2007 : WHAT ARE YOU OPTIMISTIC ABOUT?

jordan_pollack's picture
Professor and Chairman of Computer Science, Brandeis University
Computer Scientist, Brandeis University

AI Will Arise

I often attack the original "we can program it" direction of the field of Artificial Intelligence, but am still optimistic that our primitive electromechanical and computing machines will one day become intelligent enough to treat as living creatures.

I have a predictive sketch for how intelligent machines might arrive. My definition of a robot is any device—controlled by software—interacting with the physical world. An economically viable robot is such a system that earns a consistent return on investment (ROI). The ATM, the ink-jet printer, and the disk drive are today's omnipresent robots, they just don't appear as the robots of science fiction.

I see three streams which can come together in the future to allow the emergence of intelligent human-like robots.

The first stream is the lowly "web bot." Moving beyond "Eliza's" in text chat rooms, these animated humans often exist on web pages, attracting customers, helping them navigate, or selling. Another class of virtual humans are employed as extras in video games and movies. These software puppets will enter a positive feedback loop as they are programmed to exploit human psychological weaknesses. In other words, when a virtual human becomes an effective sales machine for, say, life insurance or securities, they will have achieved the ROI.

The second stream is animatronics. The word, coined by Disney, is about making mechatronic puppets which entertain us, i.e. Chuck E Cheese and Big Mouth Billy Bass. But animatronics is fairly expensive, and mainly used as Hollywood props in big-budget movies. The toy industry is capable of delivering inexpensive animatronics, from Chatty Cathy, Teddy Ruxpin and Furby to Robosapien and Pleo, but each one of these robot toys is a standalone success derived from cheap manufacturing and mass marketing, rather than the result of an ever-more capable practice. Nevertheless, I expect that eventually some kind of animatronic toy platform, like the animated Chimp and Elvis heads from Wowwee, will eventually "catch," and profits will drive efficiency until humanoid puppets are reasonably priced.

These two streams will then become symbiotic, where the best salesbot software running on inexpensive humanoid animatronics can start to displace human salespeople from car showrooms and furniture stores. But, these saledroids will still be empty puppets unless inhabited by low wage workers over broadband.

The third line has to identify and attack the core problem of AI, that sentient life forms are several orders of magnitude more complex than the most complicated systems designed and built by humans: Our Software. Building software using best engineering practices always bogs down between ten and one hundred million lines of code, before it becomes unmanageable by human corporations. Assume that a sentient animal mind would take tens of billions of lines of code, just like bodies are made of tens of billions of living cells cooperating to form a whole. In order to understand how nature could design systems of far greater complexity than human engineers, we must focus not on simulating human cognitive faculties, nor on trying to understand the brain, but on the process which can design such minds and brains.

Through work on evolutionary and co-evolutionary machine learning, we have identified missing principles of evolutionary theory as implemented computationally. We've developed systems which surpass human designs of sorting networks and cellular automata rules, shown how co-evolving robot bodies and brains could be automatically manufactured, and developed new incentive structures which can motivate a community of learners to become each other's teachers.

The third stream, sentience, I believe won't be programmed directly, but will be a result of successfully replicating how evolution has achieved an open-ended self-organizing process on a computationally universal substrate. Once sentience is achieved, it will happily reside, and earn a living selling used cars, in the aforementioned electronic puppets.