How to Make a Mind BY

ANDREW OSTROVSKY / ISTOCKPHOTO

Can nonbiological have he mammalian has a distinct aptitude not found in any other class of animal. We are capable of ­hierarchical thinking, of real minds of their own? In understanding a structure composed of diverse elements arranged this article, drawn from his in a pattern, representing that arrangement with a symbol, and then using that symbol as an element in a yet more elaborate latest book, futurist/inventor configuration. Ray Kurzweil describes the TThis capability takes place in a brain structure called the neocortex, which in humans has achieved a threshold of sophistication and capacity such that future of intelligence— we are able to call these patterns ideas. We are capable of building ideas that are ever more complex. We call this vast array of recursively linked ideas artificial and otherwise. knowledge. Only Homo sapiens have a knowledge base that itself evolves, grows exponentially, and is passed down from one generation to another.

From How to Create a Mind by Ray ­Kurzweil. Copyright © 2012, Ray Kurzweil. ­Reprinted by arrangement with Viking, a member of Penguin Group (USA) Inc.

14 THE FUTURIST March-April 2013 • www.wfs.org © 2013 World Future Society • 7910 Woodmont Avenue, Suite 450, Bethesda, MD 20814, U.S.A. • All rights reserved. We are now in a position to speed own private stores of personal data driver of impending dangers is al- up the learning process by a factor of today. ready being installed in cars. One thousands or millions once again by Last but not least, we will be able such technology is based in part on migrating from biological to nonbio- to back up the digital portion of our the successful model of visual pro- logical intelligence. Once a digital intelligence. It is frightening to con- cessing in the brain created by MIT’s neocortex learns a skill, it can trans- template that none of the informa- Tomaso Poggio. Called MobilEye, it fer that know-how in minutes or tion contained in our neocortex is was developed by Amnon Shashua, even seconds. Ultimately we will backed up today. There is, of course, a former postdoctoral student of create an artificial neocortex that has one way in which we do back up Poggio’s. It is capable of alerting the the full range and flexibility of its some of the information in our driver to such dangers as an im- human counterpart. brains: by writing it down. The abil- pending collision or a child running Consider the benefits. Electronic ity to transfer at least some of our in front of the car and has recently circuits are millions of times faster thinking to a medium that can out- been installed in cars by such manu- than our biological circuits. At first last our biological bodies was a huge facturers as Volvo and BMW. we will have to devote all of this step forward, but a great deal of data I will focus now on language tech- speed increase to compensating for in our brains continues to remain nologies for several reasons: Not sur- the relative lack of parallelism in our vulnerable. prisingly, the hierarchical nature of computers. Parallelism is what gives language closely mirrors the hierar- our brains the ability to do so many The Next Chapter in chical nature of our thinking. Spoken different types of operations—walk- language was our first technology, ing, talking, reasoning—all at once, with written language as the second. and perform these tasks so seam- Artificial intelligence is all around My own work in artificial intelli- lessly that we live our lives blissfully us. The simple act of connecting gence has been heavily focused on unaware that they are occurring at with someone via a text message, language. Finally, mastering lan- all. The digital neocortex will be e-mail, or cell-phone call uses intelli- guage is a powerfully leveraged ca- much faster than the biological vari- gent algorithms to route the infor- pability. , the IBM computer ety and will only continue to in- mation. Almost every product we that beat two former Jeopardy! cham- crease in speed. touch is originally designed in a col- pions in 2011, has already read hun- When we augment our own neo- laboration between human and arti- dreds of millions of pages on the cortex with a synthetic version, we ficial intelligence and then built in Web and mastered the knowledge won’t have to worry about how automated factories. If all the AI sys- contained in these documents. Ulti- much additional neocortex can phys- tems decided to go on strike mately, machines will be able to ically fit into our bodies and brains, tomorrow, our civilization would be master all of the knowledge on the as most of it will be in the cloud, like crippled: We couldn’t get money Web—which is essentially all of the most of the computing we use today. from our bank, and indeed, our knowledge of our human–machine We have about 300 million pattern money would disappear; communi- civilization. recognizers in our biological neocor- cation, transportation, and manufac- One does not need to be an AI ex- tex. That’s as much as could be turing would all grind to a halt. For- pert to be moved by the performance squeezed into our skulls even with tunately, our intelligent machines are of Watson on Jeopardy! Although I the evolutionary innovation of a not yet intelligent enough to orga- have a reasonable understanding of large forehead and with the neocor- nize such a conspiracy. the methodology used in a number tex taking about 80% of the available What is new in AI today is the vis- of its key subsystems, that does not space. As soon as we start thinking cerally impressive nature of publicly diminish my emotional reaction to in the cloud, there will be no natural available examples. For example, watching it—him?—perform. Even a limits—we will be able to use bil- consider ’s self-driving cars, perfect understanding of how all of lions or trillions of pattern recogniz- which as of this writing have gone its component systems work would ers, basically whatever we need, and over 200,000 miles in cities and not help you to predict how Watson whatever the law of accelerating re- towns. This technology will lead to would actually react to a given situa- turns can provide at each point in significantly fewer crashes and in- tion. It contains hundreds of inter- time. creased capacity of roads, alleviate acting subsystems, and each of these In order for a digital neocortex to the requirement of humans to per- is considering millions of competing learn a new skill, it will still require form the chore of driving, and bring hypotheses at the same time, so pre- many iterations of education, just as many other benefits. dicting the outcome is impossible. a biological neocortex does. Once a Driverless cars are actually already Doing a thorough analysis—after the single digital neocortex somewhere legal to operate on public roads in fact—of Watson’s deliberations for a and at some time learns something, Nevada with some restrictions, al- single three-second query would however, it can share that knowl- though widespread usage by the take a human centuries. edge with every other digital neocor- public throughout the world is not One limitation of the Jeopardy! tex without delay. We can each have expected until late in this decade. game is that the answers are gener- our own private neocortex extenders Technology that intelligently ally brief: It does not, for example, in the cloud, just as we have our watches the road and warns the pose questions of the sort that ask

