It has been a while since I have checked it out, but you give it training data (examples + values) and then it will be able to predict values for unseen examples. Eg. predict if a comment on a blog post is spam.

Actually the google prediction api is already really old; I don't know why it shows up now on HN