Pattie Maes and Her Agents Provocateurs
Artificial intelligence research used to focus on teaching machines to think. Until a Belgian MIT professor taught them to shop.
For a leading researcher at one of the United States' top universities, Pattie Maes is surprisingly uninterested in the technological underpinnings of her own work. That's not to say that her basic research in artificial intelligence (AI) isn't impressive. The 38-year-old associate professor at the Massachusetts Institute of Technology's Media Lab has engineered major breakthroughs in software agents, programs that are changing the face of Internet shopping and are on the verge of turning retailing on its head. But to Dr. Maes (pronounced "Mahs"), it's not the technology but its effect that is important. She is the prototype of the New Economy scholar/entrepreneur: a scientist in academia committed to seeing her ideas become commercial successes.
In navigating between the theoretical and the practical, Dr. Maes has been on a nonstop journey for the last 10 years to show that AI's promise is more than just an academic exercise. In a discipline with one disappointing commercial venture after another (remember Danny Hillis's Thinking Machines Corporation and Roger Schank's Cognitive Systems Inc.?), Dr. Maes has now launched two successful Internet startups, Firefly Network Inc. and Frictionless Commerce Inc. Both have spawned imitators and solidified the reputation of software agents as a revolutionary force in both Web commerce and bricks-and-mortar retailing. Dr. Maes's third venture, Open Ratings, is a ranking system to improve the effectiveness of e-commerce relationships between business buyers and suppliers, using advanced databases that dig deeply into individual transactions on Web sites.
"I realized that computers could augment people," Dr. Maes says. "That's what all of my research is really about. Our lives are so busy and there are so many choices to make, so many opportunities to keep track of, and so much accessible information; we end up feeling overwhelmed. If we could make computers smarter and better serve people, we'd be on to something. I think of this as intelligence augmentation — IA rather than AI."
The promise of intelligence augmentation led Dr. Maes to found the Software Agents Group at the Media Lab in 1991. She saw this as her chance to take artificial intelligence well beyond what it was best known for — expert systems. These programs essentially are force-fed heuristics — the "if I do this, then this is likely to happen" rules — that guide the decisions doctors and lawyers, among other professionals, make every day. The computer programs recreate in a mostly stolid, uneducable fashion how an expert would react to a situation.
Software agents, by contrast, are designed for average computer users who want to perform more mundane but information-intensive tasks, such as online matchmaking or comparison shopping. Most importantly, they adapt and learn, depending on how their human "master" answers questions, responds to offers, makes purchases, and expresses preferences, among many other criteria.
The fruit of Dr. Maes's work, embodied in these software agents (or "bots," as they are also called), is already playing a significant role in e-commerce, and by this holiday season they may make a dent in the bricks-and-mortar retail environment. One scenario a few years down the road: A shopper is interested in a bike at Sears or a car at a Ford dealer, and, via a personal digital assistant (PDA) or a cell phone connected to the Web, sends out a bot in search of a better deal. This digital alter ego negotiates a purchase directly with other retailers' software agents and directs the shopper to the physical location of the retailer to see the item. After the shopper approves the purchase, the bot works out the final details of the deal — perhaps cutting a bit more off the price, taking an extended warranty, or adding additional features — and completes the transaction as the shopper drives home with the product.
"It's a whole new demographic category: machines that buy," says Thornton May, senior vice president of research for Cambridge Technology Partners, a consulting firm. "This will change the whole nature of marketing. Think Blade Runner meets Filene's Basement. A retailer isn't going to be able to play if it's not good at writing these bots."
Dr. Maes isn't the only successful bot researcher, but she's certainly the most famous. Her celebrity was cemented in the last few years when both Time and Newsweek tabbed Dr. Maes a member of the "cyber elite" and a name to watch in the new millennium. In 1997, People magazine even named her one of the 50 Most Beautiful People in the World.
Dr. Maes rolls her eyes at the memory and quickly changes the subject. What matters most to her, she says, is seeing her ideas make a big difference: "I like to really change things. I'm not interested in incremental changes. I like to create little earthquakes that shake up the whole way something is done."
Smitten with Technology
Her journey to MIT's prestigious artificial intelligence program started in 1989 with a determined fan letter to AI "bad boy" Rodney Brooks. "I have money," she wrote. "I'm coming over for two months. I love your work, and this is what I want to be working on with you."
The original AI intelligentsia, brilliant computer scientists and mathematicians like MIT's Marvin Minsky and John McCarthy, created the discipline in the mid-1950s to prove that machines could be made to think and behave like humans. Their approach was that knowledge is the key to intelligence. If things such as vision, understanding natural language, and comprehending the front page of a newspaper could be broken down into discrete packets of knowledge, these activities could be programmed into computers using millions of lines of code. But the payoff for AI researchers involved in this brute-force method of teaching computers to think has been limited, mainly because these systems lack the ability to learn and adjust beyond their programmed knowledge.
