Emergence

by Steven Johnson

  Indeed, the adoption rate for these feedback devices is accelerating so rapidly that I suspect in a matter of years a Web page without a dynamic rating system attached will trigger the same response that a Web page without hyperlinks triggers today: yes, it’s technically possible to create a page without these features, but what’s the point? The Slashdot system might seem a little complex, a little esoteric for consumers who didn’t grow up playing D&D, but think of the millions of people who learned how to use a computer for the first time in the past few years, just to get e-mail or to surf the Web. Compared to that learning curve, figuring out the rules of Slashdot is a walk in the park.

  And rules they are. You can’t think of a system like the one Malda built at Slashdot as a purely representational entity, the way you think about a book or a movie. It is partly representational, of course: you read messages via the Slashdot platform, and so the components of the textual medium that Marshall McLuhan so brilliantly documented in The Gutenberg Galaxy are on display at Slashdot as well. Because you are reading words, your reception of the information behind those words differs from what it would have been had that information been conveyed via television. The medium is still the message on Slashdot—it’s just that there’s another level to the experience, a level that our critical vocabularies are only now finding words for.

  In a Slashdot-style system, there is a medium, a message, and an audience. So far, no different from television. The difference is that those elements exist alongside a set of rules that govern the way the messages flow through the system. “Interactivity” doesn’t do justice to the significance of this shift. A button that lets you e-mail a response to a published author; a tool that lets you build your own home page; even a collection of interlinked pages that let you follow your own path through them—these are all examples of interactivity, but they’re in a different category from the self-organizing systems of eBay or Slashdot. Links and home-page-building tools are cool, no question. But they are closer to a newspaper letters-to-the-editor page than Slashdot’s collective intelligence.

  First-generation interactivity may have given the consumer a voice, but systems like Slashdot force us to accept a more radical proposition: to understand how these new media experiences work, you have to analyze the message, the medium, and the rules. Think of those thousand-post geek-Dionysian frenzies transformed into an informative, concise briefing via the Slashdot quality filters. What’s interesting here is not just the medium, but rather the rules that govern what gets selected and what doesn’t. It’s an algorithmic problem, then, and not a representational one. It is the difference between playing a game of Monopoly and hanging a Monopoly board on your wall. There are representational forces unleashed by a game of Monopoly (you have to be able to make out the color coding of the various properties and to count your money) but what makes the game interesting—indeed, what makes it a game at all—lies in the instruction set that you follow while playing. Slashdot’s rules are what make the medium interesting—so interesting, in fact, that you can’t help thinking they need their own category, beyond message and medium.

  Generically, you can describe those rules as a mix of positive and negative feedback pushing the system toward a particular state based on the activities of the participants. But the mix is different every time. The edge cities of Paul Krugman’s model used feedback to create polycentric clusters, while other metropolitan systems collapse into a single, dense urban core. The networks in CNN-era television have engendered runaway positive feedback loops such as the Gennifer Flowers story, while a system like Slashdot achieves homeostatic balance, at least when viewed at level 5. Different feedback systems produce different results—even when those systems share the same underlying medium. In the future, every Web site may well be connected to a rating mechanism, but that doesn’t mean all Web sites will behave the same way. There may be homeostasis at Slashdot’s level 5, but you can always choose to read the unfiltered, anarchic version at level -1.

  Is there a danger in moving to a world where all our media responds directly to user feedback? Some critics, such as The Control Revolution’s Andrew Shapiro, worry about the tyranny of excessive user personalization, as in the old Nicholas Negroponte vision of the Daily Me, the newspaper perfectly custom-tailored to your interests—so custom-tailored, in fact, that you lose the serendipity and surprise that we’ve come to expect from reading the newspaper. There’s no stumbling across a different point of view, or happening upon an interesting new field you knew nothing about—the Daily Me simply feeds back what you’ve instructed the software to find, and nothing more. It’s a mind-narrowing experience, not a mind-expanding one. That level of personalization may well be around the corner, and we’ll take a closer look at its implications in the conclusion. But for now, it’s worth pointing out that the Slashdot system is indifferent to your personal interests—other than your interest in a general level of quality. The “ideal state” that the Slashdot system homes in on is not defined by an individual’s perspective; it is defined by the overall group’s perspective. The collective decides what’s quality and what’s crap, to use Rob Malda’s language. You can tweak the quality-to-crap ratio based on your individual predilections, but the ratings themselves emerge through the actions of the community at large. It’s more groupthink than Daily Me.

