There was a time when all you needed was an implementation of an AI algorithm. Today you need one that can process lots and lots of data. Mahout is just such a library. It is based on Hadoop and a new version is available.
The Mahout library is a collection of AI algorithms implemented in Java and released as open source code. One of its important characteristics is that it is scalable and many of its algorithms make use of Hadoop map/reduce.
It currently has four general approaches - recommendation mining, clustering, classifying and "item set" mining. Recommendation and item set mining have direct application to the sort of web-based technique that is becoming so important. Recommendation mining takes users' behavior and tries to find other items they might like. Itemset mining takes data like shopping carts and tries to identify items that tend to occur together.
Version 0.4 has added new algorithms that are worth knowing about: spectral clustering, minhash clustering, a vector classifier, new naive Bayes classifier and additions to the number crunching routines such as a distributed Lanczos SVD implementation.
As most of the algorithms are implemented on top of Hadoop it is harder to get started than with a simple non-distributed implementation, but as the dataset gets bigger you can simply add more machines without having to rewrite your software. You can get started with a single machine Hadoop implementation.
Mahout is not only a great resource for research and learning but it could also be the basis of some real AI-based systems.
Hadoop: The Definitive Guide (Book Review)