Prof. Brooks's approach was more pragmatic. Instead of trying to recreate cognitive processes, he built small autonomous systems to perform simple tasks. Through personalization, mining, and filtering of their databases, these systems adapt to conditions by repeating successful decisions. Prof. Brooks's technique grew out of traditional AI, but its emphasis was on using information and the speed with which a computer can organize and race through data to achieve practical, if not intelligent, results. Most of his early work — which was criticized by AI purists because it didn't try to decipher and encode cognitive processes — was in building small robots, called "bugbots," that mimicked insect behavior.
Dr. Maes had earned a Ph.D. degree in computer science from the Vrije Universiteit Brussel in Belgium, studying AI and human/computer interaction. But she was frustrated at the slow pace of practical innovation in AI and was attracted to Prof. Brooks's radical attempts to wake up the discipline.
Almost immediately after she arrived at MIT, it became clear that her enthusiasm for Prof. Brooks's work and her own talent were irrepressible. As her stint at the Media Lab was ending, Prof. Brooks asked her to stay on for a year as a visiting professor. "I couldn't imagine going back to Belgium," Dr. Maes says. "The level of energy and excitement at MIT is just addictive."
Her excitement about research and academia grew out of her upbringing in a Brussels suburb. She was the third daughter in a family of six children and had a large extended family of professors, teachers, and engineers. But Dr. Maes's undergraduate decision to major in computer science reflected less passion than pragmatism. She was interested in biology and architecture, but a recession in Europe in the late 1970s made jobs in those areas difficult to find. With technology booming, a computer science degree was more likely to guarantee work, Dr. Maes figured. An AI course during her junior year gave her the clear direction she needed.
The collaboration between Dr. Maes and Prof. Brooks produced virtually instant results, particularly in designing bugbots that were able to move haltingly but could adapt after making mistakes and gradually improve their ambulatory skills based on increased knowledge about their environment. "They produced some impressive demos very quickly," says Oren Etzioni, a University of Washington computer scientist who also specializes in software agent research. "The intelligent insect stumbled at first but eventually learned to walk. A lot of the component- or knowledge-based work was done in simulation. But Brooks said, 'Let's not work on components, but on a whole agent, and let's do it in the real world.'"
Prosthetics for the Mind
What sets Dr. Maes apart from other successful researchers is her keen intuition for technology trends. In the early 1990s, long before the Internet was a commercial medium, she sensed an opportunity to marry her ideas about intelligence augmentation to the Web. She envisioned a direct connection between the robotic bugs and autonomous software systems or agents that would live and evolve in a purely digital world. The collective intelligence of the Internet — a vast, pliable, and constantly expanding database in itself — was a perfect home for her style of artificial intelligence, Dr. Maes thought. Science had long ago found solutions for physical and sensory shortcomings — eyeglasses, hearing aids, artificial limbs. Why not prosthetics for the mind?
"There are just so many things that people aren't good at and that machines are good at," Dr. Maes says. "For example, multitasking, keeping track of many, many things at once. Or doing extensive searches and remembering things. A lot of my work has been about using these fairly simple agents — a little army of software entities — to augment people's cognitive capabilities."
The concept of software agents, espoused by digital futurists such as Nicholas Negroponte, the director and founder of the MIT Media Lab, and Alan Kay, a fellow at Apple Computer Inc. and Walt Disney Imagineering, had been discussed for more than a decade, but few breakthroughs were achieved in the pre-Internet world. Software agents were more concept than reality since they lacked a ubiquitous network. Finally, here was the platform they needed.
Dr. Maes founded the Software Agents Group at the Media Lab to piggyback on the promise of the Web. Central to the group's research is basic pattern and response recognition. Early on, for example, Dr. Maes's group built an Internet browser capable of remembering the Web pages a person downloads. The software notices certain key words and information in which the user appears interested and is programmed to look for other pages that display these words or ideas.
As Dr. Maes puts it, her software agents "watch over the shoulder of users and notice certain patterns and regularities in their habits and interests that can be automated. Since many users have similar habits and interests, software agents can learn how better to assist their own users by exchanging information with and learning from other agents that assist like-minded people."
Most important to Dr. Maes, this type of software is capable of independence from its user. "The software agent is a dynamic entity that can be acting on our behalf while we are talking or sleeping or working," she says.
