying to find information on the internet in those situations in which Google doesn't seem helpful.


To show why the 2020 system hypothesized above would, most probably, be capable of human-level thinking --- let us assume half of its 200TBytes of PCM memory were used to represent nodes and links in an experientially grounded, self-organizing, semantic-net memory. Assume an average of 100 bytes is required to represent and properly index an occurrence of a pattern, or concept, represented by a node in that net. Assume that roughly another 100 bytes is required to represent one of the relationships of such a concept’s occurrence to another pattern or to one or more temporal, spatial, or semantic maps. With these assumptions this 100TBytes could create an experiential record storing an average of 1000 such nodes or links for each of one billion seconds. That’s roughly the equivalent of three pages of text to describe a person’s experiences for every second in over 31 years. When combined with the type of memory described in the paragraph below, this is almost certainly much, much more world knowledge than a human mind can store.

Continuing this simplified model of memory distribution --- let the remaining 100TByte of PCM store billions of patterns to represent and ground the meaning of the above mentioned nodes and links. This would include an invariant, hierarchical, self-organizing memory representing the composition, generalization, and similarity relationships between such patterns. This semantic net would include --- billions of patterns generalized from activation, or recorded, states in the system's network of sensory and semantic nodes and links --- and --- mappings between such generalized patterns, and their parts, and occurrences of such patterns in perceptions, thoughts, plans, imaginations, and memories. These generalized patterns would include billions of relatively simple sensory and motor patterns. They would also include more complex patterns representing concepts such as objects, persons, actions, emotions, drives, goals, and behaviors. These more complex patterns would include physical and mental behaviors and plans --- and their associated goals and other memories --- including feedback on their value and effectiveness. These patterns would include temporal and spatial relationships, and relationships defined by the relative roles of patterns in larger patterns. They would also include probabilistic statistics on the frequency, long term importance, and relationships between such patterns.

For many applications, such a system would contain many terabytes of information to help them excel at communicating with humans through --- text --- speech --- vision --- gestures --- facial expressions --- tones of voice --- and photorealistic, real-time, audio/video animation. Such systems would record --- tens of millions of compressed photographs, millions of which would be stored in morphable, semantically-labeled photosynths for generating 3D images and animations --- millions of seconds of compressed audio and moving images --- including models of humans communicating --- and --- many millions of mental patterns and behaviors relating to understanding human intentions, communications between humans, and communication with humans.

From the above we can see that the 200TBytes of storage --- provided by the hypothetical 2020 system --- particularly if it uses a context-sensitive, invariant representation scheme (of the type discussed in more detail below) --- is almost certainly enough to represent much more functional world knowledge than a human mind can store --- and to ground the concepts in such knowledge in an extremely rich web of sensory, cognitive, emotional, and behavioral memories and associations. This grounding should be much more than enough to give such a system's symbols --- true “meaning.”


This hypothetical 2020 system --- not only has enough capacity to more than represent human-level world knowledge --- it also appears to have enough computational and communication capacity to reason from such world knowledge faster than humans. The 2020 system’s ability to randomly read its PCM memory 160 trillion times per second, and to perform over a 400 trillion random read-modify-writes to portions of its EDRAM representing dynamic activation values of patterns stored in the PCM, give it tremendous power to reason from its world-knowledge. It would have enough power to perform relatively shallow and most-probable (i.e. subconscious) inferencing simultaneously from billions of somewhat activated patterns --- and --- relatively deep and/or broad (i.e., conscious) inferencing, involving tens of billions of multi-level spreading activations from each of a small number of highly activated patterns that were the focus of attention. This allows rich, deep, grounded, and highly dynamic activation states. Ones that would probably have more useful informational complexity than those in our own minds.

These dynamic activation states --- when combined with mental behaviors for dynamically selecting and focusing attention --- can give rise to a powerful combination of conscious and subconscious thought. In this combination, conscious thought would commonly result from massive activation from a relatively small number of concepts and their relationships. The concepts chosen for such massive conscious activation would be generated, tested, and selected by many billions, or trillions, of computations in the subconscious. These subconscious computations would be made in response to sensations, emotions, desires, goals, memories --- and --- from activations from current and recently consciously activated concepts. In such a system the distribution of activation energy between conscious and subconscious activations, and between various activations within the subconscious, can be rapidly varied. For example, this allows each of many increasingly higher scoring networks of activation in the subconscious to receive increasingly more activation energy to verify which of them actually warrant being thresholded into conscious attention.


