www.computer.org/intelligent

Mindless Intelligence

Jordan B. Pollack

Vol. 21, No. 3 May/June 2006

This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

© 2006 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

For more information, please see www.ieee.org/portal/pages/about/documentation/copyright/polilink.html. The Future of AI

Mindless Intelligence

Jordan B. Pollack, Brandeis University

ost of my 25 years of professional involvement in AI have been focused on M research far from its mainstream, not because of any antisocial tendencies on my part, but because of certain dilemmas inherent in the field. The first dilemma con- fronting AI is that both single-celled and multicelled animals survive and reproduce very

well without any nervous system at all, and “lower tems, animals, and humans and has spun out indus- animals,” even insects, organize into thriving soci- tries such as Lisp machines, expert systems, data AI has stalled because eties without any symbols, logic, or language, bee mining, and even Internet search. dancing and birdsong notwithstanding. These phe- of its preoccupation nomena led me to delve into nonsymbolic models Don’t promise the practically and ask how complex hierarchal representations and impossible with simulating the sustained state-changing procedures might naturally We all agree on AI’s fundamental hypothesis, that emerge from iterative numeric systems such as asso- physical machines have the capacity for intelligence. human mind. By ciative or connectionist neural networks. Unfortunately, this hypothesis can neither be proven The second dilemma is that the kind of mind we nor refuted scientifically, but realized only by studying intelligence in AI seek to discover, one that “runs” on the human demonstration. And until it has been convincingly brain yet might be portable to another universal demonstrated, it must remain in scientific limbo. in natural systems, machine, wouldn’t even exist without having co- Ordinary citizens and funding bureaucrats don’t evolved with the brain—a chicken-and-egg problem. know whether AI is tardy, like mechanical flight, outside the mind, we So, while many of my connectionist colleagues which emerged from limbo after several hundred migrated with US National Institutes of Health fund- years of failure, or magical, like ESP or the can reinvigorate the ing into cognitive or computational neuroscience, alchemists’quest to turn lead into gold. Perhaps there trying to understand how the human brain works, I is even an impossibility proof waiting around the cor- field. focused instead on what natural process could design ner, as has put to rest quixotic notions such as time and fabricate machinery as complex as the brain. travel (Einstein) and perpetual motion (Ludwig I ended up working closer to the field of artificial Boltzmann). Who wants to fund a field that might be life, seeking to understand how evolution, a mind- proven impossible tomorrow? less iterative reproduction system, could eventually So AI, which represents one of the greatest intel- lead to machines whose complexity and reliability lectual and engineering challenges in human his- dwarfs the product of the largest teams of human tory—and should command the same fiscal re- engineers. sources as efforts to cure cancer or colonize Mars—is On this 50th birthday of artificial intelligence, I sometimes relegated to a laughingstock, because we would like to reflect on what I feel has been its great can’t prevent bogus claims from cropping up in mistake, and propose a corrective course for the next newspapers and books. We cannot seem to convince 50 years. But before analyzing this mistake, I want the public that humanoids and Terminators are just to say that AI is a great human endeavor with a col- Hollywood special effects, as science-fictional as the orful cast and many partial successes. It has provided little green men from Mars! frameworks for formally studying biological sys- Still, some want to keep pursuing the same old AI

