AI as a Complexity Interface

Developing windows for our information spaceship

October 20, 2020

Imagine for a moment that our world was managed by a large scale, short-sighted and very dull bureaucracy. One day, our zealous but drudging administrators would send a committee to investigate the results of the frenzied scientists and engineers working on Artificial Intelligence.

My bet is that they would be unimpressed. Enabled by their natural defenses against excitement to look beyond hyped presentations and flashy demos, they would see a technology that despite huge investment is often brittle, easy to fool, sensitive to minor specificities, expensive and risky to deploy, support and maintain. “What is the big deal with this AI?” would probably ask the Chief Inspector towards the end of his visit at a major lab.

At this point, after many had failed to impress the Chief Inspector with speculative promises about the future abilities of their creations, one shy scientist might step forward and admit: “Our work is not only motivated by the abilities of current and future AIs, but also by the limitations of our alternatives, with human cognitive limitations on top of that list”.

We often talk about developing AI as the idea of replicating human abilities using computers, but that’s only half of the story, because AI is not only an aesthetic exercise of replicating something that exists but mostly an investment in pushing the boundaries of what technology can enable us to do. As Robert Miles puts it, “AI is about making machines do cognitive tasks that we did not think they could do”. Going beyond, among those tasks that we did not think machines could do, those tasks that even humans cannot do is where the value of our investment may be.

These cognitive tasks (that we cannot do or can do at limited scale or speed) share a common denominator: complexity. In fact, a cognitive task that is difficult to do is complex by definition. There is a consensus that, exhausted from the outstanding exponential progress of the last centuries, most of the remaining open problems in science, engineering and society are all very complex. In the sense that their structures, patterns and mechanisms are very difficult to grasp using human cognitive abilities, which we need the help of technology to overcome.