Creating a marketplace for artificial intelligence

An alternative way would be for small companies/individuals to start producing small fragments of functionality and offering it in a per for use basis. (as in a cat recognizer in an image, or a recognizer for the word "hello" in an audiofile).

To create the above recognizer would require gathering data (such as hundreds/thousands of pictures of cats) and feeding it to a supervised learning algorithm like svm, boosting etc.

Thousands (perhaps millions) of people creating their own "classifiers" and charging for them (perhaps 1000 uses for 1 dollar) over the web might lead rapidly to seeing/thinking machines.

(A similar model can also be used to pose competition to entrenched monopoly software like mathematica/maya/autocad - thousands of people creating and charging for small bits of functionality)

As for the complexity, that's what the market would sort out. People would build on other's work, and everyone would get compensated for as much as their work is used.

Numenta is an example of the monolithic comapny doing all the work, and that's why you can't use or build on any of their technology yet (incidentally, they use bayesian type algorithms while i'm advocating easy to use standard algorithms that have many open source implementations - boosting, random forests, support vector machines, and maybe neural networks).

A startup opportunity would be for a company to host the learning algorithms, and just have people pay for and submit data from which classifiers (academic jargon for recongizers) can be built and returned. (The company would have to be trusted though as they could simply keep a copy of the created classifier and resell it).

and ATT ~ http://www.research.att.com/~ttsweb/tts/index.php#top

There is a recent startup doing image recognition using humans, on photos at flickr.

Search engines relied on letting a single algorithm go out and process data. There isn't yet an artificial intelligence algorithm that can just go out and do what it needs (unsupervised learning algorithm that works well). The currently available algorithms (supervised learning) require labelled data, and since there are millions of categories of things (that we as humans can percieve and so are useful to recognize), it would require at least a billion labellings of the data (1000 data labellings for each category).

The trouble with companies and groups engineers is that you would have to fund them, organize etc. and only a few groups would form (there are only an handful of video analytics companies probably, all of them probably do recognition of a handful of the same things). Besides, the system advocated above is pretty simple, hardly any engineering in the sense of specialized technical skill is required. You would just collect and feed data into the classifiers. Others would combine data from classifers into yet more classifiers.