50 1541-1672/06/$20.00 © 2006 IEEE IEEE INTELLIGENT SYSTEMS Published by the IEEE Computer Society goals: “What are the missing pieces neces- associative and matrix models of mathemat- cognitive structures such as verb conjuga- sary to achieving human-level common ical psychology, to Markovian models, to tion. Why is simulating the human mind sense?” “Let’s do a project to gain human- both game and decision theories, to early more important than simulating cellular level performance in a (nonchess) domain.” neural networks (the perceptron disaster), to metabolisms, insect or animal intelligence, “We will build natural language software simulations of evolution and organic self- complex pattern formation, or distributed that’s human-level in ability.” “Soon com- organization. The early success of low-hang- control of complex ecologies? It must be puters will be fast enough to supply human- ing symbolic fruit through Lisp program- because, as a mirror of our own intelli- level intelligence to humanoid robots.” ming led to the pursuit of the “mythical man gence, the mindless iterative and numeric AI won’t be a gift of more CPU time. If it module,” a computer program that has the computing we scientifically uncover in were, we would have already glimpsed real “look and feel” of human cognition yet is nature doesn’t compare to the perfectly log- AI on supercomputers or large clusters, yet something more than an Eliza. ical indefatigable mind of Hollywood char- nothing of the kind has occurred. We don’t John Searle’s “Chinese Room” argument5 acters such as Mr. Spock and Commander need faster chips to make robots smarter, is hateful because, in fact, he’s correct. Nei- Data, NP-completeness notwithstanding. since we can link a robot’s body to its super- ther the room nor the guy in it pushing sym- To repair this mistake and move forward computer brain over wireless broadband. As bols “understands” Chinese. But this isn’t as a scientific field, we must recognize that the joke goes, even if AI requires an infinite really a problem, because nobody actually many intelligent processes in Nature perform loop, it should run in only five seconds on a “understands” Chinese! We only think we more powerfully than human symbolic rea- supercomputer. soning, even though they lack any of the The issue isn’t the speed of running a mind- mind-like mechanisms long believed neces- like program; it is the size and quality of the sary for human “competence.” Once we rec- program itself. Because we routinely underes- The scientific evidence ognize this and start to work out these timate the complexity of evolved biological scaleable representations and algorithms with- systems, and because Moore’s law doesn’t lead coming in all around us is clear: out anthropomorphizing them, we should be to a doubling of the quality of human-written able to produce the kind of results that will software,1 the same old goals are red herrings Symbolic conscious reasoning, get our work funded to the level necessary for that promise the practically impossible! growth and deliver beneficial applications to which is extracted through society, without promising the intelligent Take Mind off its pedestal English-speaking humanoid robot slaves and AI’s great mistake is its assumption that protocol analysis from serial soldiers of science fiction. human-level intelligence is the greatest intel- ligence that exists, and thus, that our com- Defining mindless intelligence putational intelligences should operate “like” verbal introspection, is a myth. I define “mindless intelligence” as intelli- human cognition. Because of this mistake, gent behavior ascribed (by an observer) to most AI research has focused on “cognitive any process lacking a mind-brain. Suppose models” of intelligence, on programs that run understand it. As anyone—even a native some black-box process (for example, math- like people think. But it turns out that we speaker—drives further down into an expla- ematical, numerical, or mechanical) exhibits don’t think the way we think we think! nation of his or her knowledge or behavior, behavior that appears to require intelligence. The scientific evidence coming in all instead of gaining sharper insights (as we However, when we scientifically study it, we around us is clear: Symbolic conscious rea- might expect in a reductionist physical sci- find no Lisp interpreter, no symbols, no soning, which is extracted through protocol ence with a better microscope), the explana- grammars, no logic or inference engine—in analysis from serial verbal introspection, is tions get blurrier and blurrier. fact, we realize that it works without any of a myth. From Michael Gazzaniga’s famous By assuming that intelligences based on the accoutrements of cognition. We can say split-brain experiments, where a patient asso- human-centric cognitive architectures such that this process is mindlessly intelligent. ciated a snow shovel with a chicken,2 through as grammars or production systems are the Now we can begin to seriously study intel- Daniel Dennett’s demolition of conscious- zenith, are the most powerful intelligences ligent performance by ness,3 through the unconscious intelligence in the world, our field has made the same described most recently by Malcolm Glad- kind of embarrassing mistake as today’s ¥ feedback-driven systems such as ther- well,4 it’s entirely clear that the “symbolic cryptocreationists, the proponents of Intelli- mostats and steam governors; mind” that AI has tried for 50 years to simu- gent Design. By doubting that a mindless ¥ pattern-action systems such as Eliza pro- late is just a story we humans tell ourselves nonlinear iterative process such as evolution grams and immune systems; to predict and explain the unimaginably com- could be responsible for irreducible com- ¥ stability and hierarchy networks such as plex processes occurring in our evolved plexity in the designs of biological life- cellular metabolisms; brains. forms, they hold that a superhuman, super- ¥ societal assemblies such as insect and Because of this preoccupation with mim- intelligent being must have intervened. colonial life-forms; icking human-level intelligence, as a scien- AI also behaves as if human intelligence ¥ utility-maximizing systems such as game tific field, we’ve ignored or excluded the con- is next to godliness. Even the neural and economic agents; tributions of many alternative nonsymbolic approach, more accepted today then ever, ¥ exquisitely iterative systems such as evolu- mechanisms. Such mechanisms range from falls into the trap of trying to model human tion, fractals, and embryogenesis; and even

