This is a beatiful idea – I am going to blog it. Very interesting. It is somewhat similar to Google Answers and other similar "expertise referral" systems, although it is higher-level – enabling you to write programs that transparently include human agents in their logic. Brilliant.
Exactly. I saw the same relation to Google Answers.
CHI might also need a function to automatically split a problem into several parts, itself again a task a human could perform. This lowers the level of expertise needed (like a Google Answers expert), as the sub-parts could then be handled using the CHI question/answer approach in a sort of recursion. Like this:
function getComplexAnswer(string question)
person anyone = find()
array subQuestions = anyone.splitQuestions(question)
array answers = array()
foreach subQuestion in subQuestions
... person anyoneElse = find()
... answers.add anyoneElse.ask(subTask)
anyone = find()
string answer = anyone.combineQuestions(question, answers)
anyone = find()
boolean soundsGood = anyone.ask("Does this answer sound good?", question, answer)
loop until soundsGood
So I as CHI helper, interpreting the program, would get only simple small tasks.
If the question would be "When did the stock market crash most heavily during the dotcom bust, and who was responsible?"
Then John would get the task to split this. He would return
"When did the dotcom stock market crash most heavily?"
"Who was responsible for the dotcom crash?"
Peter and Tom would answer these separately, as would Frank and Ann. The answer would need to be compared to validate.
In the end, Mary would merge the sub-answers to a natural-language in-deep answer of the combined question. As in:
"The stock dotcom market most heavily crashed during April 2000 and responsible were Greenspan, as well as hobby investors in a frenzy."
Possibly, this splitting of questions into sub-questions could also be triggered manually by CHI helpers and/ or be handled automatically behind the scenes.
Very nice Idea Philipp – if something like this could really be made to work you could do some very cool stuff with it.
It reminds me of something I heard about spammers doing a while back to defeat CAPTCHAs (those distorted text images used to tell humans and computers apart). Perhaps your heard about it – To get past the CAPTCHAs the spammers set up a free porn site. To enter the porn site people had to enter the text in a CAPTCHA – which the spammer had captured from somewhere else. Thus the spammer got a human to do the hard image processing task for his spam script in an automated way. With built in rewards for the human. Basically a very limited form of your system. I must say I thought that had a certain kind of genius about it, even it it was an evil genius!
I think a system like CHI would be most successful in this kind of image analysis area. The example you give of sorting a list of faces by age could certainly be made to work well, but I am not so hopeful re: the example of answering a complex information retrieval query like you discuss above. Firstly it would be hard to integrate something like that into a program, and secondly it requires a pretty high level of commitment from your human volunteers – probably best served by Google Answers. I suppose there might be applications where the task is vast but can be effectively split up.
I think where CHI could really shine are the places where the questions can be answered with a simple Yes/No or a few words, and are easy for humans, but would be intractable for a computer – like image processing, or making subjective or common sense decisions, etc
The CAPTCHA people have a related idea on their website – The ESP Game – espgame.org/ – It's couched as a game, but the purpose is to harness humans to do image tasks.
Mark, you are absolutely right with your analysis. Also, now that you mention the evil CAPTCHAS scheme, I realize how much it was a CHI application. You've also mentioned something my friend Alex was talking about to me regarding CHI – that you could pack it into games so more people will freely do it.
You are also right the biggest task – and biggest potential problem – is figuring out how to split every complex task into simple sub-tasks. Especially so that it will not rely on experts like the researchers at Google Answers. I'm uncertain but hopeful there are many applications where that too is possible.
For now, at least I'm confident there are applications in the areas you mention – like image processing. It will be always then very interesting when the result is more complex than the sum of answers. I hope to find more applications, if only theoretical at this moment, which stress-test the concept.
wow, I didnt know the evil CAPTCHAS plan..so i was helping spammers, damn..uh..wow, actualy just kidding but I didnt know half the stuff you just said, until of course I read this article wow, i should stop being so annoying
Let's get together and work on it. See super-brain.org
I think I have a piece to your puzzle:
The power of QUESTIONS. communitywiki.org/en/ThePowerO ...
I think you want to make use of mass analysis:
It is all superstitious nonsense. You are better off in not believing in things like fung shui and the lot. Wise up. And grow up.