Watson, IBM’s astonishing Jeopardy!-playing supercomputer, will take on two top human players in head-to-head competition televised on February 14-16. I predict Watson will win, and I believe the competition would be closer if the machine played against only one person rather than two. Let me explain…
I had the chance to watch Watson in action last week at Lotusphere (my keynote at the conference is here). In addition showing the game show’s normal displays (the board showing all possible topics and dollar amounts, and the text of the question being asked) the Lotusphere screens also presented Watson’s ‘thinking’ during each question. This consisted of the top three responses it was mulling, along with its estimate of the probability that each response was the right one.
In the (low-quality) photo below, which I took during a rehearsal session at Lotusphere, Watson is considering three answers to the question being asked: “Thomas Hardy,” “Oscar Wilde,” and “Tolstoy” (sorry, I forget what the question was). It’s not very confident, though, that any of these are correct: it thinks there’s only a 20% chance that Hardy is the right answer, and the other two are even lower.
In this circumstance, Watson is not going to try be the first to buzz in — to signal to the host that it wants to answer the question. In Jeopardy!, the first player to buzz in once the host is done reading the question wins the right to answer it. If the answer is correct, the player gets the $$ associated with the question; if the answer is wrong, the player loses that $$ amount and the other two players get a chance to answer. So the three critical factors for winning the game are speed, smarts, and accurate estimates about how smart one is.
As I was watching Watson rehearse, I noticed two things and concluded two others. First, I noticed the supercomputer knew a lot, and also knew what it knew. In other words, it very often came up with the right answer — not always, but very often — and when it did it also typically assigned that answer a high probability of being correct. The rehearsal only lasted a few minutes and so didn’t provide a big sample, but I saw no cases where Watson was falsely confident or falsely unconfident.
The second thing I noticed was that the supercomputer was (duh) fast. It generated its answers and associated probabilities very quickly after being given the text of the question. Watson’s answers and probabilities appeared onscreen almost immediately after the question was revealed, and well before the human host was done reading the question aloud.
This matters a great deal in Jeopardy! because players (human or machine) can’t buzz in until the host is done reading the question aloud. If my casual observations are at all accurate, Watson is ready well before the host is done reading – it knows if it wants to buzz in, and it knows what it wants to say. And (again, if my observations are accurate) it’s rarely too confident, or not confident enough.
Now, there’s no way a human is going to buzz in more quickly than a confident Watson. When the players get the signal that it’s OK for them to buzz in, silicon and fiber optics are just always going to be faster than muscle and nerve fiber. So a confident Watson is always going to get to answer first.
Given this permanent speed advantage, the outcome of the game basically depends on only one thing: is Watson confident/smart enough on enough questions? It doesn’t matter how smart or fast its human opponents are; it only matters how smart Watson is, since it’s always fastest. And my first conclusion is that it’s smart enough. As I wrote earlier, the Watson team has done an astonishingly good job at building a Jeopardy!-playing supercomputer — one that can answer natural language questions on a huge and unspecified range of topics.
This is a true milestone, and I’ll have a lot more to say about it later. For now, let’s stick to the upcoming competition. Watson will buzz in first on all the questions where it has high confidence, and will answer them correctly. The two human players will be left to fight it out on only the questions Watson couldn’t answer. They’ll be fighting over the machine’s table scraps.
If we assume the two human players are of roughly equal ability, they’ll each get about half of Watson’s leftovers. This is great news for the computer; it lowers the percentage of questions it has to answer in order to have a reasonable shot of winning, from somewhere around 50% to somewhere just above 33%. To oversimplify a bunch (and to ignore wrinkles like the Daily Double and Final Jeopardy!), the computer needs to get only a bit more than a third of the total $$ at play if it can be confident that the humans will split the other 2/3.
So my second conclusion is that the competition would be much closer if Watson were only playing against a single human champion. The supercomputer would still get to answer first whenever it was confident, but the lone human would get his chance every time Watson wasn’t confident.
Over time, of course, even this won’t be much of a competition. Watson is only going to get smarter over time, and it’s not going to get slower, so it will someday (probably soon) become the world’s best Jeopardy! player, just as computers are now easily the world’s best standalone chess players.
As I’ll write later, I find more good news than bad news in this. For now, though, I’m just going to pop some popcorn, plop myself on my couch, and watch the competition on TV to see if my predictions hold up.
What do you think? Do you agree with my observations and conclusions, or am I missing anything important here? And who do you think is going to win? Leave a comment, please, and let us know…