MAY/JUNE 2006 www.computer.org/intelligent 51 The Future of AI

(a) (b) (c)

Figure 1. Three generations of evolved robots: (a) Pablo Funes’ evolution of Lego discovered the cantilever,14 (b) Hod Lipson’s evolution of dynamic trusses invented the ratchet,15 and (c) Gregory Hornby’s evolution using L-systems to describe machines invented a kayaking motion.16

¥ mind-erasing collectives such as academic surpass—in complexity and reliability— Ectomental learning committees, crowds, and bureaucracies. anything architects, engineers, novelists, ven- One of the oldest AI paradigms is a self- ture capitalists, or teams of software pro- learning or autodidactic system, a program To give you a broader sense of the field, grammers can achieve. that begins with a tabula rasa and, when I’ll briefly cover several kinds of natural Human teams can build systems with only dropped into an environment, gets better and processes that appear intelligent yet lack 10 million to 100 million unique moving better over time. Perhaps the best example any cognitive apparatus. John Kolen and I parts before the entire structure collapses, yet of such a system is Gerald Tesauro’s TD- showed how an iterated dynamical system biological forms can have 10 billion unique Gammon.17 He started with essentially a ran- could appear to generate a context-free or moving parts. dom neural network that could return a value context-sensitive language, depending on the For the past decade, my lab’s goal has for any backgammon position. Rather than observer.6 The dynamical system lacked any been to understand how evolution can pro- training the network against an encyclope- cognitive architecture for “generative capac- duce more complex designs than a human dia of human expert games, he essentially ity,” which has been assumed by all natural engineering team, while lacking human- trained it against itself. After about a month language processing systems since Noam level symbolic cognition. We’ve focused of computer time on an IBM supercomputer, Chomsky. specifically on coevolutionary machine with the weights adjusted as a result of each Wherever we look in Nature, we see amaz- learning systems. While we haven’t yet game, his network, with further refinements, ingly complex processes to which we can achieved a fully open-ended design process, became one of the best players in the world. ascribe intelligence, yet we observe symbolic we have Humans can verbalize backgammon strat- cognition in only one place, and only there as egies. We consider only a few plausible a result of introspection. Many of these nat- ¥ shown coevolutionary systems that have sur- moves and then estimate whether one move ural processes have been studied under the passed human performance in sorting net- is better or worse than another on the basis aegis of complex systems or have been given works and cellular-automata optimization;7 of strategic goals from models of the game the prefix “self” or “auto.” Because these sys- ¥ developed theories such as Pareto coevo- (running, blocking, back-game), using all tems have no mind, and thus no self, I’ve lution,8 emergent dimensionality,9 and kinds of approximate and exact calculations taken the liberty of replacing those prefixes computational models of symbiogene- about probability. I was a professional-level with the new term ectomental, which means sis;10 and backgammon player in 1975 and felt that outside (Greek) of mind (Latin). ¥ revealed the possibility of motivating a there were about seven different human- community of learners11 to become their player “types” who, at the top of their game, Ectomental organization own Ideal Teachers,12 resulting in novel achieved a rock-scissors-paper parity. Evolution is the primary example of educational software.13 On the other hand, TD-Gammon is a an intelligent designer who lacks a mind. mindless intelligence that dominates all There’s no grammar, set of rules, library of Perhaps our best-known research is on the human players. It uses a function to estimate CAD parts, or physics simulation. Simply coevolution of robot bodies and brains, values and uses a one- or two-ply look-ahead put: a mindless reproductive system operates, known as the Genetically Organized Lifelike with a greedy selector to make a move. It has transcription errors occur, and selection locks Electro-Mechanics, or GOLEM, project. This no logic or symbols, no strategy that looks in a statistical advantage for the marginally research resulted in three generations of self- far ahead or back in time, and no language better—or luckier—members of a popula- designed systems that discovered irreducibly component to discuss its strategy. Yet it’s tion. And yet this iterative process has auto- complex components and processes such as stronger than any rule-based strategy. matically designed machines of incredible the cantilever, ratcheting, and kayaking (see My lab had worked on self-learning for beauty and complexity, objects that far figure 1). tic-tac-toe,18 and we became interested in

