Image by Francesco Romoli

What are humand good for? Members of SFI’s Applied Complexity Network (ACtioN) confronted this question again and again over the course of the November symposium on natural and artificial intelligence.

In the most pessimistic interpretation, the answer is: not much. Machines beat us in chess in 1997 and now, in 2017, in the notoriously complex game of Go. Algorithms are producing nostalgic playlists, tweeting prolifically, designing fonts, and co-authoring scientific papers. And although their sonnets remain stilted for now, robots’ weather and sports reports have become the norm, whether we realize it or not.

Of course, there are also algorithms that spend their days massaging cat photos into the form of a loaf of bread. But even these are a fertile breeding ground for machine learning, suggests Y Combinator’s Michael Nielsen. The real question isn’t what computers can do, but rather what computers are for. Are they simply machines for answering questions — what Nielsen terms “cognitive outsourcing?” But can machines change the range of thoughts that we can think? Can they transform our cognition?

Since long before the days of dunce caps, we have associated intelligence with computation powers far more than with kinesthetic or emotional abilities. But SFI President David Krakauer offered an alternative definition: “Intelligence,” he declared, “is making hard problems easy.”

The simple act of walking across a room without having to think about each step represents a form of intelligence, according to SFI Professor Nihat Ay. His research group is building a theoretical framework for understanding “embodied intelligence,” which could one day be used to re-create natural movement in robots.  For humans, this form of intelligence emerges from a learning process — a self-organized interplay between brain, body, and environment. Robotics, by contrast, is still dominated by the paradigm of pre-programmed control from a central computer.

Machines do not experience the open-endedness of childhood — a developmental period which, arguably, is responsible for making us human. Most machines learn using human-generated datasets. Even AlphaGo, which was not trained using human data, still operates entirely within the game’s rigid parameters.

As External Professor Melanie Mitchell (Portland State University) pointed out, babies spend all their time simply discovering the physical world around them — touching, drooling, biting, and developing common sense. AI can beat us at Go, but it can’t tell us “whether Michael Phelps’ hair was wet when he got out of the pool.”

“In robotics, you get regularly humbled by the real world,” says Philip Heermann, who attended with Sandia National Labs. Things that come naturally to humans — sitting, walking — are often comically difficult to replicate. But should we be replicating them? Or splitting our tasks? 

If humans were only ever good for playing Go, tweeting, and recognizing voices, there wouldn’t be much use for us anymore. The key now, it seems, is to play to our strengths: problem-solving, innovation, and play. Esther Dyson (Way to Wellville), a former SFI Trustee, remarked, “I would like to see all of those truck drivers become gym teachers and soccer coaches, and pay gym teachers and soccer coaches more.”

Technically, we’ve been commandeering “other” — if not “artificial” — intelligence for millennia. We took a hard problem — traveling fast, carrying heavy goods — and recognized that it was not our strong suit. We domesticated the horse. And here we are today.


ACtioN members can access video of the symposium by logging in and navigating to the symposium the event page (November 3-4, 2017)