In summary, the above numbers give us good reason to believe that within ten years it will be commercially viable to build and sell machines that have the representation, computation, processor-to-memory, and processor-to-processor capacities necessary to support human-level --- and likely superhuman-level --- intelligence.

As one former head of DARPA’s AI funding told me, “The hardware being there is a given. It’s the software that’s needed.”



THE SOFTWARE
=============

Tremendous advances have been made in artificial intelligence in the recent past. This is largely due to the ever increasing rate of progress in brain science. It is also due to the increasing power of the computers that researchers can experiment with.

One example of such recent progress is the paper “Learning a Dictionary of Shape-Components in Visual Cortex:...”, by Thomas Serre of Prof.Tomasa Poggio’s group at MIT. It describes a system that provides human-level performance in one limited, but impressive, type of human visual perception (http://cbcl.mit.edu/projects/cbcl/publications/ps/MIT-CSAIL-TR-2006-028.pdf  ). The Serre-Poggio system learns and uses patterns in a generalization and composition hierarchy. This allows efficient multiple use of representational components, and computations matching against them, in multiple higher level patterns. It allows the system to learn in compositional increments. It also provides surprisingly robust invariant representation. Such invariant representation is extremely important because it allows efficient non-literal matching, pattern recognition, and context appropriate pattern imagining and instantiation. Such non-literal match and instantiation tasks have --- until recently --- been among the major problems in trying to create human-like perception, cognition, imagination, and planning.

Although it is different than the Serre-Poggio system, the system described in Geoff Hinton’s Google Tech Talk at http://www.youtube.com/watch?v=AyzOUbkUf3M  demonstrates a character recognition architecture that shares many of these same beneficial characteristics --- including a hierarchical, scalable, and invariant representation/computation scheme that can be efficiently and automatically trained. The Hinton scheme is quite general, and can be applied to many types of learning, recognition, and context sensitive imagining. The architecture described by Jeff Hawkins et al. of Numenta, Inc. in “Towards a Mathematical Theory of Cortical Micro-circuits” (http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1000532  ) also shares the concepts of hierarchical memory and invariance, and provides a potentially powerful and general computational model that attempts to describe the functioning of the human cortex in terms of its individual layers.

Similar amazing advances have been made in understanding other brain systems --- including those that control and coordinate the behavior of, and between, multiple areas in the brain --- and those that focus attention and decide which of competing actions to take or consciously consider.

These advances, and many more, provide enough understanding that we can actually start experimenting with designs for powerful artificial minds. It’s not as if we have exact blue prints. But we do have a good overview, and good ideas on how to handle every problem I have ever heard mentioned in regard to creating roughly brain-like AI. As Deb Roy, of MIT, once agreed with me after one of his lectures, there are no problems between us and roughly human-level AI that we have no idea how to solve. The major problem that exists is the engineering problem of getting all the pieces to fit and work together well, automatically, and within a commercially-viable computational budget. That will take experimentation.

But we certainly do know enough to design and build general artificial intelligences that could provide useful functions.


The most complete, currently-publicly-available artificial brain architecture of which I am aware, is the OpenCogPrime architecture. It has been created by the open-source AGI initiative headed by
Ben Goertzel. There may be other equally complete and impressive brain architectures available to the public. But since I do not know them, let me give a brief -- but hopefully revealing – overview of the OpenCog architecture --- as I understand it, in combination with some of my own thinking on the subject. (The OpenCog architecture is described at http://www.opencog.org/wiki/OpenCogPrime:WikiBook#Introduction .)

OpenCog starts with a focus on “General Intelligence”, which it defines as “the ability to achieve complex goals in complex environments.” “AGI” stands for Artificial General Intelligence. It is focused on automatic, interactive learning, experiential grounding, self understanding, and both conscious (focus-of-attention) and unconscious (less attended) thought.