www.wfs.org • THE FUTURIST March-April 2013 15 contestants to name the five primary pha, along with continually updated themes of A Tale of Two Cities. To the data on topics ranging from econom- extent that it can find documents ics to physics. that do discuss the themes of this In a private conversation I had novel, a suitably modified version of with him, Wolfram estimated that Watson should be able to respond to self-organizing methods such as this. Coming up with such themes those used in Watson typically on its own from just reading the achieve about an 80% accuracy when book, and not essentially copying they are working well. Alpha, he the thoughts (even without the pointed out, is achieving about a words) of other thinkers, is another 90% accuracy. Of course, there is matter. Doing so would constitute a self-selection in both of these accu- higher-level task than Watson is ca- racy numbers, in that users (such as pable of today. myself) have learned what kinds of It is noteworthy that, although questions Alpha is good at, and a Watson’s language skills are actually “We could give our new similar factor applies to the self- somewhat below that of an educated organizing methods. Some 80% ap- human, it was able to defeat the best brain a more ambitious pears to be a reasonable estimate of two Jeopardy! players in the world. It goal, such as contributing how accurate Watson is on Jeopardy! could accomplish this because it is queries, but this was sufficient to de- able to combine its language ability to a better world. A goal feat the best humans. and knowledge understanding with along these lines, of It is my view that self-organizing the perfect recall and highly accurate methods such as I articulate as the memories that machines possess. course, raises a lot of pattern-recognition theory of mind, That is why we have already largely questions: or PRTM, are needed to understand assigned our personal, social, and the elaborate and often ambiguous historical memories to them. Better for whom? Better hierarchies we encounter in real- Wolfram|Alpha is one important in what way?” world phenomena, including human system that demonstrates the language. Ideally, a robustly intelli- strength of computing applied to or- gent system would combine hierar- ganized knowledge. Wolfram|Alpha chical intelligence based on the is an answer engine (as opposed to a PRTM (which I contend is how the search engine) developed by British human brain works) with precise mathematician and scientist Stephen codification of scientific knowledge Wolfram and his colleagues at per person?” (Answer: Monaco, with and data. That essentially describes a Wolfram Research. For example, if $212,000 per person in U.S. dollars), human with a computer. you ask Wolfram|Alpha, “How or “How old is Stephen Wolfram?” We will enhance both poles of intel- many primes are there under a mil- (he was born in 1959; the answer is ligence in the years ahead. With re- lion?” it will respond with “78,498.” 52 years, 9 months, 2 days on the gard to our biological intelligence, al- It did not look up the answer, it com- day I am writing this). Alpha is used though our neocortex has significant puted it, and following the answer it as part of Apple’s ; if you ask Siri plasticity, its basic architecture is lim- provides the equations it used. If a factual question, it is handed off to ited by its physical constraints. Put- you attempted to get that answer us- Alpha to handle. Alpha also handles ting additional neocortex into our ing a conventional search engine, it some of the searches posed to Micro- foreheads was an important evolu- would direct you to links where you soft’s Bing search engine. tionary innovation, but we cannot could find the algorithms required. Wolfram reported in a recent blog now easily expand the size of our You would then have to plug those post that Alpha is now providing frontal lobes by a factor of a thou- formulas into a system such as successful responses 90% of the time. sand, or even by 10%. That is, we can- Mathematica, also developed by He also reports an exponential de- not do so biologically, but that is ex- Wolfram, but this would obviously crease in the failure rate, with a half- actly what we will do technologically. require a lot more work (and under- life of around 18 months. It is an Our digital brain will also accom- standing) than simply asking Alpha. impressive system, and uses hand- modate substantial redundancy of Indeed, Alpha consists of 15 mil- crafted methods and hand-checked each pattern, especially ones that oc- lion lines of Mathematica code. What data. It is a testament to why we cre- cur frequently. This allows for robust Alpha is doing is literally computing ated computers in the first place. As recognition of common patterns and the answer from approximately 10 we discover and compile scientific is also one of the key methods to trillion bytes of data that has been and mathematical methods, comput- achieving invariant recognition of carefully curated by the Wolfram ers are far better than unaided hu- different forms of a pattern. We will, Research staff. You can ask a wide man intelligence in implementing however, need rules for how much range of factual questions, such as, them. Most of the known scientific redundancy to permit, as we don’t “What country has the highest GDP methods have been encoded in Al- want to use up excessive amounts of

