dave glasser [08.14.06 08:07 AM]

You mean like "solve these CAPTCHAs for us to see porn"?

  Tim Connor [08.14.06 09:49 AM]

You do realize, how big this could be if they played it right, right? I mean, in theory, this could become THE web currency. If it got adopted wide enough as a micro-payment solution and they allowed exchanging it in other ways, you could see the beginning of a lot of cyber-punk genre economic trends in this. I don't think they'll take it that way, but if they are smart they will see the potential - becoming the banker for the processor cycle based economy.

  BPAndrew [08.14.06 10:02 AM]

In regards to the captcha busting, its been attempted - just like paying people to Digg up stories on www.digg.com. They don't allow stuff like that.

Audio transcribing and text translation are the two examples of what could be done on AMT. However more people need to be thinking about AMT and ways to outsource fractions of work there. "AMTSense" could be what gets people interested and thinking about these business solutions.

Thanks for the link btw!

  jonblock [08.14.06 12:13 PM]

Interesting model. Unless the term has already been taken for something else, might this be considered a "Distributed Mechanical Turk" (DMT)?

One closely-related, early-Internet example of the DMT notion ("You can use this service, if you agree to help make it useful for someone else.") comes to mind right away:

The Usenet Oracle (now known as the Internet Oracle -- Wikipedia entry).

Although two-way participation was technically optional and interactions were more typically humorous than actually useful, the concept is very similar to this DMT suggestion, and it dates back to 1989. The idea was that you would e-mail a question to the Oracle, and you could expect an answer within a day or two. In exchange for this service, the Oracle would send you somebody else's question to answer.

It wasn't particularly Web 2.0 (even accounting for the fact that the necessary technologies didn't exist then), but the user-generated content model certainly worked.

  Jeroen Wenting [08.16.06 12:58 AM]

I wonder how this will ever be policed. Who's going to decide whether the work was performed, and thus payment made? And if it was performed, who's to decide that it was performed to sufficient quality to warrant counting as payment?

The "seller" could just classify everything given to him (no matter how good) as being substandard if there's no oversight, and the buyer could claim he did work that never happened and threaten the seller to file complaints unless he gets the goods anyway.

For an escrow agency to work as a middleman, they'd need a massive amount of resources with many skills to judge the work submitted, which is hardly a good businessmodel.

  Madison adams [08.16.06 06:13 AM]

In regards to the captcha busting, its been attempted - just like paying people to Digg up stories on www.digg.com. They don't allow stuff like that.

Audio transcribing and text translation are the two examples of what could be done on AMT. However more people need to be thinking about AMT and ways to outsource fractions of work there. "AMTSense" could be what gets people interested and thinking about these business solutions.

  BPAndrew [08.16.06 07:29 AM]

"I wonder how this will ever be policed. Who's going to decide whether the work was performed, and thus payment made?"

To ensure quality there are a few techniques being used

1) mass sampling. By asking more people the same question you can use trends for correct answers.
2) tiered sampling. Having levels of people answering and reviewing answers is another method. Using the qualifications system for MTurk you can reward and promote good works and block bad ones.

The API is pretty flexible, you can implement any technique you wish to ensure the quality of the work. Altho most of the techniques require issuing more jobs(HITs) to the system, in the end you are still paying less than having the same amount of full time people on staff.

  Jeff Barr [08.16.06 05:14 PM]

Wow, great discussion. I'd like to make a few points:

1. Each Requester (the organization putting work in to the system) has the ability to accept or reject each piece of work. Accepted work is paid; rejected work is not.
2. The Amazon Mechanical Turk tracks the acceptance rate for the work done by each worker. Workers with low acceptance rates can be disqualified from doing certain types of work.
3. A commonly used quality control method involves plurality. The same work unit (HIT) is sent to more than one Worker. A "majority rules" model is then used to check results. This could be 2 out of 3, 3 out of 5, and so forth.
4. We reserve the right to remove work from the system under certain conditions.

  Jörg Müller [08.23.06 12:03 PM]

Open Source for example inherently tends to be WORKWARE, since even the user's bug reports are internet powered "outsourced" work.


BTW: The tag cloud on the right hand side is aready barely usable. Maybe the mapping algorithm could be refined to normalize the scaling. If the computed scaling exceeds the maximum font size, the tag in question might be set in a darker color.