Friday, November 11, 2005

Mechanical Turk

For many years, people have been talking about "software services". My objection has always been that this misses the point. If something is a service, hidden behind a service interface, you should no longer know or care about the delivery mechanism - software, firmware, human or white mice.

Amazon's launch of the Mechanical Turk service reinforces this idea. The idea is that human microtasks can be rendered as services and orchestrated into business processes, just as if they were software-based. Amazon refers to this as "artificial artificial intelligence", which is a nice conceit. It's also called Human Intelligence Tasks (HIT).

But not necessarily human. Some of the examples of services rendered via Mechanical Turk involve simple picture recognition, and could possibly be done by a trained animal or bird. (After all, homing pigeons are pretty good at direction-finding, and a hawk can spot a tiny animal in the undergrowth.) All they need is something they can peck or paw to record their answer, and they can earn enough to pay for that gilded cage on eBay.

Some people are sceptical about the economics of human microservices. For example, Patrick Tanguay [update: URL added] says
Problem is, with the prices they offer so far, you’ll be lucky if you can make 4-8$ an hour. Seems like after years of various people trying to crack the micropayment formula, Amazon is now bringing us micro-outsourcing where any third world person with access to the web can log in and start churning out answers for 3 cents a pop.
But in my view, the real interest of Human Intelligence Tasks is the potential for aggregating and composing them in complex ways with non-human services. Think of the security applications - get a hundred people around the world to vote whether an airline passenger looks guilty or merely uncomfortable, or to vote whether she looks enough like last year's picture. And we don't have to send every case out for human review - merely a random sample to help calibrate the accuracy of the machine, and to create fear and uncertainty for the bad guys. Think of the marketing and business intelligence applications.

And what about Internet search? I find I waste huge amounts of time using search engines, wading through pages that have nothing but keywords, cached pages from previous searches, links that no longer exist, pages with no significant content that have been cleverly designed to trick the search engines. I might be willing to pay someone to filter the search results for me - let's say five dollars for twenty minutes of reasonably intelligent work, results delivered within an hour. Or am I asking too much here?

Technorati Tags:

No comments: