MIT professor Pattie Maes has created a stir by making agents a household word, Firefly a $100 million business, and Newsweek's "100 most important people to watch" list.
Pattie Maes sits on a coffee table in a small room in the MIT Media Lab called the Pond, talking about the sea. "That's me!" she exclaims, watching a video image of herself projected on a nearby TV. "I look like a monster."
It's true. The elegant and poised MIT professor resembles a blowfish in the footage. She's scuba diving in Hawaii, having slipped out of a computing conference to pet moray eels and feed sea turtles. Maes (pronounced "mahs") describes the different fish and decries the tackiness of the made-for-tourists video, which is lavished with Waikiki helicopter shots and Muzak. She seems excited showing off reef life and embarrassed talking about herself. She's also completely unfazed by the phalanx of industry reps making their way around the murky room looking at demos from the Things to Think With program while catching sneak peeks of the video. Demos are interesting, but, hey, there's a woman in a bathing suit flitting in their periphery.
Earlier, as she was making the five-minute walk between her offices at the Media Lab and those of her fledgling software company, Firefly Network Inc., Maes mused about the possibilities presented by cuttlefish. The elephantine cephalopods shift their coloring and markings in milliseconds, instantly changing from zebra stripes to blank slate to marbleized pebble skin. Squid too, she remarked, signal with similar speed and repertoire; there has to be some way to apply this molluscan talent to technology, some way rooms could sense mood and emit different colors, some way software could do the same thing.
It's pure Maes. One foot in an ecosystem, the other in an application - both feet firmly planted. Which, given her growing fame, would seem hard to maintain. The 36-year-old software-agent maven has a two-and-a-half-year-old company worth an estimated US$100 million, a new baby boy, new interests, and a newfound place on myriad lists: Newsweek's 100 most important people to watch, the World Economic Forum's 100 people to listen to, People's 50 most beautiful people, and the Association for Computing Machinery's 15 most perspicacious visionaries.
In the past two years, Maes and Firefly have done more for software agents than the semi-intelligent bits of software have ever done for us. Agents - small programs that do electronic tasks for their masters and that can, ideally, learn by watching their user's activities - have been dogged by hype for the past 20 years. The approach that AI researchers had generally used - the deliberative thinking paradigm - had not yielded serviceable autonomous agents, so the promise of the servile bots was never followed by the real thing.
Now, however, agents are finding their way into the world in large part because of Maes's pioneering work. Her radical approach flew in the face of traditional knowledge-based AI research. This new generation of agents has found a niche as people flail around in information overload and crave personalized Web experiences - and as advertisers seek to reach out and touch just you and just you and just you, not just a homogenized lowest common denominator.
Agents can work as technological machetes, clearing the lush data tangle of the Web. Some, such as Jango (née Bargain Finder and made by NetBot Inc.), scan through hundreds of Web sites and find the least expensive price for a toaster or a Toyota. Others, like Firefly, take stock of our buying and browsing habits and then offer purchase recommendations based on the behavior of like-minded people, while also facilitating the delivery of finely tuned ads.
But far from being physical alter egos or butlers - as they were originally described by Media Lab director Nicholas Negroponte in the 1970s, and as they have been envisioned ever since - agents are manifesting themselves invisibly, blending into the network gestalt.
Dehype, rebuild
In some ways, agents seem a perfect American solution: create a bot in your own image and go back to sleep on the couch. If you can define your interests and have your agent serve them up, you have the potential to simplify the world, narrowing it down to comfortable ideas and people who think like you.
But this is the antithesis of how Maes views agents. "Agents are a new way of thinking about software that is more proactive," she explains, her Belgian accent lingering. "Sometimes I envision an agent as having extra eyes, hands, or brains which are looking out for my interests. I am convinced that software will treat us in a more personalized way." This has already happened with Microsoft Office 97, for example. The system observes what the user does and then gives recommendations that are context sensitive.
Her thoughts are echoed by many in the field of agent development, including Alper Caglayan, president of Open Sesame in Cambridge, Massachusetts. "Agents by themselves as an entity will not happen, but there will be agents in the interface, so every application will be sort of agentized," says Caglayan, coauthor of the Agent Sourcebook.
In the seven years Maes has been at the Media Lab, she has managed to both dehype and rebuild the agent legacy. "Maes has played a leadership role in the development of the technology and for the cadre of experts, who are mostly her students," says Anthony Rutkowski, vice president of Internet business development for General Magic, a company developing mobile agents. "On the commercial side, she has captured the public imagination, and that is extremely important for a nascent technology."
