The two hundred people packed into a small screening room in Midtown Manhattan on a recent Tuesday night made quite a throng. Engineers, venture capitalists, and entrepreneurs sipped Sam Adams and nibbled bits from a fruit plate. They were there to learn about CrowdControl, a New York startup that is melding human workers with artificial intelligence to create the next paradigm for global labor: crowd computing.
The crowd filtered into the theater, and Kirill Shenykman, a venture capitalist who had recently led a $2 million investment in CrowdControl, took the stage. “What we are trying to do is to transform human labor into something that scales like software,” he explained. “We’re trying to take people and make them into bits.”
A few, dark chuckles went through the crowd. With his deep voice, slight Russian accent, and coiffed silver hair, Shenykman seemed like a bit of a James Bond villain describing a master plan for world domination. “I don’t mean that in a negative way, a diminutive way,” Shenykman said, waving his hand. “But just as Amazon can provide computing power on demand to a growing startup, we want to be able to offer an elastic marketplace for human labor.”
CrowdControl takes large complex jobs and breaks them into tiny pieces, then sources the piecework out to millions of micro-task workers around the world.
The company’s founder, Max Yankelevich, joined Shenykman onstage. “What we are doing is tapping into the world’s cognitive surplus,” he said, referencing a concept first laid out by digital intellectual Clay Shirky. “When you stop to think about the amount of brain power we have on demand, it’s kind of staggering. If we wanted to, with all the available excess on hand, we could recreate Wikipedia from scratch in a single day.”
– – –
The star of the show that night was Mechanical Turk, a service Amazon initially created in 1995 to deal with its own problems sorting its massive inventory.
“We came at it programmatically from every angle we could think of, but it wasn’t working,” said Heidi Bretz, a business development manager at Amazon. “We would have a pair of red shoes listed on the site, with a description for a red sofa, priced like a different pair of rouge shoes. It was a mess.”
Eventually the company conceded that there is some work that humans are simply better at doing than machines. So Amazon decided to offer the work up to its customer base. They would post a job to their site — sorting the color of shoes or types of chandeliers — and web users from around the world picked up the work, often for just a few pennies per chore. The result was such a success that Amazon decided to open up the marketplace to other companies. “When things work internally, we like to turn around and sell them as a tool,” said Bretz.
The name for this platform came from The Turk, a chess-playing automaton which cut a swathe through the ranks of 18th century gamers in Europe and America. It was only decades later that this brilliant thinking machine was later revealed as a hoax: a chess master hidden inside a box. Amazon’s system attempts a similar sleight of hand, organizing people from across the globe to work as efficiently as a giant machine.
“The idea is you might sit down to a computer, log onto the web, order up a task, just as you would with any other piece of software,” says Prof. Rob Miller, who works on crowd computing at MIT. “You get the data back, all without ever realizing there was a human being on the other end of that transaction.”
Anyone can put a HIT — human intelligence task — onto Mechanical Turk, and anybody else can do the work. But the complexity in this kind of system was limited. “They should call them SHITs,” Shenykman, the investor. “Stupid Human Intelligence Tasks. With CrowdControl, we’re going to go much further.”
Who are the humans completing these tasks? They began as bored Amazon shoppers looking for a little spare change, first in the United States and then in India. The the labor pool has grown rapidly in the last five years. Chances are you’ve done some micro-tasking yourself in the last year, probably without knowing it.
The history of crowd computing could be traced back as far as the end of the 18th century, when the British Royal Astronomers distributed spreadsheets by mail, asking the crowd to help them create maps of the stars and the seas. It reached its height in the United States during the 1930s, when the government employed hundreds of “human computers” to work on the WPA and the Manhattan Project. The word “computer,” as it applies to our modern-day devices, came directly from these mathematically-minded men and women.
With the emergence of the modern day microchip, using large crowds for mechanical computation declined steadily in the second half of the twentieth century. But as the volume of data online grew, it became clear to companies like Amazon and Google that there were some things humans were simply better at doing than machines.