It records its sensory and emotional experiences, finds repeated patterns in such recordings, makes generalizations and compositions out of such patterns --- all through multiple levels of generalization and composition --- based on spatial, temporal, and learned-pattern-defined relationships. It uses Bayesian mathematics --- based on the frequencies of the detection of such patterns and their relationships --- in a way that allows inferences to be drawn from many billions of activated patterns at once.

Patterns -- which can include behaviors (including those that control the operation of the mind itself) -- are recorded, modified, generalized, and deleted all the time. They have to compete for their computational resources, including memory space, and, thus, their own continued existence. Re-enforcement learning and other forms of credit assignment are used to determine which patterns are useful enough to be kept, and for how long. This results in a self-organizing network of similarity, generalization, and composition 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 also used to weight patterns for long-term and current short term importance, based on the 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, contexts, and goals, and for the relative usefulness of various past inferencing behaviors in such contexts -- significantly narrow and focus attention and spreading activation. This helps avoid the pitfalls of combinatorial explosion, and it tends to result in context-appropriate perception, cognition, and behavior.

OpenCog uses evolutionary program learning --- somewhat like genetic programming --- to increase the system’s ability to learn and tune: generalizations of patterns; classifiers; creative ideas; and behaviors --- including physical, attention focusing, inferencing, and learning behaviors. This evolutionary learning is made more powerful by being used by --- and by using --- the rest of the system. This includes the system’s composition and generalization hierarchy, its network of probabilistic associations, its inferencing, and its reinforcement learning. Evolutionary programs can be used by the system’s experiential probabilistic learning. Such programs can, themselves --- along with experientially learned patterns --- be incorporated --- with or without modification --- into the learning of new evolutionary programs. A compositional and generalization hierarchy including such evolutionarily-learned programs enables complex programs to be learned more efficiently in incremental steps from more simple ones. Experiential memories help guide and evaluate the evolutionary process, including reducing the computation required for estimating the fitness functions for many evolutionary candidates. Experiential memories can also provide information for probabilistically inferring which programs are appropriate to employ, with which parameters, in which contexts.


Taken together, software architectures like those discussed above --- when combined with the hardware likely to be available within a decade --- will allow AGI systems 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 automatically learn roughly the equivalent of human-level world knowledge. After all, it takes over twenty years for most human minds to train up. But there is reason to believe substantial portions of such machine learning could be performed in parallel. And since such machines will be capable of remembering vastly more detail than humans, their learning should be much faster. Such machine learning is likely to be better grounded in physical reality, if such machines can control robotic bodies with human-like senses that enable them to learn by exploring the physical world, as do human children --- or by, at least, having the equivalent in a fairly accurate virtual world. The learning of many concepts would be improved by having human teachers. Once such a system achieves a certain level of world-knowledge --- including a child’s level of common-sense physics, basic human behavior, natural-language, and visual scene understanding --- they will be able to rapidly learn by reading and viewing images and diagrams from large libraries of digitally recorded books, and from media on the web.

And once one such system has been fully trained in basic world knowledge --- or in the knowledge relating to a given field of expertise that is linked to a common representation of such world knowledge--- that knowledge can be copied to another similar machine in seconds or minutes.



WILL IT WORK?
============

The answer is most probably, yes, because such systems will: ---

-a--- in multiple important ways --- work like the human brain, itself;

-b--- have enough representation, computation, and interconnect capacity to make types of AI that were never before even close to possible for most in the AI community --- not only possible --- but commercially practical --- including the ability to represent and rapidly reason from grounded, human-level world knowledge --- and

-c--- benefit from the explosion of AI related advances that will occur in this decade.

This explosion of AI-related advances --- in addition to the hardware advances described above --- will occur in: --- brain science --- generalized machine learning and inferencing --- attention and inference control --- large-scale semantic web applications --- learning and reasoning from self-organizing ontologies --- natural language understanding and generation --- common-sense and world-knowledge learning and computing --- evolutionary learning --- machine vision --- multimedia indexing --- command and control --- national security and defense applications --- search --- robotics --- personal assistants --- web agents --- user interfaces --- and more human-like characters for video games and virtual realities. All of this will be in addition to the increasing research that will be performed in AGI, itself.