There's a confidence and a sense of purpose about Dr. Maes that academic researchers often don't have, say those who know and work with her. Bruce Blumberg, a former Apple Computer engineer who was Dr. Maes's first Ph.D. student in 1991, says she reminds him of Apple CEO Steve Jobs. "Like Steve, Pattie has the ability to ask just the right question, out of which the correct answer will emerge," Dr. Blumberg says. "And also like Steve, she has the fundamental belief that it is not as complicated as it seems. Confidence in one's own ideas can be good or bad. In Pattie's case, it's probably the source of her intuition, an ability to believe that just because someone says it is so, doesn't make it so."
Bots Take Charge
In the mid-1990s, with virtually every technology campus infected with startup fever, the commercial possibilities for software agents excited Dr. Maes and her students. Well-trained and targeted bots offered e-commerce ventures a way to personalize customer contact, maximize shopping efficiency on the Web, and give people more reasons to buy online.
With its potentially powerful and lucrative impact on e-commerce models, Dr. Maes's research easily attracted attention outside academia. IBM used her work in part as the inspiration for its Agents and Emergent Phenomena group at the company's Institute for Advanced Commerce. Jeff Kephart, one of six researchers on the IBM team, says he's certain the future of e-commerce will teem with billions of software agents haggling with each other on the Internet. These bots, he says, will evolve naturally from facilitators into decision-makers, and their degree of autonomy and responsibility will increase over time. Ultimately, transactions among these economic agents will constitute an essential and perhaps even dominant portion of the world economy. That could result in significant business discontinuities, such as price wars that break out suddenly and last 10 minutes or 10 hours, depending on the balance of supply and demand.
Wary of how this could shift the balance in commercial environments, the IBM team wants to simulate in the lab the impact of such an agent dynamic. The purpose, Dr. Kephart says, is to figure out "what effect this will have on the economy. We may find behaviors that may not be very good and we'd like to try to solve these problems in the lab before we let them loose in the real world."
Because of Dr. Maes, these agents are already skittering around the Web in increasing numbers. Her first commercial release of bots occurred in 1995, when she launched Firefly Network. Firefly agents perform what Dr. Maes calls an "electronic word of mouth" function for its users, who are consumer-shoppers. The technical name for this, collaborative filtering, refers to the bots' ability to provide an individual consumer with customized recommendations for products like compact disks, books, and videos, based on his or her past purchasing or browsing behavior. Most good portals and retail sites now use some form of collaborative filtering, which enables sites to greet visitors by name and steer each person to individually targeted product suggestions.
Although a founder of Firefly, Dr. Maes initially worked there only one day a week and never served as chief executive. She adamantly refuses to get involved in the operational side of her business forays, finding excitement in generating ideas and in watching a company evolve during its first six months. But through her contacts she helped Firefly consummate a deal with Microsoft, which bought Firefly in a stock swap for $40 million in 1998. (Dr. Maes won't say how much she pocketed in the Microsoft deal.)
Before this deal, Firefly had negotiated an exclusive agreement with Barnes & Noble, allowing the bookseller to use its collaborative filtering technology on the B&N Web site. Seeing Firefly hook up with Barnes & Noble spurred Amazon.com Inc. to find its own partner, and Amazon soon signed a deal with a startup named Net Perceptions Inc., which was founded a year after Firefly and uses similar collaborative filtering agents. As Amazon's popularity soared, so did the value of Net Perceptions. Net Perceptions went public in 1999; its share price reached about $60 in January 2000, giving the company a market capitalization of more than $1.7 billion. Even after slipping to $17 per share in June, Net Perceptions is still worth nearly $400 million — 10 times what Microsoft paid for Firefly. Had Firefly waited a bit longer to close the Microsoft deal, instead of jumping into the arms of the first suitor with cash, it might have sold for a higher price.
What Microsoft wanted wasn't just Firefly's agents. In building Firefly, Dr. Maes's group added an enhanced profiling capability they dubbed Firefly Passport, which lets Web users control personal data and decide which sites can use it. Since its acquisition of the company, Microsoft has been much more aggressive about using Firefly Passport than the company's collaborative filtering agents, because of users' concerns about privacy on the Web.
Soon after the Microsoft deal, Dr. Maes founded Frictionless Commerce with Rob Guttman, who worked in her Software Agents Group, and Alex Kleiner, a student at MIT's Sloan School of Management. In fact, one of Dr. Maes's innovations was initiating a collaborative agreement, with Sloan professor Erik Brynjolfsson, between the Media Lab and the Sloan School to foster cross-disciplinary approaches to the New Economy.
In many ways, Frictionless is the next technical step beyond Firefly. Its software agents are designed to do one key thing that collaborative filtering can't: help buyers decide how to buy. Online shopping choices can be overwhelming. Thousands of similar products are available at retail sites, Web auctioneers, electronic exchanges, and more. To cut through this, Frictionless bots act as extremely knowledgeable one-on-one digital salespeople, completely aware of all the products and selling styles, models, and techniques of every store in the vast Internet mall.
