I spent some time with Watson over the past week or so at various conferences and I came away impressed. This is IBM’s Jeopardy-solving computer. At the St. Louis Gateway to Innovation conference, IBM had a kiosk showing off its prowess. You had your chance to try to come up with the correct question (remember, the game gives you the answers) before the computer did. Now, I consider myself a fairly decent Jeopardy player: when I watch the show on TV, I can usually get the right question more than half the time. But that is sitting in the comfort of my own living room.
When it came to standing in front of the Watson kiosk, I choked. Big time. It was interesting, because you get to see how Watson “thinks” and formulates its replies. I could barely start typing before it had finished figuring out its (correct) question. The kiosk doesn’t contain the actual Watson computer: it is just a remote interface. You can see a picture of part of the kiosk here, and how it judges among three different possible responses, along with assigning confidence values to each.
Then, at the IBM Impact conference in Vegas, I got to attend one of the sessions where Bob Madey, one of Watson’s developers, was talking about their next steps. Watson isn’t a product: with all the processors and storage it isn’t exactly portable. And you can’t really buy it anyway. It isn’t quite a service yet, although IBM is working on commercial versions for healthcare and financial services, and eventually for other industries such as sales contact management. But what is interesting is how IBM is going to make money from their efforts with this research project.
Anyone who signs up for Watson (and you better believe there is a waiting list) has to agree to a unique arrangement: IBM will take 40% of the calculated savings from its use. So if Watson is helping to diagnose diseases (as was being shown in the session that I attended), the hospital and IBM will come to some agreement as to what the cost savings will be from this activity. It is an intriguing way to sell computer services. The client only pays if Watson delivers.
This differs from usual computer services types of deals where you pay by the CPU minute or the amount of storage or other resources consumed. But that wouldn’t do for Watson, because there are a lot of resources involved in processing its natural language queries and storing all the data it needs to figure out the context of the query and possible answers to your questions. And one other thing: IBM will own all the data that is fed into Watson. You just get the results, and that’s it.
There are 220 people involved in the Watson project and this large staff is needed. It took them more than a week to load in a new database for each Jeopardy build. For the medical apps they are currently working on, it still takes them several days to retrain Watson. Part of the problem here is in understanding how to retire old data, or to assign lesser confidence intervals as you learn new things about the particular subjects you are covering. One of the audience members at Impact raised that issue, and it was an important one. Another asked how long it would take IBM to develop the Watson equivalent of a WebMD.com that anyone could query its medical data online. While Madey was cagey about answering this, if Watson would be allowed to reply the machine would probably agree that was a reasonable goal. I started thinking of the HAL9000 computer that powers the spaceship in 2001. “Dave, that’s a really good idea.”
It was fun to watch Watson on TV last year, although none of the human contestants stood a chance and eventually it was a total silicon rout. I know a bit how those fellow carbon-based life forms all felt: it was humbling just to try my hand at a couple of questions. But it is exciting to see where IBM is taking the project, and I suspect we will be hearing more about this curious machine sometime soon.