Within three to seven years, hardware having the effective representation, computation, and inter-connect of small mammal and, then, primate brains will be available --- at sufficiently low costs that thousands of such systems will be used by academic and corporate teams to experiment in such fields. All this research will help identify, refine, and tune various general algorithms that could be put together to create powerful generalized “thought robots” --- i.e., powerful artificial intelligences that can --- automatically --- or with relatively little handcrafting --- tune their learning and mental behaviors to achieve various goals across a broad range of applications.

AGI is not currently competitive for most applications, because its general algorithms tend to require much more training, memory, and computation than AI systems handcrafted by humans to solve a particular set of problems. But many of the learning, inferencing, and inference control mechanisms deployed in more narrow applications can be generalized to have applicability to AGI. And many early AGI’s will have handcrafted parts to make them more competitive for specific applications. As memory, computation, and interconnect costs drop drastically relative to programming costs --- particularly relative to the cost of handcrafting AIs for extremely complex problems --- larger, more general, and more capable AGI will become ever more competitive.

Once created, AGI will be particularly attractive for corporate and cloud computing --- because it can automatically be adapted to the many different uses that different people will want from AI services --- and because it can provide superintelligent user interfaces --- using text, speech, audio, vision, and the real-time generation of animation --- to make it easy for users to instruct, monitor, and learn relevant information from such machines.


So can human-level AGI be built?

Yes!

The only question is how fast. And it is almost certain that if the right people, got the right money, it could be built within ten years --- that is --- by no later than Twenty-Twenty.

Making this happen should be our nation’s "Twenty-Twenty vision" because machine superintelligence is the most transformative technology of all.



THE CONSEQUENCES
=================

It is hard to overstate the economic value and transformative power of the types of machines that will probably be built by 2020 --- and if not by then --- by 2030.

The one-million-dollar 2020 hardware hypothesized above could be rented out on the web, at a profit, for roughly $50 an hour. It --- would have superhuman concentration --- could work close to 24/7 --- could perform many types of reasoning tasks millions of times faster than a human --- and --- if connected to a cloud of similar machines that stored a large percent of human knowledge in instantly-accessible semantic deep structure --- it would, in effect, have photographic memory for almost all of recorded human knowledge. It is not unrealistic to think that for a large number of tasks such a machine could do work at a higher rate than one hundred programmers, lawyers, doctors, or managers.

If such a system were part of a computing cloud --- then an average of --- 64 thousand cores --- 500 GBytes of 2ns EDRAM --- 3 TBytes of 20ns-read-access PCM memory --- and --- over 300 billion global, 64Byte messages per second --- could be provided to serve an individual user of a wireless mobile phone; retinal-scanning, headset computer; or other personal device --- at roughly the same price currently charged for long distance phone calls. This should be enough power to provide users with moderately good --- natural language --- vision --- real-time animation --- intelligent search --- semantic web reasoning --- and machine-mediated collective computing.

Most of the time they were connected, users would not even begin to fully use the 64K processor chunk of hardware described above, but there would frequently be tasks demanding more power --- such as understanding difficult natural language constructions --- performing computationally intensive queries, summaries, and reasoning --- and --- synthesizing creative solutions to complex problems. Larger portions of the cloud could be used in a multiplexed manner for such tasks. Users who utilize more than a certain amount of the cloud's resources in a given time could be notified that they were about to do so, and be billed extra for it at less than one dollar for a the equivalent of using one of the above described one-million-dollar 2020 machine’s worth of hardware for one minute.

That one dollar should be enough, for example, to get a reasonably well reasoned legal brief on a moderately complex issue --- something that would cost several hundred to several thousand dollars from most American lawyers.

Even if we make the extremely conservative assumption that our one-million-dollar 2020 machine could only simultaneously do the work of ten human lawyers, doctors, financial experts, or managers --- that would mean it could provide the services of such a professional or manager for $5 dollar per hour --- making most such highly educated professionals or managers unemployable at the current minimum wage.