Using gigantic databases with robust decision-support technology, Frictionless tracks down items, no matter where they're offered on the Web, and pulls them into a product comparison mix for the shopper. It also allows for an open-ended number of criteria to guide purchase decisions — such as price ranges, special options, delivery times, and warranties. The customer gets a deep, rich, side-by-side comparison among products, even if items are described differently by retailers. First-stage customers of Frictionless include Lycos's Compare Products, Wingspanbank.com, and Webhelp.com.
Without skipping a beat after launching Frictionless, Dr. Maes turned her attention to a new interest: business-to-business (B2B) e-commerce. Her latest venture, Open Ratings, is less a software agent than an automated B2B version of eBay Inc.'s trust and reputation system, offering ratings services for e-commerce sites.
"One of the major ingredients to really make these exchanges take off is trust," Dr. Maes says. "If you are a chemical company in need of supplies, you don't want just any company to bid on that RFP and deliver those chemicals to you. You want to make sure the company has a solid reputation for customer satisfaction. You might care more about on-time delivery or quality of the products than you care about price, for example."
Online exchanges and their customers need a reliable independent ratings system that incorporates real feedback from actual customers, Dr. Maes figures. To that end, Open Ratings licenses its ranking system database to online exchange sites as well as to companies doing business on those sites, and manages the database for them.
Although Open Ratings is not itself an agent, it will be a crucial link for software agents that actually are empowered to do buying and selling on a company's behalf. A bot, for example, might be instructed to negotiate only with businesses that have an Open Ratings four-star rating.
Open Ratings is a logical extension of Dr. Maes's notion that today's e-commerce is little more than a hint of what's to come. It may seem like retailing, but what has occurred will pale against the revolutionary effect technology will yet have on commerce in general and shopping in particular. Like many other Web futurists, she believes the New Economy — some of it taking place on wireless media, some in the bricks-and-mortar world — will be characterized by a preponderance of haggling and negotiating, with much more valuable information available to help people make decisions. But the negotiating, data analysis, and buying will be conducted by agents. "The power has been shifting to the buyer, and this will help shift it even more toward the buyer," Dr. Maes says.
For a hint of what Dr. Maes means, the latest round of research in her group includes a program called "Shopping on the Run," a system that broadcasts shopping requests over mobile networks directly from cars and lets local shops and restaurants respond with information about sales, price, reservations, directions, and the like.
Community Ware Is Next
Despite her devotion to bots and to launching a shopping revolution, Dr. Maes confesses that she's quickly bored. Consequently, her next project, something known as Community Ware, will take her away from hard products in the business-to-consumer and B2B environments and bring her closer to a more esoteric commodity — information. Dr. Maes foresees an "electronic market for knowledge," in which software agents match up students and experts and make deals for on-the-fly assistance.
As she describes it, to participate in this knowledge marketplace, a person can send out a selling agent that knows his or her expertise, availability, and rates. Someone looking for an expert would, in turn, create a buying agent that knows what information or background is being sought and when it is needed — in the next three minutes, in three days, etc. The buyer's agent might have a comparison shopping bot built into it to judge price, reputation, level of expertise, and availability.
At the heart of this, Dr. Maes says, is the idea that computer networks peppered with software agents make it increasingly easier for individuals to share knowledge with others and cut down on redundancy among people. Very few of the problems we solve as individuals are truly original, Dr. Maes points out; in 99.9 percent of the cases, the problem has been dealt with before. Someone else has read the article that you are about to read and could help you decide whether it is worth reading or not. Someone else has solved the type of differential equation you need to solve and could help you reach the right answer.
There's a certain bittersweet quality to Dr. Maes's discussions about her work these days. She's planning to leave the Media Lab within 18 months to spend more time on commercial ventures and with her husband and two young sons.
It's not that Dr. Maes is planning to sit in an executive suite anytime soon. But she now believes that she can work more effectively from outside the walls of academia to use technology to change commercial relationships and augment people's lives.
"Ultimately, what excites a researcher is not just to get approval from your peers," she says, "but to get your stuff out there and have an impact. These days, that kind of work can be done outside the university. You don't have to be at a university anymore to introduce totally new technologies and concepts. You can do it directly in the commercial world. It is such a great time to innovate."
Reprint No. 00307
Glenn Rifkin, firstname.lastname@example.org
Glenn Rifkin has covered technology for the New York Times and has written for the Harvard Business Review and Fast Company. He is coauthor of Radical Marketing (HarperBusiness, 1999) and The CEO Chronicles (Knowledge Exchange, 1999).