Maes says she did it all because of her own information overload - and that it started with animals. "Chickens and rabbits and fish," recounts Maes's mom, Lieve Maes, over the phone from Belgium. "She bred rabbits. She always brought home sick cats and dogs and so on. It cost me so much money. I had to take them first to the vet, to sterilize the cats and dogs. She wanted to be a vet. Otherwise she wanted to have a kennel. It was only later that she became interested in computers." (The family of eight - with four daughters and two sons - showed great restraint and ate only one of Pattie's pets, an aged rooster too tough, in the end, to chew.)
Maes is not in love with technology and gadgets. "On the contrary, I wanted to affect people and the way they communicate and socialize," she says. Maes decided to get her bachelor's and then her PhD in computer science because technology factored into so many disciplines: business, architecture, science, art. She has always felt torn between her strong interests in these fields.
"I am not one of those people who lives for work," says Maes, who spends some of the free time she has going to art galleries. "I enjoy sculpting and photography and tennis and swimming. I simply do not have enough time. My life is not in control."
As we move on to the verdant Jungle, the autonomous agents suite at the Media Lab, Maes recounts the travels that brought her to MIT. During her third year at the University of Brussels, Maes became passionate about AI. Rodney Brooks of MIT's AI Lab heard her present a paper at a conference in 1987. "She touched on a whole range of issues that I had been struggling with in behavior-based AI - having it be a more dynamic, fluid system," Brooks remembers. "She had a way to make the system learn. It was very exciting. It was very new."
Brooks invited her to work with him and AI pioneer Marvin Minsky for a two-month stint back at MIT. On her last day, she was offered a visiting professorship. "I couldn't believe it," she remembers. "I couldn't say no because the opportunity was just too good, but I was worried because I looked up to all these people here and was thinking, 'How am I going to live up to this?'"
Brooks and Minsky were busily shaping the next AI wave. Instead of trying to model high-level functions like logic and program for every contingency, they would instead aim for interaction and learning. They set out to use this bottom-up approach, building robots, or "animats," that could perform simple tasks. Maes's job was to teach Brooks's six-legged insectlike robot Genghis - currently housed in the Smithsonian Institution - how to walk. It was tough, often frustrating work.
Next year, in 1990, she moved from the AI Lab to the Media Lab, where she set about sending bots skittering into electronic networks instead of across the Lab's floor. "I started all of this software-agents work because I realized that our computer environments and our lives are becoming incredibly complex," says Maes, who routinely forgets names and dates and places and locations of various materials in her overstuffed office. "Our brains were simply not designed for this chaos. We need help dealing with all the problems."
Intelligence is a social phenomenon
Maes posited that an agent could learn from its user's patterns, if they were repetitive enough. Her first bot did just that: it observed the way she scheduled her myriad meetings, and, over time, scheduled them itself. It learned, for instance, that Maes is not an early morning person - she never took a meeting before 9 a.m.
Her second agent, Maxims, did the same for email, but it took the burgeoning technology one step further. Maxims was a filter that "looked over the shoulder of the user," remembering all the pieces of mail deleted or read or forwarded and then prioritized them accordingly the next time around.
Working with a sharp and talented student by the name of Max Metral - he's now the 25-year-old chief technical officer of Firefly - Maes programmed the agents to learn from each other. When one user's agent encountered an email for which it didn't have a memory, it would communicate with other Maxims agents in the office, finding out whether a message from, say, Nicholas Negroponte was given a lot of attention, or just a little. "Intelligence is really a social phenomenon," Maes explains.
Collaborative filtering was born. Rather than having to be programmed for every possibility and every detail about a user's choices (as knowledge-based agents must), collaborative filtering agents fill the gaps in their knowledge by learning from their fellow agents.
Maes, along with Metral and some other students, next set about making agents that could apply collaborative filtering to compare the musical tastes of different users. If someone liked Suzanne Vega and Shawn Colvin, for instance, they could be linked to a user whose preferences included Vega, Colvin, and Natalie Merchant.
Why start with a musical agent? "All the projects we do are personally motivated," Maes happily admits. Ringo, as the prototype was dubbed, solved her music problem: she couldn't find an interesting mix of artists on Boston radio, nothing like the eclectic selection in Brussels.
What Maes liked especially was the possibility of collaborative filtering fostering community. "Maybe it's a bit cliché, but I think that women are more interested in building and maintaining communities," she says. Maes's own sense of community is strong: she visits her family in Belgium four times a year.
Interestingly, although her father was a physician and her mother a dentist, two of Maes's five siblings are computer scientists. Leaving her family was the hardest part of moving across the ocean and over to MIT. "I am the only one who is more than 20 minutes from my mother's house," she says. Maes works hard to keep her students and employees unified. "Everyone in the group is a lot stronger if we work with the same vision," she explains.
