Artificial Artificial Intelligence
Aug. 4th, 2006 10:26 am![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
jducoeurOkay, this has to be one of the most amusing technology plays I've seen in quite a while -- it'll be interesting to see if it works.
Amazon is currently in beta with a new tool called Amazon Mechnical Turk. Those with a good sense of the history of geekdom will get the reference -- it refers to the supposedly intelligent 18th century chess-playing "Mechanical Turk" robot, which actually had a human hiding inside of it. And that's exactly what they've built.
The idea is that some problems are hard for even the cleverest computer program, but very easy for a person. So they've built a set of Web APIs that allow you to submit requests to an "AI" that actually has humans hiding inside of it. Programs submit questions to the Mechanical Turk system, including the question, exactly what they're looking for, how much they're willing to pay for the answer, and how soon they need to get it. People come in the other end of the system, pick off questions, answer them, and get paid for it. Amazon takes a 10% cut of each question answered.
Remarkably clever -- if they can get the economics right, they may well have hit upon a powerful new business. At the moment, I don't think they have the pricing structure right: sampling a few questions on the site, it looks like they're all paying 1 cent for questions that take about a minute for a moderately competent person to answer, so they're underpricing it by at least a factor of five. But these things will tend to balance out over time: indeed, I see a future enhancement that adds auction-like mechanisms for providing answers, which would get much better pricing structures quickly. (And probably time-adjusted mechanisms, where the price starts higher and drops over time.)
There are some dangers, of course. For example, this system makes it straightforward to implement automated captcha-breaking, albeit costing a small amount of actual money. Most spam factories are loathe to spend any money at all, which will reduce the danger here, but this system will tend to drive down that expense to something pretty minimal. I'd bet that Amazon will be forced to implement policies against such things.
Mostly, though, I'm hugely amused (if perhaps slightly perturbed) by the phrase "Artificial artificial intelligence"...
Amazon is currently in beta with a new tool called Amazon Mechnical Turk. Those with a good sense of the history of geekdom will get the reference -- it refers to the supposedly intelligent 18th century chess-playing "Mechanical Turk" robot, which actually had a human hiding inside of it. And that's exactly what they've built.
The idea is that some problems are hard for even the cleverest computer program, but very easy for a person. So they've built a set of Web APIs that allow you to submit requests to an "AI" that actually has humans hiding inside of it. Programs submit questions to the Mechanical Turk system, including the question, exactly what they're looking for, how much they're willing to pay for the answer, and how soon they need to get it. People come in the other end of the system, pick off questions, answer them, and get paid for it. Amazon takes a 10% cut of each question answered.
Remarkably clever -- if they can get the economics right, they may well have hit upon a powerful new business. At the moment, I don't think they have the pricing structure right: sampling a few questions on the site, it looks like they're all paying 1 cent for questions that take about a minute for a moderately competent person to answer, so they're underpricing it by at least a factor of five. But these things will tend to balance out over time: indeed, I see a future enhancement that adds auction-like mechanisms for providing answers, which would get much better pricing structures quickly. (And probably time-adjusted mechanisms, where the price starts higher and drops over time.)
There are some dangers, of course. For example, this system makes it straightforward to implement automated captcha-breaking, albeit costing a small amount of actual money. Most spam factories are loathe to spend any money at all, which will reduce the danger here, but this system will tend to drive down that expense to something pretty minimal. I'd bet that Amazon will be forced to implement policies against such things.
Mostly, though, I'm hugely amused (if perhaps slightly perturbed) by the phrase "Artificial artificial intelligence"...