The Perils of Prediction

In a column two months ago I reminisced about a natural-language breakthrough from 1971. Terry Winograd’s program SHRDLU was a simple graphical environment, laughably primitive by today’s standards. But users could command the program to manipulate the boxes, pyramids, and blocks in the environment using everyday written English.

Anyone viewing SHRDLU in 1971 would have predicted great things for computers’ ability to understand language and, more generally, for artificial intelligence (AI). But the closest that most of us come to AI today is the grammar checker in a word processor.

The other technology in SHRDLU, built simply to showcase the natural-language processing as the main act, was the blocks world. This was a simple 3D wire-frame graphical environment. Though not bad for its day, it gave no indication of a bright future for computer graphics.

And yet, in sharp contrast to AI, graphics from that point made remarkable progress. Computer graphics have become so important that video games are the only major PC application that still justifies ever more processing power. The quality has become so good that creating truly realistic human characters is one of the few remaining challenges for computer-generated movie special effects.

In hindsight, we can see why graphics succeeded but AI didn’t. The obstacle for realistic graphics was not ideas and algorithms but computing horsepower, and Moore’s Law has provided that in abundance over the past 35 years. This isn’t to dismiss the creativity and genius behind the software that gives us dazzling movie special effects and video games, but simply to note that software wasn’t the bottleneck.

On the other hand, the stumbling block for AI has been the software. Who cares how fast your processor is when you don’t have the algorithm to run on it? We simply don’t know how to program language understanding, vision or human creativity.

Predicting the Future

What a surprise to have serious and practical AI proceed slowly while computer graphics, driven by frivolous games and movies, proceeded quickly. The problem is that you can’t schedule a breakthrough. Progress goes at its own pace.

AI researchers were hit with what I think of as a sound barrier problem. Progress is steady until “Wham!” you hit an unforeseen difficult obstacle.

These problems have sprung up in many disciplines. Virtual-reality researchers didn’t expect that users would occasionally suffer vertigo, sometimes hours after leaving a VR simulation. Civil engineers built caissons in rivers to allow workers to dig down to bedrock for bridge foundations, but they didn’t expect workers to suffer from the bends. Airplane designers didn’t expect the sound barrier to be so challenging, and even today the speed of sound remains a barrier for commercial airplanes.

Computer graphics and AI illustrate two of the four possible technology outcomes; that of long-term success and long-term disappointment. Here are examples of the final two outcomes:

In 1967, the 100-year-old company Keuffel & Esser was commissioned to study the future. A major failure of its analysis was not seeing that its own flagship product would become obsolete in just a few years. K&E was the country’s leading slide-rule manufacturer, and it was blindsided by the product it failed to see, the electronic calculator.