Technology Predictions: Intelligence and brute force

By Andrew Odlyzko

Technology prediction is inherently hard.  And it is even harder to
predict how society will react to a new product or service.

Potential customers may sneer at a new technology, as happened a decade
ago with application service providers (ASPs). Or they may embrace it,
as seems to be happening with today’s incarnation of ASPs, cloud
computing.  (Of course it is still too early to tell if what we see is
truly an enthusiastic embrace, or simply hype generated to stimulate an
enthusiastic embrace.)  The presence of complicated feedback loops—hype
can inspire creation of new applications, which will make a service more
attractive and persuade people to try it—makes the task of prediction
even harder.  So it is no wonder that "progress by mistake" is not just
frequent, but almost a rule.

An unexpected killer app Technology can surprise on the upside as well
as the downside.  Email, which was specifically excluded from the design
criteria for the Arpanet, became the "killer app" of that network as well
as its descendant, the Internet.  Who could have known at the time that
the computer mouse, demonstrated by Doug Engelbart over 40 years ago,
would today still be the key device for human-computer interaction? And
the World Wide Web, now 20 years old, spread slowly for several years,
until the release of the Mosaic browser made it widely accessible—and
then it caught fire.  But even then, in the first few years, there was
considerable speculation that even better tools for accessing information
over the Internet might emerge.

What can we conclude from the long history of failed technology
predictions?  Wide experimentation is certainly called for, as well as
maximizing the flexibility of new technologies, in order to accommodate
demands that one did not foresee initially.  One should not count on
serendipity, but be prepared for it.  And of course, we should ride
the technology curve, taking advantage of the various Moore's laws that
provide predictable progress in information technologies, at rates that
vary from field to field.

Under- and overestimating Aside from the widely accepted principles
above, there are a few other patterns that one can discern in the
history of predictions about technology.  Thus, although general
technology forecasting is unreliable, some predictions have proven
correct over an extended period of time. A fairly persistent pattern
is the underestimation of the continuing increases in processing power,
storage capacities, and communication bandwidth, and overestimation of
the extent to which computers can be made to reason like people.

A striking example of this dichotomy is provided by J. C. R. Licklider's
book Libraries of the Future, published in 1965.  Licklider has the best
claim of anybody to be called the "grandfather of the Internet," as he
was the first one to point to computers as being primarily communication
devices, not just computing ones, and he set up the program that led to
the creation of the Arpanet.  In his book, he made many predictions. Some,
about development of computer networks, and about digital libraries
becoming feasible around the year 2000, are among the finest examples
of futurology.  But those were based primarily on extrapolations from
basic technology trends.  Many of his forecasts were wrong, in particular
those based on expectations that computers would acquire intelligence.

A similar pattern appears in other areas:  Speech recognition has
made great strides, primarily by exploiting more powerful technology
to do massive pattern matching, rather than by the methods pursued
in the 1960s of trying to get computers to understand human speech.
Language translation followed the same pattern. And so did chess. The best
computer chess programs can handily beat the best human players today,
but not by imitating human thought processes (which presents us with a
mystery "Why are there no contests involving pairings of people and
computers on each side?").

Human, not artificial, intelligence With the Web, too, brute force has
triumphed, although that brute force is directed by human intelligence,
in the form of clever algorithms. (Clever algorithms were also needed for
the advances in speech recognition, language translation, and chess.).
The popularity of the Web obtained a substantial boost from the
appearance of AltaVista, the first popular search engine. AltaVista's
breakthrough, later improved on by Google, was to demonstrate that
with sufficient computing, storage, and communications resources, one
could do effective automated crawling and indexing.  But AltaVista's
managers, for what seemed to be good business reasons at the time,
made the misstep of switching their focus to making AltaVista a portal,
and thus facilitated Google's rise to dominance. Google succeeds largely
through use of massive resources, with direction from clever methods,
but not ones drawn from conventional AI.

The Web is evolving rapidly.  And there are hopes for major breakthroughs
based on computer understanding of the growing volume of digitized
data. Yet, if we go by historical precedents, such hopes will be
disappointed. Computing, storage and communications are all progressing
rapidly, even if somewhat less rapidly than a decade ago.  Hence it is
most reasonable to expect the incremental improvements they will provide
(together with improvements in standard data mining, visualization,
data base, and related algorithms) to be the main contributors to the
Web's evolution.