Maes also wanted her work to be used in the real world, not sitting listless in some rarefied academic sink tank. Through a chance meeting on an airplane between her student David Waxman and recent business-school grad Nicholas Grouf, she soon got her wish. Agents Inc. - a private company that would market collaborative-filtering and personalization technology - was founded in March 1995.
To date, the company has generated an impressive $18 million from venture capital sugar daddies, including Esther Dyson, Merrill Lynch, Softbank, Trident, and Dun & Bradstreet. One of the earliest financiers was Atlas Venture's Christopher Spray, a man who, coincidentally, had previously bankrolled Maes's sister Kathleen and her electronic publishing operation, SoftCore Inc., in Belgium.
The company built a Web site, rechristened the Ringo prototype Firefly, and created an area where visitors could set up a personal profile that described their musical interests (it also managed to include their ZIP code, age, and gender). They then could step back and just let the algorithm do its thing, hooking them up with music they might enjoy. They also had the option of chatting with people who showed similar musical tastes, a feature that made the site very popular, Maes explains. (And the demographic and psychographic information in the Firefly profile - which is called Passport - in turn allows the marketers to crisply target their advertisements to, say, a 20-year-old woman living in Manhattan's Upper West Side, or a 55-year-old man in Miami.) The company changed its name to Firefly Network Inc. in August 1996.
Privacy battles
While Maes's style is quiet and steadfast, her company's is anything but. Firefly's marketing strategy embodies its name: brilliant and flashy. This winter, the company will sell its Web site (source of fame and fortune, but still not bug-free) to Launch, a Santa Monica, California-based company that creates original music for the computer. Meanwhile, it has licensed the software to My Yahoo!, ZDNet, Barnes & Noble, America Online's Greenhouse Networks, and Reuters New Media. Visitors to these sites are issued a Passport that allows them to find books, articles, Web sites, or other information. The valuable demographic information held in these profiles - 3 million and counting - is managed by Firefly.
And then there's Firefly's coup de grâce: the Open Profiling Standard. OPS defines the technical specs for user profiles on the Web - such as Passport - as well as the rules governing the exchange of personal information. Users can either allow or not allow personal information to be shared with other companies or services. As Saul Klein - an energetic former ad exec who serves as Firefly's senior vice president of marketing - explains it, OPS "is about the individual getting control of their online experience and personalizing it."
On June 2, OPS (endorsed by Netscape and VeriSign) was submitted to the World Wide Web Consortium, which will review the work as part of a larger privacy effort called P3P, or the Platform for Privacy Preferences Project.
Microsoft has joined the alliance, agreeing to work on the same standard. A final decision on OPS could come as late as a year from now.
In the meantime, Firefly must stay sharp and focused as the field of agent suppliers becomes increasingly crowded. Although Minneapolis-based Net Perceptions is the company's most direct competitor - its GroupLens collaborative filtering software is the agent behind the scenes at the Amazon.com site - other agent makers abound, including Autonomy, AgentSoft, Open Sesame, BroadVision, Empirical Media, and Excite. Their variations come in several flavors. While GroupLens uses collaborative filtering, some others, such as Autonomy, use neural networks to compare the user's interests to rules embedded in the software.
Each design has advantages and drawbacks. "Collaborative filtering, aka personalization, works very well for the more defined things, like CDs or films," says Peter Dushkin, a technology analyst with New York-based Jupiter Communications. "But what isn't clear is how well the personalization idea works for less-defined items, like news stories. The neural nets have an advantage there because they can understand the type of content."
Open Sesame, for instance, positions itself as a company that offers the best of both possible worlds: collaborative filtering and neural networking. Dushkin says that future agents will combine many different approaches.
"We are likely to see more partnerships between search-engine firms and agent companies," he notes.
Finally, General Magic's mobile agents are a different genre altogether. Mobile agents do not stay at home, working out of your computer and talking to servers; rather, they sally forth, take up residence on different servers, and then return with the information you asked for, or after accomplishing a requested task, such as purchasing an airplane ticket. But mobile agents make Maes skittish. She believes the possibility for security problems is greater if an agent actually moves into, or resides in, a server - even if temporarily.
Ecommerce tryout
These days, Maes seems content to generally leave Firefly alone and spends only one day a week there. She still sits on the board and, according to her colleagues, plays a major role in brainstorming sessions. She maintains there is no conflict of interest between running her own business and generating research projects at MIT. "I try to keep them separate. Even though a lot of things we do at the Lab could be commercialized and could potentially fit within the product, Firefly is like every start-up: it must focus. The company is totally occupied with what it's already doing," says Maes. Firefly pays the Media Lab royalties and has to keep its paws off anything that comes out of the research - at least for the first two years, in which the sponsors get first licensing dibs.