52 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS (a) (b)

Figure 2. (a) An iterated-function-systems fractal is like a feedback loop on a copy machine that makes more than one reduced copy of an image, resulting in the same limit for a speck of dust or a full page of ink. (b) The IFS theory explained the “strange automata” that emerged when recurrent neural networks were trained to recognize languages. understanding why TD-Gammon worked. man’s earlier mathematical work. mindless actions can’t help but keep return- We were able to replicate the Tesauro effect Perhaps many mindlessly intelligent pro- ing to? In other words, the answer to self- using simple hill-climbing,19 which led to the cesses in Nature are similar instances of repair is that there’s no blueprint or explicit question of why coevolutionary self-learn- mathematical ideals that can lead to conver- diagram; there’s just a framework and a set of ing worked so well for backgammon. Game gence, complexity, and optimal performance parameters that mathematically define a theorists such as Richard Bellman recog- in the limit. complex attractor. Mindless and far-flung nized many years ago why a purely numeric distributed operations can’t help themselves; backgammon player works better than a log- Ectomental repair they must gravitate toward it. ical game.20 He proved the existence of a A marvelous characteristic of natural sys- Such dynamical systems with complex value table for optimal sequential choice in tems is that they can heal, or self-repair. A attractors driven by parameters are well Markovian games, where opponents can naïve computerized view would be to envi- known. One example is the Mandelbrot set, choose strategies yet are buffeted by random sion the algorithmic equivalent of a team of a truly exquisite iteration where the parame- elements such as dice. Moreover, iterated repairpersons who, under centralized super- ters define a window and each pixel com- approximation of the value table, through a vision, consult a system model and are then putes its own color. Another example is iter- single-ply expectimax look-ahead, leads to deployed to a disturbance’s site to apply cog- ated function systems, a union of a set of its convergence. So, an optimal value table nition, logic, and spare parts to return the sys- contractive maps that Michael Barnsley combined with a one-ply greedy choice leads tem to model behavior. However, imagining proved has a single fractal limit attractor akin to the strongest-possible player. a system that contains a deployable model of to Cantor dust.24 In order to study the success of learning itself can lead to logical conundrums.22 Barnsley showed, much analogous to Bell- backgammon, I recently invented Nannon¨, How might we understand self-repair in man’s proof, that some nonlinear iterative the smallest version of backgammon that natural systems? In artificial-life research on processes, despite having many adjustable maintains its core behaviors, yet only uses six “algorithmic chemistry,” Walter Fontana and parameters, have a single, yet complicated, points, three checkers, and one die per side. Leo Buss described systems of simple limit, defined by the interaction of the para- There are only 2,530 different board positions, lambda calculus programs that consume and meters and rules. Simply put, an IFS fractal and the value table converges in 15 sweeps to produce each other, forming a metabolism.23 attractor is like repeatedly copying an image an error of 107.21 While the full game of back- When such an artificial-chemistry network with a special copying machine that makes gammon is much larger than Nannon, so a had a steady-state dynamic, perturbations multiple shrunken and transformed copies of table can’t be stored, Tesauro’s choice of input would return to the same attractor, like the the input page (see figure 2). All nonblank representation and network size from earlier memories in a Hopfield network. starting pages, from a speck of dust to a piece experiments led to a fortuitous convergence Is the Bauplan of an animal a similar of black construction paper, end up converg- between TD reinforcement learning and Bell- attractor, which the myriad of microscopic ing to the same attractor in the limit.