The advent of Mechanical Turk spawned startups like the Silicon Valley-based Crowdflower and Odesk, who offered to manage the outsourcing of work to micro-taskers around the globe. But the idea of people in third-world countries doing tasks for a few pennies raises all kinds of fair labor concerns. At the CrowdControl event, a number of audience members raised their hands to ask how Amazon could justify paying workers so little. “This is piecemeal work and the wages are set by supply and demand,” said Amazon’s Bretz. “We are just the platform.”
Crowdflower is careful to foreground the work it has done with companies like SamaSource, which relies on a pre-screened group of marginalized workers — women, youth, and refugees — who are living in poverty. But perhaps the most interesting class of micro-task workers contributing to the brainpower of large-scale crowd computing are the ones who don’t realize they are a cog in the machine.
Have you visited a website in the past month or two that showed you a bunch of distorted words and asked you to type them into a box in order to prove you were human? This security measure, known as a captcha, is meant to screen out spam bots from signing up for and abusing services like Gmail or Facebook. But the words displayed on the screen aren’t random. Often they are pieces of text from books that Google is trying to digitize. By correctly answering the captcha, you’re helping Google to complete its project. Recently the company began adding street signs from its global mapping project to captcha fields.
There exists another group of laborers, the growing ranks of casual gamers interested in buying some virtual goods. The players in a single game, World of Warcraft for example, have put in over 10 million hours of manpower fighting virtual battles. Companies like Crowdflower have tapped that audience by offering to exchange virtual goods for their labor. Instead of paying 99 cents for that shiny new sword, a gamer can help translate a sentence or organize a group of objects. It is the realization of Clay Shirky’s dream, that humanity’s cognitive surplus might be turned into productivity.
“More than half of Crowdflower’s workers are now made up of these gamers,” said Miller. “How do you decide on a standard for fair wages when people are doing your work for fun?”
– – –
While the HITs performed on Mechanical Turk have so far largely been restricted to simple, objective tasks — sort the shoes from the sofas, translate this sentence from English to German, or, rarely, draw a picture of a sheep — many in the field see greater possibilities for crowd computation. “People are starting to ask what kind of creativity the crowd is capable of,” says Prof. Michael Bernstein of MIT. “Ask a simple human crowd to design a chair, and they are terrible at it. But structure that workflow with the help of a computer, take the product and iterate on it multiple times with different crowds, and you can get some quite interesting results.”
In a study entitled “Cooks or Cobblers,” two professors from the Stevens Institute of Technology gave a crowd of 1,047 people a very complex task: design a chair for children from scratch. The first edition, created by a section of the crowd, was predictably terrible. But the study, which took its name from the old Chinese proverb that “three cobblers combined make a genius mind,” took that initial product and passed it on to another portion of the crowd, who refined the chair, and passed it on to another group for their contribution. With each generation, the object improved.
Others have found the power of iterative improvements through the crowd. Amazon found that when it asked Mechanical Turks to improve the length and grammar of review users left on products, they could actually boost sales of those items. Some MIT professors cooked up a program called Soylent — because there are people inside! — which integrated Mechanical Turk and Microsoft Word to provide on-demand copy editing via the crowd. Major sites like Facebook and Twitter rely on the crowd to power the translation that spreads their service around the globe.
The next wave will be crowd computing that works in real time. In the video below you can see Jasmine, a study participant, at her home cooking dinner, and in the mood for soup. The tricky bit is, she’s blind, and can’t remember which can in her cabinet was the coconut milk. Her sense of touch alone won’t help her solve this problem, so she pulls out her iPhone and opens up the WizViz app, a program created by crowdcomputing researchers at the University of Rochester. She snaps a picture of the cans, and in less than a minute, gets back an answer: it’s the can on the right. A human worker halfway around the world got her query and answered it in real time.