16 THE FUTURIST March-April 2013 • www.wfs.org memory on very common low-level tinual background task. It would be known scientific methods and applies patterns. very beneficial if human brains did them to carefully collected data. This the same thing. type of system is also going to con- Educating Our Nonbiological Brain I would also provide a module tinue to improve, given Stephen that identifies open questions in ­Wolfram’s observation of an exponen- A very important consideration is every discipline. As another contin- tial decline in error rates. the education of a brain, whether a ual background task, it would search Finally, our new brain needs a pur- biological or a software one. A hier- for solutions to them in other dispa- pose. A purpose is expressed as a se- archical pattern-recognition system rate areas of knowledge. The knowl- ries of goals. In the case of our bio- (digital or biological) will only learn edge in the neocortex consists of logical brains, our goals are about two—preferably one—hierar- deeply nested patterns of patterns established by the pleasure and fear chical levels at a time. To bootstrap and is therefore entirely metaphori- centers that we have inherited from the system, I would start with previ- cal. We can use one pattern to pro- the old brain. These primitive drives ously trained hierarchical networks vide a solution or insight in an ap- were initially set by biological evolu- that have already learned their les- parently disconnected field. tion to foster the survival of species, sons in recognizing human speech, As an example, molecules in a gas but the neocortex has enabled us to printed characters, and natural-­ move randomly with no apparent sublimate them. Watson’s goal was language structures. sense of direction. Despite this, vir- to respond to Jeopardy! queries. An- Such a system would be capable of tually every molecule in a gas in a other simply stated goal could be to reading natural-language documents beaker, given sufficient time, will pass the Turing test. To do so, a digi- but would only be able to master ap- leave the beaker. This provides a tal brain would need a human narra- proximately one conceptual level at a perspective on an important ques- tive of its own fictional story so that time. Previously learned levels would tion concerning the evolution of in- it can pretend to be a biological hu- provide a relatively stable basis to telligence. Like molecules in a gas, man. It would also have to dumb it- learn the next level. The system can evolutionary changes also move self down considerably, for any sys- read the same documents over and every which way with no apparent tem that displayed the knowledge of over, gaining new conceptual levels direction. Yet, we nonetheless see a Watson, for instance, would be with each subsequent reading, similar movement toward greater complex- quickly unmasked as nonbiological. to the way people reread and achieve a ity and greater intelligence, indeed More interestingly, we could give deeper understanding of texts. Billions to evolution’s supreme achievement our new brain a more ambitious of pages of material are available on of evolving a neocortex capable of goal, such as contributing to a better the Web. Wikipedia itself has about 4 hierarchical thinking. So we are able world. A goal along these lines, of million articles in the English version. to gain an insight into how an appar- course, raises a lot of questions: Bet- I would also provide a critical- ently purposeless and directionless ter for whom? Better in what way? thinking module, which would per- process can achieve an apparently For biological humans? For all con- form a continual background scan of purposeful result in one field (bio- scious beings? If that is the case, who all of the existing patterns, reviewing logical evolution) by looking at an- or what is conscious? their compatibility with the other pat- other field (thermodynamics). As nonbiological brains become as terns (ideas) in this software neocor- We should provide a means of capable as biological ones of effect- tex. We have no such facility in our stepping through multiple lists simul- ing changes in the world—indeed, biological brains, which is why taneously to provide the equivalent ultimately far more capable than un- people can hold completely inconsis- of structured thought. A list might be enhanced biological ones—we will tent thoughts with equanimity. Upon the statement of the constraints that a need to consider their moral educa- identifying an inconsistent idea, the solution to a problem must satisfy. tion. A good place to start would be digital module would begin a search Each step can generate a recursive with one old idea from our religious for a resolution, including its own search through the existing hierarchy traditions: the golden rule. ❑ cortical structures as well as all of the of ideas or a search through available vast literature available to it. A reso- literature. The human brain appears About the Author lution might mean determining that to be only able to handle four simul- Ray Kurzweil is an inventor, one of the inconsistent ideas is simply taneous lists at a time (without the writer, and futurist. Among incorrect (if contraindicated by a pre- aid of tools such as computers), but his honors are the MIT- ponderance of conflicting data). More there is no reason for an artificial neo- Lemelson Prize, the Na- constructively, it would find an idea cortex to have such a limitation. tional Medal of Technology, and, in 2002, at a higher conceptual level that re- We will also want to enhance our induction solves the apparent contradiction by artificial brains with the kind of intel- into the U.S Patent Office’s providing a perspective that explains ligence that computers have always National Inventor’s Hall of each idea. The system would add this excelled in, which is the ability to Fame. resolution as a new pattern and link master vast databases accurately and This article was excerpted to the ideas that initially triggered the implement known algorithms quickly from his most recent book, search for the resolution. This critical and efficiently Wolfram|Alpha How to Create a Mind thinking module would run as a con- uniquely combines a great many ­(Viking, 2012).

www.wfs.org • THE FUTURIST March-April 2013 17