One of the Lab projects that could have direct applications for Firefly is Yenta, an agent-based personal introduction service being developed by Leonard Foner, one of the research assistants in Maes's agents group. A user's Yenta talks to other Yentas and then picks a possible partner who has similar interests. Another research assistant is creating a remembrance agent that tracks articles, emails, or files that are related to whatever its user is working on. Maes says that she uses this agent when she writes reports and that it often finds articles she neglected.
Bradley Rhodes, who is developing this agent, has it running on a purple wearable computer - color-synched to his purple T-shirt - that hangs from the corner of his black beret. Maes isn't ready to go that far, "not until they become fashionable enough."
Last fall, Maes used her agents in a real-world situation in setting up an elaborate electronic marketplace called Kasbah, a Media Lab agent extravaganza. Two hundred symposium attendees were given a bag of goods and fake money, plus access to agents that could be instructed to buy and sell. People went wild, Maes recalls. "Some of the participants really should have been thrown in jail - they promised to buy certain things when they did not have the money."
A black market emerged - most of the Japanese sponsors skipped the symposium and haggled instead. Maes and some colleagues have just implemented a permanent system called MIT Marketplace, in which people can electronically buy and sell CDs, books, and other small items via software agents.
Blame it on the bot
The Kasbah experiment raised some of the familiar and troubling integrity issues surrounding agents: Are you responsible for your agent's purchases? How can you trust an agent? And how do you keep your agent abreast of your offline transactions?
Maes notes that the possibility for the misuse, co-option, or failure of agents is ever-present and foresees the development of reputation servers that would automatically verify the credibility of an agent and help earn the public's trust in the new and generally unfamiliar technology.
But the benefits are also impressive, she observes, as agents could, say, do comparative shopping or help users organize a consumer cartel.
None of these agents have individual personalities. "For a long time, people assumed we were trying to build an agent that behaved like a real person,"Maes says. This agent personification could be "really powerful because people have very strong emotional relationships with characters. If we can impart personality to software, especially educational software, it could be a very motivating factor."
Such visions worried those who thought that users would trust these agents too much. "Some people even argued that if we were used to sort of delegating all these tasks to little people in our computers that maybe we would start bossing around people in the real world as well," says Maes.
Ben Shneiderman of the University of Maryland and virtual reality pioneer Jaron Lanier believe that agents are a disturbing idea when they are presented as being "autonomous," which suggests that people are no longer in control of the software.
But both say they've made their peace with Maes. "As she defined her terms a couple of years ago, I thought she was on a dreadful antihuman track," says Lanier. "But Pattie has changed her definitions so that what she says no longer troubles me so much." Lanier says he is comfortable with agents "as long as there is no deception. As long as nothing is hidden from the user. And as long as obfuscation isn't the interface."
Both Lanier and Shneiderman say that they admire Maes's intellectual openness, as well as her concerns about privacy and about the place of agents in our psyche. And besides, adds Shneiderman, "How do I debate someone who is among the 50 most beautiful and 100 most influential? That's a tough job."
The People article, describing Maes as "a download diva," managed to win her some detractors. The article also quoted her as saying that MIT "is almost a wasteland in terms of beauty." She immediately sent out an email saying that her words had been taken out of context and altered.
"I haven't heard anything bad around here," notes Rosalind Picard, a professor at the Media Lab.
Maes, a part-time model in college, says that she agreed to be profiled because she thought it would be good for readers to see a woman recognized for her brains. Even now, she gets upset about the article, and her homepage offers a link to an explanation for anyone who has read the piece.
Getting noticed for her looks may have backfired, but Maes says getting attention for being a woman has generally been helpful. "If you're woman in a field where there aren't as many women, you get more attention, rather than less. Or more attention than men at the same level," admits Maes. "So I haven't experienced it to be a negative thing at all. I have never made a professional distinction between men and women. I've never compared myself with other women. I have always compared myself with everybody else."
Maes possesses an incredible drive and an insatiable curiosity. Not content to have a quiet summer, she's teaching a new fall course on intelligence augmentation. "I just like diving into new things, especially getting involved in starting a whole new field," she explains. Her newest target is medicine. Maes is up for tenure this winter - if all goes well, she will take a sabbatical next year to think about healing and agents. Again, she smiles, the motivation is purely personal: "I am getting older and thinking more about my health."
On another walk across the campus, this one for lunch, Maes contemplates offers from a Japanese cosmetics company and a Japanese electronics firm to appear in television commercials.
She quickly steps into the moving wedge of a revolving door. Out the other side, she talks about traveling more. "I always fear my life will be boring - then something happens." She has already visited reefs in Belize, Fiji, Palau, and Hawaii, and she plans to dive at least once in every ocean on the planet. Just as soon as the agents release her, she'll be down among the brain coral.