MAY/JUNE 2006 www.computer.org/intelligent 53 The Future of AI

I came across IFSs while working to interesting threads is the relationship be- search for more complex reproductive forms understand the relationship between recur- tween robotic assembly with errors and holds some hope for understanding how a rent neural networks and finite-state ma- noise, and the kinds of tasks that Bellman mindless reproductive process can become chines. As the result of trying to learn a lan- proved could iteratively converge to opti- more capable over time to sustain complexity guage, a recurrent network generated an mal.31 This might provide a self-construction in the design of reproducible machinery. infinite-state machine with the states located theory involving not a blind watchmaker but on a fractal attractor.25 Subsequent research a blind chessmaster who continuously opti- Ectomental recognition, control, used these structures for memory and hier- mizes assembly processes to maximize its and regulation archal representations.26 own chances for successful reproduction. Obviously, intelligence arises outside the The mindless intelligence of self-defining mental sphere in so many other places in and self-repairing, or autopoetic,27 biologi- Ectomental reproduction nature that I can’t list them all. cal forms is a big leap from Fontana’s Another great mystery of Nature is com- The immune system is an ectomental chemistries and Barnsley’s fractals. Yet I am plex self-reproduction. Shy of a magical chemical recognition system that filters and certain that biological form will one day be reverse-engineering theory (which would let separates millions of chemicals along the scientifically explained as an attractor that us genetically engineer flying horses), we me/not-me boundary, without a central data- changes its parameters over time while it’s have little or no grasp on the algorithmic base listing which compounds are in or out. constantly and mindlessly repaired by dis- processes involved in the major transition Self-control of physical movement, of tributed processing at a microscopic level. individuals and groups, is often mindless. This isn’t only because time constraints push Ectomental assembly nervous-system controls to the edge but also Fetal development, or embryogenesis, is Fetal development, or because it’s hard to find a valuable use for perhaps the perfect place to recognize the cognitive symbols inside mainly numeric profound scale of complex behavior achiev- embryogenesis, is perhaps the models such as pattern generators and feed- able by mindless intelligence. back loops. Herb Simon introduced Tempus and Hora perfect place to recognize the Finally, the zenith of self-regulation is as two different kinds of watchmakers who probably the planet itself. Similar to Adam suffer from interruptions: one uses modular profound scale of complex Smith’s “invisible hand” idea that markets are construction; the other works with basic mindlessly intelligent regulators and alloca- parts.28 Richard Dawkins introduced the idea behavior achievable by tors of goods and services, the Gaia hypoth- of the Blind Watchmaker.29 Both researchers esis proposes that the whole biosphere oper- comfortably anthropomorphized what is a ates so as to maintain the right conditions for mindless assembly process. mindless intelligence. life as we know it.36 A trivial and kooky inter- Every assembly factory depends critically pretation is that Gaia is a Goddess with a on human minds both as labor as well as mind of her own, complete with symbols, supervision to monitor, correct, and repair from single cells reproducing, through colo- logic, and language, so she might talk to us ongoing processes. Yet a developing fertil- nialization, to multicellular creatures with one day through a burning bush or a statue of ized egg is also an assembly factory, without differentiated tissues and functions. her likeness. A deeper interpretation is to rec- any human supervisors or any brain, that pro- I think it’s another case of dramatically ognize that the algorithmic complexity of bal- duces an exquisite, custom product with 10 underestimating the amount of intelligence ancing resources, encouraging growth, and billion moving parts in only nine months! in a seemingly obvious natural process. We managing the network of species to maintain Where’s the mind inside the fertilized egg? have many simple examples of reproduction the “sweet spot” for life is a huge job requir- Even Intelligent Design proponents might be in software, from straight data copying to ing such intelligence that we better not entrust hard pressed to defend the existence of an self-reproducing code as shown by evaluat- it to any elected human officials! omniscient “Intelligent Factory Foreman” ing this ditty in Common Lisp: Under the mindless-intelligence viewpoint, who supervises every embryo developing in both evolution itself and the global-regulation the world simultaneously, deciding which ((LAMBDA (X) (LIST X (LIST ‘QUOTE X))) system known as Gaia are intelligent beyond creatures live or die. ‘(LAMBDA (X) (LIST X (LIST ‘QUOTE X)))) and outside the mental framework based on Other than basic work on pattern forma- the symbol manipulation that AI has chosen tion, related to work by, for example, Alan Following John Von Neumann’s challenge as its first 50-year focus. Turing and Stephen Wolfram, we have a long of finding self-reproduction in cellular au- way to go in understanding the mindless tomata, Christopher Langton helped birth the intelligence in a process that could self- field of artificial life with his more elegant assemble into a biological form with billions automata,32 and Jason Lohn and James Reg- ’m neither alone nor unique in wishing for of parts. My lab is working on replacing the gia showed how easy it is to discover the rules Ia stronger scientific basis for the field. idea of a perfect robotic factory with evolu- for such automata.33 Yet so far, all our com- These comments certainly hearken back to tionary processes that must evolve both form puting reproducing systems, including Tom many earlier calls.37 Much of the world has and formation and overcome noise and error Ray’s Tierra34 and Hod Lipson’s cubes,35 are changed in the last decade. For example, after in physical assembly.30 One of the more very simple. I’m hopeful that evolutionary so many years of chasing generative linguis-