  Perhaps, then, the danger lies in too much groupthink. Malda designed his system to evaluate submissions based on the average Slashdot reader—although the karma points tend to select moderators who have a higher-than-average reputation within the community. It’s entirely possible that Malda’s rules have created a tyranny of the majority at Slashdot, at least when viewed at level 5. Posts that resonate with the “average” Slashdotter are more likely to rise to the top, while posts that express a minority viewpoint may be demoted in the system. (Technically, the moderation guidelines suggest that users should rate posts based purely on quality, not on whether they agree with the posts, but the line is invariably a slippery one.) From this angle, then, Slashdot bears a surprising resemblance to the old top-down universe of pre-cable network television. Both systems have a heavy center that pulls content toward the interests of the “average user”—like a planet pulling satellites into its orbit. In the days before cable fragmentation, the big three networks were competing for the entire television-owning audience, which encouraged them to serve up programming designed for the average viewer rather than for a particular niche. (McLuhan observed how this phenomenon was pushing the political parties toward the center as well.) The network decision to pursue the center rather than the peripheries was invariably made at the executive level, of course—unlike at Slashdot, where the centrism comes from below. But if you’re worried about suppressing diversity, it doesn’t really matter whether it comes from above or below. The results are the same, either way. Majority viewpoints get amplified, while minority viewpoints get silenced.

  This critique showcases why we need a third term beyond medium and message. While it’s true that Slashdot’s filtering software creates a heavy center, that tendency is not inherent to the Web medium, or even the subset of online communities. You could just as easily build a system that would promote both quality and diversity, simply by tweaking the algorithm that selects moderators. Change a single variable in the mix, and a dramatically different system emerges. Instead of picking moderators based on the average rating of their posts, the new system picks moderators whose contributions have triggered the greatest range of responses. In this system, a member who was consistently rated highly by the community would be unlikely to be chosen as a moderator, while a member who inspired strong responses either way—both positive and negative—would be first in line to moderate. The system would reward controversial voices rather than popular ones. You’d still have moderators deleting useless spam and flamebait, and so the quality filters would remain in place. But the fringe voices in the community would have a stronger presence at level 5, because the feedback system wo
uld be rewarding perspectives that deviate from the mainstream, that don’t aim to please everyone all the time. The cranks would still be marginalized, assuming their polemics annoyed almost everyone who came across them. But the thoughtful minorities—the ones who attract both admirers and detractors—would have a place at the table.

  There’s no reason why centrist Slashdot and diverse Slashdot can’t coexist. If you can adjust the quality filters on the fly, you could just as easily adjust the diversity filters. You could design the system to track the ratings of both popular and controversial moderators; users would then be able to view Slashdot through the lens of the “average” user on one day, and through the lens of a more diverse audience the next. The medium and the message remain the same; only the rules change from one system to the other. Adjust the feedback loops, and a new type of community appears on the screen. One setting gives you Gennifer Flowers and cyclone-style feeding frenzies, another gives you the shapeless datasmog of Usenet. One setting gives you an orderly, centrist community strong on shared values, another gives you a multiculturalist’s fantasy. As Wiener recognized a half century ago, feedback systems come in all shapes and sizes. When we come across a system that doesn’t work well, there’s no point in denouncing the use of feedback itself. Better to figure out the specific rules of the system at hand and start thinking of ways to wire it so that the feedback routines promote the values we want promoted. It’s the old sixties slogan transposed into the digital age: if you don’t like the way things work today, change the system.

  5

  Control Artist

  On the screen, the pixels dance: bright red dots with faint trails of green, scurrying across a black background, like fireflies set against the sky of a summer night. For a few seconds, the movement on-screen looks utterly random: pixels darting back and forth, colliding, and moving on. And then suddenly a small pocket of red dots gather together in a pulsing, erratic circle, ringed by a strip of green. The circle grows as more red pixels collide with it; the green belt expands. Seconds later, another lopsided circle appears in the corner of the screen, followed by three more. The circles are unlike any geometric shape you’ve ever seen. They seem more like a life-form—a digital blob—pulsing haphazardly, swelling and contracting. Two blobs slowly creep toward each other, then merge, forming a single unit. After a few minutes, seven large blobs dominate, with only a few remaining free-floating red pixels ambling across the screen.

  Welcome to the world of Mitch Resnick’s tool for visualizing self-organizing systems, StarLogo. A descendant of Seymour Papert’s legendary turtle-based programming language, Logo, StarLogo allows you to model emergent behavior using simple, English-like commands—and it displays that behavior in vivid, realtime animations. If decentralized systems can sometimes seem counterintuitive or abstract, difficult to describe in words, StarLogo makes them come to life with dynamic graphics that are uniquely suited for the Nintendo generation. If a calendar is a tool for helping us think about the flow of time, and a pie chart is a tool for thinking about statistical distributions, StarLogo is a tool for thinking about bottom-up systems. And, in fact, those lifelike blobs on the screen take us back to the very beginnings of our story: they are digital slime molds, cells aggregating into larger clusters without any “pacemaker” cell leading the way.

  “Those red pixels are the individual slime mold cells,” Resnick says, pointing at the screen, sitting in his Cambridge office. “They’re programmed to wander aimlessly around the screen space, and as they wander, they ‘emit’ the green color, which quickly fades away. That color is the equivalent of the c-AMP chemical that the molds use to coordinate their behavior. I’ve programmed the red cells to ‘sniff’ the green color and follow the gradient in the color. ‘Smelling’ the green pixels leads the cells toward each other.”

