Message: 12
Date: 6.1.94
From: <nicholas@media.mit.edu>
To: <lr@wired.com>

Less Is More: Interface Agents as Digital Butlers

Al Gore need not be right or wrong in his conception of details. It almost doesn't matter whether he calls it an information superhighway, an infobahn, or a National Information Infrastructure. What matters is his personal and sincere interest in computers and communications and the fact that his enthusiasm has raised our popular consciousness of telecommunications. The media cacophony over phenomena like the Internet fosters an open architecture and emphasizes access by all Americans.

The clamor, however, has perpetuated a tacit assumption that more bandwidth is an innate, a priori, and (almost) constitutional good. The right to 1,000 channels of TV! Continental Cable, the local cable company in Cambridge, Massachusetts, now offers Internet access at 500,000 bits per second. With that service, The Wall Street Journal takes sixteen seconds to transmit in its entirety (as structured data mostly, not fax, please!). When fiber reaches the home, by some estimates, we will have access to as much as 100 billion bits per second. Hmmm. Most people generally make a false assumption that more bits are better. More is more.

In truth, we want fewer bits, not more. Our needs fall along a spectrum. Consider a newspaper: Our requirements are very different on Monday morning from what they were on Sunday afternoon. At 7 a.m. on a workday, you are less likely to be interested in browsing stories. Serendipity just does not play a key role then. In fact, you would most likely be willing to pay The New York Times US$10 for ten pages vs. $1 for 100 pages. If you could, you would opt for a heavy dose of personalized news.

It's simple: Just because bandwidth exists, don't squirt more bits at me. What I really need is intelligence in the network and in my receiver to filter and extract relevant information from a body of information that is orders of magnitude larger than anything I can digest. To achieve this we use a technique known as "interface agents." Imagine a future where your interface agent can read every newspaper and catch every broadcast on the planet, and then, from this, construct a personalized summary. Wouldn't that be more interesting than pumping more and more bits into your home?

Why do people pay 85 cents to find out whether their one daily lottery ticket won? TV Guide has been known to make larger profits than all four networks combined. What do these things tell you? It says that the value of information about information can be greater than the value of the information itself. From that and other similar observations (American Airlines makes more from its reservation system than from carrying passengers) I am willing to project an enormous new industry based on a service that helps navigate through massive amounts of data.

When we think of new information delivery, we tend to cramp our thoughts with concepts like "info grazing" and "channel surfing." These concepts just do not scale. With 1,000 channels, if you surf from station to station, dwelling only three seconds per channel, it will take almost an hour to scan them all. A program would be over long before you could decide whether it is the most interesting.

I am fond of asking people how they select a theatrical, box-office movie. Some pretend they read reviews. I hasten to interject my own solution - which is to ask my sister-in-law - and people quickly admit that they have an equivalent. What we want to build into these systems is a sister-in-law, an interface agent which is both an expert on movies and an expert on you.

Your Model of Its Model of Your Model of It
The key to agent-based systems is learning. It is not a matter of a questionnaire or a fixed profile. Agents must learn and develop over time, like human friends and assistants. It is not only the acquisition of a model of you; it is using it in context. Timing alone is an example of how human agents distinguish themselves. But it is all too easy to wave your hand and say "learning." What constitutes learning?

The only clue I have found goes back two decades to the work of the English cybernetician Gordon Pask, who taught me to look at the second- and third-order models. In human-to-computer interaction, your model of the computer is less telling than its model of your model of it. By extension, your model of its model of your model of it is even more critical. When this third-order model matches the first (your model of it), we can say that you know each other.

Swiss Banking of Network-Based Agents
All of us are quite comfortable with the idea that an all-knowing agent might live in our television set, pocket, or automobile. We are rightly less sanguine about the possibility of such agents living in the greater network. All we need is a bunch of tattletale or culpable computer agents. Enough butlers and maids have testified against former employers for us to realize that our most trusted agents, by definition, know the most about us.

I believe there is a whole new business in confiding our profiles to a third party, which will behave like a Swiss bank. I fear this will not be one of my credit card companies, which have sold my name for all sorts of purposes, and have thus shot themselves in the foot. It must be a credible third party, perhaps a local telephone company, perhaps a long distance company like AT&T, perhaps a new venture altogether. What we should be looking for is an entity which is able and willing to keep our identities confidential while at the same time passing along newsworthy advertising and information.

Such services will only work with a high degree of machine learning. While it is important to postulate such learning, how does this relate to human learning?

Next Issue: Learning vs. Teaching

[Back to the Index of WIRED Articles | Back to Nicholas Negroponte's Home Page | Back to Media Lab Home Page]
[Previous | Next]

[Copyright 1994, WIRED Ventures Ltd. All Rights Reserved. Issue 2.06 June 1994.]