54 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS tics’ focus on parsing and syntax, the main digestion of the food in the Chinese Room. 3. D. Dennett, Consciousness Explained, Little, thrust of both natural language processing and I don’t mean to imply that human cogni- Brown and Co., 1991. speech recognition has been to drive mind- tion isn’t worth studying. I just want to re- 4. M. Gladwell, Blink: The Power of Thinking less statistical responses from large corpora iterate that cognitive reporting is an always- without Thinking, Little, Brown and Co., 2005. rather than to establish carefully wrought incomplete story, a simplified verbalization rules and features. Intelligent-control research of a partial insight of the working patterns of 5. J.R. Searle, “Minds, Brains, and Programs,” Behavioral and Brain Sciences, vol. 3, no. 3, is also moving in a mindless way, from robot- our brains. And brains aren’t instruction set 1980, pp. 417Ð457. ics that use shaky logical algorithms to more computers; they’re complicated biological mathematically sophisticated nonlinear con- networks with all kinds of feedback at all lev- 6. J.F. Kolen and J.B. Pollack, “The Observers’ trol systems.38 Much cognitive-modeling els, like metabolisms, gene regulatory net- Paradox: Apparent Computational Complex- ity in Physical Systems,” J. Experimental and research takes seriously the idea that algo- works, and immune systems. The software Theoretical Artificial Intelligence, vol. 7, no. rithms should be not only cognitively plausible and systems that emerge from and control 3, 1995, pp. 253Ð269. but also neurally plausible. Finally, machine these networks, like evolution, embryologi- learning research has progressed from its early cal-development protocols, Gaian ecologi- 7. H. Juille and J.B. Pollack, “Coevolutionary Learning and the Design of Complex Sys- efforts at matching human learning curves, to cal regulation, or mind, will be much harder tems,” Advances in Complex Systems, vol. 2, building strong algorithms for extracting to reverse-engineer than the artifacts of no. 4, 1999, pp. 371Ð393. knowledge from large statistical sources. human engineering culture. Yet these fields often must defend them- 8. S.G. Ficici and J.B. Pollack, “Pareto Optimal- ity in Coevolutionary Learning,” Advances in selves from the charge that they aren’t really Artificial Life: 6th European Conf. (ECAL AI. George Dyson recently visited Google 2001), LNAI 2159, Springer, 2001, pp. and wrote that he has long considered that Symbolic Mind is a self-aggrandized 316Ð325. when “real” AI arrives on the scene, it will fiction told to make sense of a few 9. A. Bucci and J.B. Pollack, “Focusing versus be surrounded by “a circle of cheerful, con- Intransitivity: Geometrical Aspects of Co- tented, intellectually and physically well- evolution,” Proc. 2003 Genetic and Evolu- nourished people.”39 Certainly Google is pounds of mindlessly intelligent tionary Computation Conf., LNCS 2723, based on a very large database and uses Springer, 2003, pp. 250Ð261. statistical machine learning techniques to meat. It’s time we wean ourselves 10. R.A. Watson and J.B. Pollack, “A Computa- choose which keywords are important in dif- tional Model of Symbiotic Composition in ferent contexts. Does Google software have from the fiction and start working Evolutionary Transitions,” Biosystems, vol. any of the cognitive aspects that AI has stud- 69, nos. 2Ð3, 2002, pp. 187Ð209. ied for many years? The mindless market on the science. 11. E. Sklar and J.B. Pollack, “A Framework for doesn’t care. Enabling an Internet Learning Community,” As we’ve seen, mindless intelligence J. Educational Technology & Society, vol. 3, abounds in Nature, through processes that no. 3, 2000, pp. 393Ð408. channel mathematical ideals into physical Emphatically then, as AI arises, it won’t 12. E.D. De Jong and J.B. Pollack, “Ideal Evalu- processes that can appear optimally designed be organized like a good computer program, ation from Coevolution,” Evolutionary Com- yet that arise through and operate via exquis- it won’t speak English, and it certainly won’t putation, vol. 12, no. 2, 2004, pp. 159Ð192. ite iteration. act like a humanoid robot from a science fic- The hypothesis for how intelligence arises tion movie. Symbolic Mind is a self-aggran- 13. A. Bader-Natal and J.B. Pollack, “Motivating Appropriate Challenges in a Reciprocal Tutor- in Nature is that dynamical processes, driven dized fiction told to make sense of a few ing System,” Proc. AI in Education 2005 by accumulated data gathered through iter- pounds of mindlessly intelligent meat. It’s (AIED 05), IOS Press, 2005, pp. 49Ð56. ated and often random-seeming processes, time we wean ourselves from the fiction and can become more intelligent than a smart start working on the science. 14. P. Funes and J.B. Pollack, “Evolutionary Body Building: Adaptive Physical Designs adult human, yet continue to operate on prin- for Robots,” Artificial Life, vol. 4, no. 4, 1998, ciples that don’t rely on symbols and logical pp. 337Ð357. reasoning. The proof lies not only in Mar- Acknowledgments kovian situations where a greedy sequential- Thanks to my PhD students, past and present, 15. H. Lipson and J.B. Pollack, “Automatic Design and Manufacture of Robotic Life- choice algorithm driven by values converged for their collaborations, to the editors for their encouragement in finishing this manuscript, and forms,” Nature, vol. 406, no. 6799, 2000, pp. under Bellman’s equation, but also in the reli- to Carl Feynman for help coining a word. 974Ð978. ability, complexity, and low cost of biologi- cally produced machines. 16. G.S. Hornby and J.B. Pollack, “Creating High-Level Components with Generative Because our minds aren’t what they seem, References Representation for Body-Brain Evolution,” symbolic explanations of our behavior that Artificial Life, vol. 8, no. 3, 2002, pp. 223Ð were extracted from protocol analysis and con- 1. J. Lanier, “One Half of a Manifesto,” Wired, 246. scious introspection are misleading at best and Dec. 2000; www.wired.com/wired/archive/ 8.12/lanier.html. 17. G. Tesauro, “Temporal Difference Learning complete fabrications at worst. Most of what of Backgammon Strategy,” Proc. Int’l Conf. our brains are doing involves mindless chem- 2. M.S. Gazzinga, The Social Brain: Discovering Machine Learning (ICML 92), Morgan Kauf- ical activity not even distinguishable from the Networks of the Mind, Basic Books, 1985. mann, 1992, pp. 451Ð457.

