Subscribe to
Posts
Comments

Archive for April, 2010

[berkman] Luis von Ahn on free lunches, captcha, and tags

Luis von Ahn of Carnegie Mellon University is giving a Berkman lunchtime talk. [NOTE: I’m liveblogging. I’m making mistakes, leaving stuff out, paraphrasing, getting things wrong. This is an unreliable record.]

Luis invented captchas, the random characters you have to type in to convince a web page that you are a human and not a hostile software program. (He shows randomly generated sequences that happened to spell out “wait” and “restart.”) Captchas are useful, he says, when you’re trying to prevent people from gaming a system by writing a program to enter data robotically. They’re also useful to prevent spammers from signing up for free email accounts. To get around this, spammers have started up sweat shops where humans type captchas all day long; it costs the spammers about $0.33/account. And some porn companies ask users to type in a captcha to see photos; the captchas are drawn from email account applications. Damn clever!

He shows some variants. A Russian asks you to solve a mathematical limit. In India one asks you to solve a circuit. Luis says these aren’t all that effective because compputers can solve both problems, but they’re still better than the “what is 1 + 1?” captchas he’s found on US sites.

He says that about 200M captchas are typed every day. He was proud of that until he realized it takes about 10 seconds to type them, so his invention is wasting 500,000 hours per day. So, he wondered if there was a way to use captchas to solve some humungous problem ten seconds at a time. result: ReCAPTCHA. For books written before 1900, the type is weak and about 30% of the text cannot be recognized by OCR. So, now many captchas ask you to type in a word unrecognized when OCR’ing a book. (The system knows which words are unrecognized by running multiple OCR programs; ReCAPTCHA uses those words.) To make sure that it’s not a software program typing in random words, ReCAPTCHA shows the user two words, one of which is known to be right. The user has to type in both, but doesn’t know which is which. If the user types in the known word correctly, the system knows it’s not dealing with a robot, and that the user probably got the unknown word right.

ReCAPTCHA is a free service. Sites that use it have to feed back the entries for the unknown word. About 125,000 sites use it. They’re doing about 70M words per day, the equivalent of 2-4M books per year. If the growth continues, they’ll run out of books in 7 years, but Luis doesn’t think the growth will continue, so it might take twenty years. (There are 100M books.)

(In response to a backchannel question, Luis tells the penis captcha story.)

The ReCAPTCHA system filters out nationalities, known insult terms, and the like, to avoid unfortunate juxtapositions. It’s soon going to be released in 40 languages. Google acquired ReCAPTCHA.

Q: When will OCR be good enough to break captchas?
A: I don’t know. We’ll probably run out of books first.

Q: Business model?,br>
A: Google Books gets help digitizing.

ReCAPTCHA “reuses wasted human processing power.” The average American spends 1.9 seconds per day typing captchas. We also spend 1.1 hours a day playing electronic games. We humans spent 9B hours spending in 2003. It took less than a day of that to build the Panama Canal. So, Luis switches topics a bit to talk about how to solve human problems by playing games.

First is tagging images with words. Image search works by looking at file names and html text, because computers can’t yet recognize objects in images very well.

Does typing two words take twice as long as typing random letters? No, it takes about the same time, he says. Luis says about 10% of the world’s population have typed in a captcha. The ESP game asks two people unknown to each other to label an image until they agree. The game taboos words that other players have already agreed on. The system passes images through until they get no new labels. They’ve gotten over 50M agreements. 5,000 players playing simultaneous could label all Google images in a month. Google has itsown version; Google has an exclusive license to the patent.

Q: Demographics?
A: For my version, average age is 29 (with huge variance), evenly split between women and men.

Q: Compared to Flickr tags?
A: Only a small fraction of Flickr images have useful tags. The tags from flickr tend to be significantly more exact, but also significantly noisier (e.g., a person tagging an image in a way that means something idiosyncratic).

Q: Bots?
A: Yes, we don’t want you to wait for a partner, so sometimes we’ll give you a bot that replays the moves a human had made with the same image.

Q: Google Images benefits from its version of your game. Who benefits from your version of the game?
A: No one.

For some images, guesses change over time. E.g., a Britney Spears photo five years ago got labels like britney and hot. About two years ago, the labels changed to crazy, rehab, and shaved head. Now they’re back to britney and hot. By watching a player for 15 mins, you can guess whether the player is male or female with 95-98% accuracy.

Why do people like the ESP game? Sometimes they feel an intimacy with their partners. They have to step outside of themselves to make the match. They can have a sense of achievement.

He ends by saying that the about the same number of people — 100,000 — have worked on humanity’s big projects, e.g., pyramids, Panama Canal, putting a person on the moon. That’s in part (he says) because it is so hard to coordinate large numbers of people. Now we can get 100M people to work on something. What can we do?

Shirky’s myth of complexity

Clay Shirky has given us a surprising number of Internet myths. And by this I mean not falsehoods but the opposite: Broad, illuminating ways of making sense of what’s going on. For example, Clay’s post about the power law distribution of links in the blogosphere (based on research by Cameron Marlow) changed how we view authority, fame, and success in the Web ecosystem, and provided the structure within which Chris Anderson could point to the Long Tail. And Clay’s Ontology Is Overrated made clear that a change in how we categorize our world affects very real power relationships; that essay was highly influential, including on my own Everything Is Miscellaneous.

Clay’s new post — The Collapse of Complex Business Models — gives us a broad way of understanding why those who used to provide us with content will not be the ones who give us content in the future…and why they cannot fathom why not.