Google's Prediction API aims to convert every app into a smart app by making AI learning algorithms a commodity that you can buy off the shelf. It's an interesting idea that has reached version 1.2 and full commercial status.

Google has released the latest version of its AI API Google Prediction. In case you have missed what the Prediction API is all about you can supply it with some training data and it will try to learn the correct responses using whatever method it thinks is best. It is a black box approach to machine learning and now it seems to be working its way towards a fully fledged commercial release allowing you to analyse lots of data using Google's hardware and AI expertise.

The biggest change in the new version seems to be the activation of the "paid-for-mode".

As reported previously the service will be free to users for six months or until the they reach a cap of 20,000 predictions - whichever comes first. Free accounts will also be limited to 100 queries and 5MB trained per day.

After the free period Google will charge:

• $10 monthly fee per project. A project is defined by a unique key on the Google APIs console.
• The monthly fee covers 10,000 additional predictions per month. Beyond that, predictions cost $0.50 per 1,000 predictions. If you intend to make more than 40,000 predictions per day, please contact us. Google Prediction has an absolute limit of 60,000 predictions per day.
• $0.002 per MB trained (maximum size of each dataset: 100MB)

The big difference is that now the software includes the billing and for some users the six month free period is up. In addition ot these charges you also have to have a Google Storage account to store the training data and this is also charged for.

The other changes in the V1.2 API are fairly slight but a new JSON format for all requests and responses effectively breaks any code you might already have written - but it should be easy to fix. Some additional stats are provided - mean square error values for regression models for example and prediction scores are normalized to fit in the range 0 -1.

One of the things that the Prediction API does is to put a price on big AI. For example, the training needed to implement Microsoft's Kinect body part recognition algorithm used roughly 4Kbytes of data from one million images i.e. about 4Gbytes of training data which would cost $8 plus the charge to upload and store the data.

Of course there is no way to know that the Prediction API would necessarily use the same learning model and currently there is a limit of 100Mbytes on any training data file - so you couldn't actually do the job using the Prediction API, but it does make the whole idea seem affordable. In fact when you take into account the amount of time training can take 20 cents per 100MBytes of training data seems almost under-priced. On the other hand 50 cents per 1000 predictions seems over priced given how simple a prediction computation usually is.

Presumably things will settle down as Google discovers what makes a profit and what doesn't.

Related articles:

Google to start charging for Prediction API

Google gets into AI with an API

All About Kinect

Kinect's AI breakthrough explained

If you would like to be informed about new articles on IProgrammer you can either follow us on Twitter or Facebook or you can subscribe to our weekly newsletter.