100 predictions a day isn't enough to test if it works. Somebody needs to make a much bigger commitment of their time to make a machine learning application that really works -- when you look at what's holding machine learning back it's not the algorithms or the hardware requirements, it's that people don't want to do the work of creating high quality training sets and validating them.
At least, by making it a for-pay service, the likelihood of Google just shutting it down in the middle of the night is lessened (although still present). However, anyone competent enough to be passing data to an API can pass data to one of the many open source ML libraries that are available for many languages. I don't see the point.
* Training models can be rather computationally expensive. Especially if your business requires training new models very often, this can be prohibitively expensive to do in ec2, whereas the prediction API solves that for you.
* Just hooking up to an open source ML library isn't the whole story. You still need to do backtesting on different algorithms and do the parameter tuning, aka you need some machine learning know-how. The Prediction API does all this for you automatically and probably uses a much larger set of algorithms than you would bother to test yourself.
It's great to see these tools available on a cloud computing basis. Just make sure you read the ToS:
1.2. From Customer to Google. By submitting, posting or displaying any Customer Data on or through the Service, Customer gives Google a worldwide, non-sublicensable, non-transferable, non-exclusive, terminable, limited license to reproduce, adapt, modify, translate, publish, publicly perform, publicly display and distribute any Customer Data for the sole purpose of enabling Google to provide Customer with the Service in accordance with the Agreement.
Isn't that the standard lawyerspeak to actually run the service? If I understood correctly, when reading those you just need to check that it is limited to the service provided ("for the sole purpose of enabling Google to provide Customer with the Service").
100 predictions a day isn't enough to test if it works. Somebody needs to make a much bigger commitment of their time to make a machine learning application that really works -- when you look at what's holding machine learning back it's not the algorithms or the hardware requirements, it's that people don't want to do the work of creating high quality training sets and validating them.