Abstract: User input on a computing device can be intelligently predicted using one or more learning algorithms. When user text input is received, an input repository comprising a plurality of text items is accessed. The plurality of text items includes a plurality of user-specific text items, which are classified according to a probabilistic model. One or more input predictions are inferred by applying the probabilistic model to the user text input. A suggested text input is presented using an output element of the computing device. The suggested text input is based on the user text input and one or more text items from the input repository. Additional input is received which overrides the suggested text. Any text associated with the user override is then classified within the probabilistic model, thereby updating the input repository based on the user override.

Abstract: Requests from multiple users for one or more shared resources can be dynamically throttled based on at least one allowance value determined for each user at a current point in time. One or more business functions or other such algorithms can be used to determine an appropriate allowance value for a user based on any of a number of appropriate factors, such as the cost of a resource, a value of the user, a request volume of the user, or other such factors. The values for each user can be fed into one or more algorithms to determine information such as request limits, burst limits, and costs per request. Such approaches provide for dynamic and fine-grained traffic shaping while allowing for variance of these limits or values over time.