Mechanical Turk can’t solve everything
I like it when I see Web 1.0 companies rollout new and innovative services. Last week, Amazon launched a new service, Mechanical Turk. According to their about pages, Amazon’s Mechanical Turk ‘solves the problem of building applications that have not worked well because they lack human intelligence’ – all those nay sayers of computers must be loving this!
Mechanical Turk is ‘artificial artificial intelligence’ – a bunch of humans perform the hard work, to give the impression that a computer is executing the required tasks. Arguably, humans are better at doing some tasks.
I played around with the service for last couple days, and pretty much all of the requests (called HITs) were for Amazon’s A9 service -which is going to face some of the similar problems that I spoke about in another post. Theoretically, there is nothing saying that they can’t expand Mechanical Turk so that instead of asking users to approve photos, they could be asked to actually submit photos in return for payment, but that still won’t solve A9’s issue of stale data completely.
As the guys at Business Week argue, there are many tasks that can’t be automated that can benefit from this, which is probably true. Michael Parekh has a good post on it here. But photographic mapping services are not currently one of them. At this stage, MT is actually highlighting the problems inherent in photographic mapping services. The costs are simply too high.
In the HITs I was working on, it was obvious that A9 obtained the business names from some directory and that the pictures were taken by a moving vehicle. Now they are trying to use MT to have users verify which of the many pics taken is the best for the business. In most cases, none of the pictures were acceptable at all, and if you think about it, it isn’t feasible to match photographs to directory name listings like this.
Back in 2000, when we were seeking additional funding for our company which was working in this area, Arts Alliance directed us to a complementary service that a Pages-Jaunes subsidiary was executing. They already took the photographs with moving vehicles and mapping them to gps co-ordinates. However, they changed the paradigm a bit. They obtained GPS co-ordinates for each building (I can not recall how they did this), and then simply took 4 pictures in the general vicinity of that building (one in each direction). Then when a request for a business address arives, they provide the photos attached to the co-ordinates. They are not trying to provide photos for each business, but photos of the general vicinity. Using this method, it would be easier for users to upload and attach their pictures as well.
The point is, companies out there have already tried this and either failed or had to lower their goals – Mechanical Turk will not make it easier for A9 either.
Mechanical Turk is definitely a cool service but the current A9 application is a poor example. Unless of course, Mechanical Turk is trying to find out how poor A9’s current process is. If you think about it, A9 just ran the cost of obtaining the pictures, paying me $.03 to check them, all for nothing, since the pictures were unusable. In the Web 2.0 world where many revenues are based on volumes of small increments, $.03 is a lot of money and as Paul Kedrosky notes, $.o3 may still not be enough…
In general I think that this is an innovative service and am glad that it came from one of the old Web1.0 darlings, but after reading many blogs about this with views ranging from skeptical to over enthusiastic, I’m still at a loss to think of really good examples that can actually afford to pay end users sufficiently.