Amazon’s new Artificial Artificial Intelligence Ryan 03 Nov 2005

12 comments Latest by Bill Seitz

Amazon’s new “Mechanical Turk” lets people answer questions that are better suited for humans than computers. These questions are called HITs.

HIT stands for Human Intelligence Task. These are tasks that people are willing to pay you to complete. For example a HIT might ask: “Is there a pizza parlour in this photograph?” Typically these tasks are extraordinarily difficult for computers, but simple for humans to answer.

Here’s a sample HIT: You are presented with the name and address of a business as well as a set of photos taken along the street where the business is supposed to be located. Your task is to identify the best photo of the business that is listed.

Pretty interesting stuff — especially from a “less software” perspective (screw the software, just let the humans do it).. Read more about it.

12 comments (comments are closed)

Don Wilson 03 Nov 05

Very good idea, but the pictures they’re using are far from decent.

Nathan 03 Nov 05

Agreed, the photos are terrible. I just spent like 10 minutes using it and I didn’t even earn .03, damn I needed it too. Great idea Amazon, but you need some good photos first, and better yet photos that match the description. Pretty interesting though.

Jamie 03 Nov 05

Wow! Interesting stuff. But what about our Robot masters? Will the prophesy be fulfilled!?

Zach 03 Nov 05

Theoretically…

I hone my photo processing skills for weeks and get to the point where I can crank out 200 photos an hour, (less than 20 seconds each) surely fame and fortune will follow.

Or maybe $6.00 an hour? With a best case scenario? I love the idea but it seems they may be aiming a bit low. I know it’s not exactly skilled labor but it seems you’d have to offer a rate competitive with what you’d pay someone to do the same work in an office. Your savings would be in other areas.

Am I missing the point?

Dave Churchville 03 Nov 05

This sort of creates a global economy for computer-based grunt work. If you’ve got a computer, you can earn 3 cents for each minute of your day.

I can just picture a high-tech sweatshop with 3 dusty computers, and 20 hungry former dot-commers fighting to split the $6/hour premiium.

Come on, man, I need a HIT…

chu 03 Nov 05

Zach, I think the idea is that $6 p/h would be a pretty good rate in China or wherever the cheapest labour is these days though with zero worker’s rights and protections.

Jeff 03 Nov 05

Can this be used to defeat CAPTCHAs? It looks like any organization can become a Requestor. A person could set up a stream of HITs that are CAPTCHAs, and then feed the answers back to whatever service they’re trying to spam; this thing has an API as well.

Adam Smith 04 Nov 05

I had a similar idea a few months back, but I didn’t think to give it an awesome name like “artificial, artificial intelligence”.

I wanted to allow computers to programmatically access data that was only available through human-interfaces. I started making a wrapper API that sits on top of the ip-relay.com applet. With all of the difficult text-to-speech and speech-to-text taken care of for free it should be fairly easy to make simple APIs for your local Chinese restaurant or family dentist — so long as you specify a simple enough grammar.

A Nonie Moose 05 Nov 05

Amazon might be on to the unspoken maintainance problem these pic-based business searches have.
Example: Joey’s Shoe Store happens to be located next to an Osco Drug. Osco Drug gets bought out, they abandon the space, and a Starbucks moves in. This process can and does now happen within days. What is Google’s/Amazon’s/Yahoo’s commitment to maintainence here and how do they know something changed?

Enter Amazon with a steady stream of human reviewers and they’ve got it covered (if they can keep the interest level)

Joseph 05 Nov 05

Jeff, great point about CAPTCHAs, that’s the first thing that came across my mind when I heard about Mechanical Turk.

But hey, this API can be used for more than just image analysis; what other tasks could you see building into an application? i.e. things computers stink at but humans can provide a quick answer?

Sam 05 Nov 05

Metadata. No one likes inputting good metadata. Micro-economics makes it viable. Image searches could be dramatically improved.

Google’s secret program to manually improve their search results with human response fulfills the same function. Humans get paid to look at and rate search results. De-spamming all of Google’s search results sounds impossible, but I imagine a few million dollars could manually de-spamify a fair bit of the popular searches.

Bill Seitz 07 Nov 05

hmm a Turing test…