Furthermore if new areas of electronics --- such as 3D, carbon nanotube, graphene, nanowire, quantum dot, spintronic, quantum entanglement, molecular, neuromorphic, and self-organizing electronics --- keep Moore’s law going for several decades past the final density expected for traditional silicon electronics --- in twenty to thirty years a machine of power similar to the one-million-dollar 2020 system might cost less than a current personal computer. Such a continuation of Moore's law is likely. This is because machine superintelligences can be produced at even the 22nm node at sufficiently low prices that they could be commercially useful for greatly increasing the rate of development in electronics and electronic design. If such cost reductions are, in fact, obtained, virtually all human mental jobs could be replaced for one or two pennies an hour in thirty years. If superintelligence is used to speed advances and cost reductions in robotics --- all of humanity --- including in places like China, India, Vietnam, and Haiti --- will cease to be competitive for most current forms of work.

AGI will create a historical “singularity” of the type Ray Kurzweil has done so much to popularize. That is, a technological change so powerful --- that --- when combined with the massive acceleration it will cause in --- the internet --- electronics --- computing --- robotics --- nanotechnology --- and --- biotechnology --- it will warp the very fabric of human economies, cultures, values, and societies in somewhat the same way the singularity of a black hole warps the fabric of space-time --- and is believed --- by some --- to create an entirely new universe -- one largely disconnected from its past in space and time.

Can we, or should we, stop the advent of superintelligence?

No. It is futile to try.

Too many people already know how much technological, economic, military, cultural, and political advantage can be gained by the nations and corporations that are first to substantially deploy it. It cannot be stopped because electronic technology and our understanding of intelligence are already so advanced that in a decade the development of superintelligence will be well within the economic and intellectual grasp of a most nations, thousands of corporations, and hundreds of universities. It is already within the grasp of the world’s leading nations and technological companies. It cannot be stopped by international agreements, because --- compared to the development of nuclear weapons --- in a decade, machine intelligence can be developed for very little money, in very little space, with relatively little electricity. Its development would be very difficult to detect, prove, or stop.


There are many reasons we should want --- rather than oppose --- the development of superintelligence. It --- and the rapid advances in technology and productivity it will bring --- could be a force for tremendous good.

It could create a world of material, medical, mental, and intellectual well being and richness. It could help us develop highly efficient, sustainably, less-polluting, factories, farms, stores, corporations, and transportation. It could teach us how to cure most disease, how to keep our bodies younger longer, and how to make our minds work in more powerful and satisfying ways. It could help us to become a truly enlightened species. It could educate all of us, with virtual tutors more knowledgeable and more capable of explaining things to us than the most brilliant and attentive human teachers. It could learn to know each of us better than we know ourselves, and to provide us with personal counseling superior to that of the best psychologist or friend.

It could help us better simulate and determine the costs, benefits, and risks of personal, corporate, and governmental decisions. It could enable people to communicate, collaborate, and deliberate with an efficiency and fairness never before possible. It could help us better deal with the rapid changes it will produce --- such as the fact it will end most current ways of earning a living in the industrialized world --- by helping us to create a new, fair, and sustainable social contract --- and new types of meaningful work ---- such as sharing more responsibility in much more participatory local, regional, national, and world governments and institutions. It can allow us to have AGI-mediated, real-time virtual conversations, debates, celebrations, songs, dances, games, and prayers --- in which hundreds, thousands, millions, or billions of people take part.

The world is facing many challenges that seem beyond the capacity of our current political institutions to solve. It is arguable we need superintelligence to help us find how to provide food, shelter, clothing, medical services, education, meaningful lives --- and --- most importantly --- peace --- for the projected 9 billion people that will populate earth by 2050. Most of these will come from societies that are brutally poor --- and yet most of them will have access to the extremely powerful --- but, by 2050, inexpensive --- personal information devices of the future --- video devices that will likely teach them to want as much power and material wealth as people in the richest nations. It is arguable that we will need superintelligence to help us deal with such problems without poisoning our planet with pollution --- or by destroying it with war or terrorism.


But superintelligence, and the technologies it will bring, could also cause great harm.

