TECHNOLOGY FOR CHANGE

What's Next for Watson?

John Jainschigg | Date: 02-18-11 | Comments

IBM's Watson supercomputer won this week's "Jeopardy!" tournament decisively. But what does it all mean and what will be Watson's next challenge?

Table of Contents:

What's Next for Watson?
After four years of work, more than a year of building buzz and an impressive performance in practice trials last spring, it was no great surprise that IBM's Watson supercomputer won this week's "Jeopardy!" tournament, decisively beating human champions Ken Jennings and Brad Rutter.

Within hours, the blogosphere and popular press were cataloging "what it all means"�with predictable and amusing veers toward the Dreyfusian ("It's just a gimmick!") and Kurzweilian ("It's the singularity!") extremes. For a uniquely information-rich short article and slide show on Watson's history and futures, read Darryl K. Taft's Feb. 14 piece in eWEEK.

IBM's Watson supercomputer competed successfully against the top two human players on "Jeopardy!" this week.

Several writers, including Taft, note that Watson represents a change in the style in which human engineering horsepower is deployed to solve AI problems. In essence, they assert that a Jeopardy! win requires a Jeopardy!-themed, "We reach the moon"�style effort. I think this assertion is debatable. The "epic contest of man versus machine" paradigm has been with us since the industrial revolution (cf. John Henry). In more recent years, efforts like IBM's "Deep Blue" campaign to beat chess champion Garry Kasparov have taken similar form, at least in the popular imagination.

Philosophically, however, I think the point is very well taken that doing well on tasks like Jeopardy!�and more generally, on any task that smart humans would agree requires "smartness"�needs an iterative effort. First, you constrain the task (or select a constrained task). Then you watch people succeed at the task, think very hard about the task and advance a general architectural solution�acknowledged to be incomplete. Only then do you develop Version 1.0 of a specific solution (which usually stinks). Then you spend a long time testing the solution, tweaking the solution and developing heuristics around the solution, slowly raising its performance on task to competitive standards.

AI has really only been successful when researchers have taken this approach. Deep Blue; the DARPA Grand Challenge in autonomous navigation; the rise of "expert systems" in the '70s and '80s; decades of gradual, painstaking work on speech recognition and machine vision now bearing fruit on smart mobiles�all are triumphs of iterative, rather than "great leap forward" efforts by (relatively humble) engineers.

It's hard to avoid thinking that the success of this approach reflects the nature and semantics of "intelligence" at a deep level. I tend to think of it as a vindication (were such necessary) of Marvin Minsky's Society of Mind theory, which views intelligence not as a monolithic, generalized ability, but as the aggregate of numerous, co-evolved task-specific competences.

In a more general sense, the development paradigm we're talking about is Agile: Use a small team, think hard, release early and often, repeat until you win. Not rocket science, as they say�though arguably, it may be where rocket science comes from.

What also interests me about Watson�particularly as the machine has been described, so often, as a purpose-engineered solution�is that it really isn't. The development effort behind it was purpose-driven, no question. But Watson is not the ENIAC or the Cray 1, lovingly hand-assembled (with solder-blobs on the boards to show for it), running against a database of all human knowledge aggregated and encoded by hundreds or thousands of patriotic clerks. It's really just a big cluster of Power7 blades, running IBM Analytics software over a database derived from the Internet. A purpose-integrated solution, perhaps. But the magic is clearly in the software, not the hardware.

Next: Watson's Next Challenge >>