Watson and the future of AI
January 31, 2011 by Hans Moravec
Radical roboticist Hans Moravec, former director of the Mobile Robot Laboratory at Carnegie Mellon University, expanded our imagination with his vision of future robots as our “mind children.” Now he’s revolutionizing industry with his enhanced-vision mobile robots. We asked him to help us put Watson in perspective. Full disclosure: Ray Kurzweil is on the board of directors of Seegrid Corporation. — Ed.
Let’s take a moment to lift our noses from the grindstone and reassess the future. Apropos is IBM’s upcoming artificial intelligence spectacular, pitting “Watson” (a 90-server, 80 teraflops, 15 terabyte supercomputer, doing natural language disambiguation and question answering from a huge mixed-format internal database — no Internet), against the two top previous human winners of the game show Jeopardy!
In 1997, IBM’s supercomputer Deep Blue defeated world chess champion Garry Kasparov, in a world first. Now small computers can do it, and chess masters learn from them.
Chess and natural-language question answering are two points on a scale of AI and robotics difficulty that runs in the opposite direction of difficulty for humans. Skills easy for humans are hard for computers, and vice versa.
We do effortlessly and effectively things that were long a matter of survival for our ancestors, the machinery for those skills having been ruthlessly optimized. Recent cultural tasks, on the other hand, often recruit ill-fitting survival skills in unnatural and often hugely inefficient ways.
Computers lack powerful specialized skills, they just perform simple operations in long sequences specified by programs. This general-purpose neutrality seems powerful when applied to tasks humans do inefficiently, but weak when imitating natural survival skills.
As computers became increasingly powerful over the last 60 years, successful programs for ever harder tasks gradually became possible.
Fully intelligent machines will change the world
Seegrid’s robotics technology focuses on the very hard end, where even the most basic functions, like our visual route memorizer, are barely possible. But that’s today.
In coming years, extensions of our techniques and more powerful computers will allow the robots to learn routes that are just pointed out (on plans, or on location), to classify what they see, to find, pick and put things, everywhere, not only in warehouses.
Several extrapolations point to about 30 years for our robots to see fully as well as people. By then the other, easier, parts of the AI problem will also be perfected, and we will have fully intelligent robots. Thirty years may seem a long time, but today’s Seegrid system is a product of already over 30 years focused research. This is the halfway point.
If the state of progress seems disappointingly modest, note that the second half of an exponential curve greatly dwarfs the first half, as it amplifies hard-won early gains.
Fully intelligent machines will, of course, change the world. Most exciting to me is that they will open the universe to exploration and discovery on a scale unimagined now.
I am doing everything in my power to be fully part of that, with or without future medical advances (inevitable though they be, and invite you to think that way also.