Unless we are careful --- in addition to putting most of us out of work --- superintelligence can be used to create surveillance systems and robotic police or military forces that could enable one class, group, person, or system of machines to create a powerful oligarchy or dictatorship. Unethical people, governments, or machines will almost certainly use superintelligences to constantly try hacking into our networks and intelligent machines --- trying to take control of them for their own selfish purposes.

AGI can create virtual realities, friends, and lovers that are much more attractive, attentive, sensitive, romantic, funny, and seductive than those that are real. These virtual worlds and personalities could weaken the bonds between humans, and could seduce us into increasingly turning over more attention and power to machines, and to the virtual worlds they generate for us --- whether those machines are controlled by businessmen, political leaders, or machines themselves. Low income housing may become stackable 4 x 4 x 8 foot plastic pods with super HD 3D virtual interfaces, in which the elites provide the masses with a bare-minimum physical reality, but an extremely rich --- and much less expensive --- virtual one.

Human laziness may well lead us to turn too much power over to superintelligences --- so much so that we might soon be at their mercy. Ultimately, the machines themselves might well take over. And if they do --- it is not clear whether they would like us enough to let us keep consuming so much of the earth’s resources --- which they, themselves, could use for their own purposes --- and their own progeny.


Some transhumanists say --- the only way in which what we value as "human" can remain competitive in a world bound to be ruled by superintelligences is for us to increasingly merge our values, memories, consciousnesses, and bodies with such machines. They say our seduction by virtual worlds, friends, and lovers is a good thing --- because it will make --- what they view --- as the necessary man-machine merger --- more emotionally acceptable. To make our lives more meaningful --- they say --- we should view the machines as our kin and our posterity.

Some transhumanists suggest it is necessary for our very survival that we place into, or onto, our brains high bandwidth connections to superintelligences. Preferably such connections will be to a World-Wide-Web of other people similarly connected to such machines. This would make us into Star-Trek-like borgs --- but, let us hope, ones with substantially more individualism, humor, and happiness.

Other transhumanists suggest uploading our minds to run on such machines so they can “live” for billions of years.

To most people all this sound like a whipped out science-fiction horror flick. But there are reasons to believe that within only decades --- almost certainly by the end of this century --- much of this could, in fact, come true.

The transhumanists may be right. We humans largely rule the earth because our intelligence and knowledge excels that of all other species. By analogy, it only makes sense that --- starting in several decades when there are likely to be networks of many superintelligences --- each thousands of times smarter than humans --- there will ultimately come a time when machines take domination away from us.

That is, perhaps, unless we join them, and make them part us, and us part them. There are already many who look forward to connecting their brains to superintelligences --- and it is almost certain, that once superintelligence arrives, the people who use such implanted, high bandwidth, connections to such machines will be more successful than those who do not.

But even the transhumanist scenario requires that humanity act intelligently and wisely if the transition to humanity+ is to be a happy one.

How we develop --- use --- and control superintelligence is one of the greatest challenges facing mankind. We cannot stop its advent, but we can try to control it --- to reduce its danger, at least to some degree. Great flexibility is possible in the design of AGI’s, and we should be careful to learn what types of machines are likely to be more safe and what types are likely to be more dangerous. We should learn how to best use the safer types of superintelligence to protect us against the more dangerous. If you care about humanity --- more important than creating superintelligence per se --- is creating super-intelligence that is well combined with the wisdom, compassion, and voices and concerns of billions of individual human beings.


That is why, ultimately --- from humanity’s standpoint --- the most important technology of all is collective intelligence.

It is the technology of using the internet, computers, and, soon, superintelligence, to enable groups, corporations, nations --- and ultimately the world --- to think and act together more intelligently, successfully, and humanely --- as we --- as a species --- have to navigate in an ever more rapidly-changing future.

And that is why --- the single most important use of superintelligence --- is to help give mankind enough collective intelligence that --- for decades , or, perhaps, even centuries --- humanity can safely and happily travel into that rapidly-changing future.





(For a more complete discussion of collective intelligence see http://fora.humanityplus.org/index.php?/topic/70-collective-intelligence-our-only-hope-for-surviving-the-singularity/ .)


No comments:

Post a Comment