MAY/JUNE 2006 www.computer.org/intelligent 55 The Future of AI

Organization,” Bull. Mathematical Biology, vol. 56, 1994, pp. 1Ð64. The Author 24. M.F. Barnsley, Fractals Everywhere, Acade- mic Press, 1988. Jordan B. Pollack is a professor of computer science and complex systems at Brandeis University and the director of the DEMO (Dynamical & Evolu- 25. J.B. Pollack, “The Induction of Dynamical tionary Machine Organization) laboratory. He has broad interests and publi- Recognizers,” Machine Learning, vol. 7, nos. cations across AI, artificial life, neural and evolutionary computing, robotics, 2Ð3, 1991, pp. 227Ð252. complex and dynamic systems, educational technology, and intellectual-prop- erty law. He has been involved in startup companies including Abuzz, 26. S. Levy, O. Melnik, and J.B. Pollack, “Infi- Affinnova, and Thinmail. His laboratory's work on automatically designed and nite RAAM: A Principled Connectionist manufactured robots made front-page news worldwide in 2000, and he was Basis for Grammatical Competence,” Proc. named one of MIT's Technology Review “TR 10” in 2001. He received his 22nd Ann. Conf. Cognitive Science Soc., Cog- PhD in computer science from the University of Illinois. Contact him at the Computer Science Dept., nitive Science Soc., 2000, pp. 298Ð303; www. Brandeis Univ., 415 South St., Waltham, MA 02454; [email protected]; www.jordanpollack.com. cogsci.rpi.edu/CSJarchive/Proceedings/ 2000/COGSCI00.pdf.

