A predictive compression engine. Fun and educational to work on, testcases are easy to write (compression followed by uncompression is the identity), lots of existing papers and code, and if you write the center part well, you can accept code drops of specialist agents, taking advantage of their expertise in some situations without allowing them to degrade your aggregate performance in other situations.
Applications: a shell that predicts what the user will type, finding patterns in logfiles, roshambo and similar games, predicting wiggly graphs, Schmidhuber's artificial curiosity, probably other things too.
So, the London community is arranging a Hackday where some of us will get together and code. In order to ensure we work on the awesomest idea(s) possible, we decided to ask LessWrong to add to our list of candidates. So here is the question:
What could a few developers do in a day or less worth of coding that will be awesome? Also, as a way of checking calibration, you can give your estimate for how long such a thing would take to build.
Note: While we will take ideas and voting here into account, there is no guarantee that we will actually end up choosing one or more of them.