  Like Gordon’s ant colonies, Resnick’s slime mold simulation is sensitive to population density. “Let’s start with only a hundred slime mold cells,” he says, adjusting a slider on the screen that alters the number of cells in the simulation. He presses a start button, and a hundred red pixels begin their frenetic dance—only this time, no clusters appear. There are momentary flashes of green as a few cells collide, but no larger shapes emerge at all.

  “With a hundred cells, there isn’t enough contact for the aggregates to form. But triple the population like so,” he says, pulling the slider farther to the right, “and you increase the contact between cells. At three hundred cells, you’ll usually get one cluster after a few minutes, and sometimes two.” We wait for thirty seconds or so, and after a few false starts, a cluster takes shape near the center of the screen. “Once they come together, the slime molds are extremely difficult to break apart, even though they can be very fickle about aggregating in the first place.”

  Resnick then triples the population and starts the simulation over again. It’s a completely different system this time around: there’s a flash of red-celled activity, then almost immediately ten clusters form, nearly filling the screen with pulsing watermelon shapes. Only a handful of lonely red cells remain, drifting aimlessly between the clusters. More is very different. “The interesting thing is,” Resnick says with a chuckle, “you wouldn’t have necessarily predicted that behavior in advance, just from looking at the instructions. You might have said, the slime mold cells will all immediately form a giant cluster, or they’ll form clusters that keep breaking up. In fact, neither is the case, and the whole system turns out to be much more sensitive to initial conditions. At a hundred cells, there are no clusters at all; at three hundred, you’ll probably get one, but it’ll be pretty much permanent; and at nine hundred cells, you’ll immediately get ten clusters, but they’ll bounce around a little more.” But you couldn’t tell any of that just by looking at the original instruction set. You have to make it live before you can understand how it works.

  *

  StarLogo may look like a video game at first glance, but Resnick’s work is really more in the tradition of Friedrich Froebel, the German educator who invented kindergarten, and who spent much of his career in the early nineteenth century devising ingenious toys that would both amuse and entertain toddlers. “When Froebel designed the first kindergarten,” Resnick tells me, “he developed a set of toys they called Froebel’s gifts, and he carefully designed them with the assumption that the object he’d put in the hands of kids would make a big difference in what they learned and how they learned. We see the same thing carried through today. We see some of our new technology as the latter-day versions of Froebel’s gifts, trying to put new sorts of materials and new types of toys in the hands of kids that will change what they think about—and the way they think about it.”

  StarLogo, of course, is designed to help kids—and grown-ups, for that matter—think about a specific type of phenomenon, but it is by no means limited to slime molds. There are StarLogo programs that simulate ant foraging, forest fires, epidemics, traffic jams—even programs that generate more traditional Euclidean shapes using bottom-up techniques. (Resnick calls this “turtle geometry,” after the nickname used to describe the individual agents in a StarLogo program, a term that is itself borrowed from the original Logo language, which Papert designed to teach children about traditional programming techniques.) This knack for shape-shifting is one of the language’s great virtues. “StarLogo is a type of modeling environment where kids can build models of certain phenomena that they might observe in the world,” Resnick says. “Specifically, it enables them to build models of phenomena where lots of things interact with each other. So they might model cars on a highway, or they might model something like a bird flock, where the kids design behavior for lots of individual birds and then see the patterns that form through all the interactions.

  “One reason that we’re especially interested in building a tool like this is that these phenomena are common in the everyday world,” he continues. “We see bird flocks and traffic jams all of the time. On the other hand, people have a lot of
trouble understanding these types of phenomena. When people see a flock of birds, they assume the bird in the front is the leader and the others are just following. But that’s not the way the real birds form flocks. In fact, each bird just follows simple rules and they end up together as a group.”

  At its core, StarLogo is optimized for modeling emergent systems like the ones we’ve seen in the previous chapters, and so the building blocks for any StarLogo program are familiar ones: local interactions between large numbers of agents, governed by simple rules of mutual feedback. StarLogo is a kind of thinking prosthetic, a tool that lets the mind wrap itself around a concept that it’s not naturally equipped to grasp. We need StarLogo to help us understand emergent behavior for the same reason we need X-ray machines or calculators: our perceptual and cognitive faculties can’t do the work on their own.

  It’s a limitation that can be surprisingly hard to overcome. Consider the story that Resnick tells of artificial-intelligence guru Marvin Minsky encountering the slime mold simulation for the first time. “One day shortly after I developed the first working prototype of StarLogo, Minsky wandered into my office. On the computer screen he saw an early version of my StarLogo slime mold program. There were several green blobs on the screen (representing a chemical pheromone), with a cluster of turtles moving around inside each blob. A few turtles wandered randomly in the empty space between the blobs. Whenever one of these turtles wandered close enough to a blob, he joined the cluster of turtles inside.”

  Minsky scanned the screen for a few seconds, then asked Resnick what he was working on. “I explained that I was experimenting with some self-organizing systems. Minsky looked at the screen for a while, then said, ‘But those creatures aren’t self-organizing. They’re just moving toward the green food.’”