27. F.J. Varela, H.R. Maturana, and R. Uribe, 18. P.J. Angeline and J.B. Pollack, “Competitive 21. J.B Pollack, “Nannon: A Nano Backgammon “Autopoiesis: The Organization of Living Environments Evolve Better Solutions to for Machine Learning Research,” Proc. 2005 Systems, Its Characterization and a Model,” Complex Problems,” Proc. 5th Int’l Conf. Int’l Conf. Computational Intelligence in BioSystems, vol. 5, no. 4, 1974, pp. 187Ð196. Genetic Algorithms (ICGA 93), Morgan Games, 2005, pp. 277Ð284; http://csapps. Kaufmann, 1993, pp. 264Ð270. essex.ac.uk/cig/2005/papers/p1055.pdf. 28. H.A. Simon, The Sciences of the Artificial, 19. J.B. Pollack and A. Blair, “Co-Evolution in 22. M. Minsky, “Matter, Mind, and Models,” MIT Press, 1969. the Successful Learning of Backgammon Proc. IFIP Congress, 1965, pp. 45Ð49; http:// Strategy,” Machine Learning, vol. 32, no. 3, groups.csail.mit.edu/medg/people/doyle/gallery/ 29. R. Dawkins, The Blind Watchmaker: Why the 1998, pp. 225Ð240. minsky/mmm.html. Evidence of Evolution Reveals a Universe without Design, W.W. Norton and Co., 1986. 20. R.E. Bellman, Dynamic Programming, 23. W. Fontana and L.W. Buss, “‘The Arrival of Princeton Univ. Press, 1957. the Fittest’: Toward a Theory of Biological 30. J. Rieffel and J.B. Pollack, “Evolving Assem- bly Plans for Fully Automated Design and Assembly,” Proc. NASA/DoD Conf. Evolv- able Hardware (EH 05), IEEE Press, 2005, pp. 165Ð170.

31. S. Viswanathan and J.B. Pollack “On the Robustness Achievable with Stochastic See the Future of Development Processes,” Proc. NASA/DoD Conf. Evolvable Hardware (EH 05), IEEE ComputingNow Press, 2005, pp. 34Ð39. 32. C.G. Langton, “Self-Reproduction in Cellu- lar Automata,” Physica D, vol. 10, nos. 1Ð2, in IEEEIntelligent Systems 1984, pp. 135Ð144. 33. J.D. Lohn and J.A. Reggia, “Automatic Dis- covery of Self-Replicating Structures in Cel- lular Automata,” IEEE Trans. Evolutionary Tomorrow's PCs, handhelds, and Computation, vol. 1, no. 3, 1997, pp. 165Ð Internet will use technology that exploits 178. current research in artificial intelligence. 34. T.S. Ray, “An Approach to the Synthesis of Life,” Artificial Life II, C. Langton et al., eds., Breakthroughs in areas such as intelligent Addison-Wesley, 1991, pp. 371Ð408. agents, the 35. V. Zykov et al., “Self-Reproducing Machines,” Semantic Web, Nature, vol. 435, no. 7038, 2005, pp. 163Ð164.

data mining, 36. J.E. Lovelock, Gaia: A New Look at Life on and natural language Earth, Oxford Univ. Press, 1979. processing will 37. R.A. Brooks, “Intelligence without Repre- revolutionize your work and sentation,” Artificial Intelligence, vol. 47, nos. 1Ð3, 1991, pp. 139Ð159. leisure activities. Read about 38. F. Zhao, “Extracting and Representing Qual- this research as it happens in itative Behaviors of Complex Systems in IEEE Intelligent Systems. Phase Spaces,” Artificial Intelligence, vol. 69, nos. 1Ð2, 1994, pp. 51Ð92.

39. G. Dyson, “Turing’s Cathedral,” Edge, 24 Oct.

SUBSCRIBE NOW! http://computer.org/intelligent IEEE 2005, www.edge.org/3rd_culture/dyson05/ dyson05_index.html.

56 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS