What Mike and I accomplished during Spring break can only be described as a sprint. We hammered out an initial release of a decent contribution to the Python community in two days. In fact, we started preparing for the next release by rewriting everything from scratch to more easily support new collaborative filtering models and make our current models more accurate. There's another project we started that I'll put online in a week or so.

Our long and focused coding sessions resulted in what I have only ever experienced after playing a single video game for prolonged hours. I fall asleep thinking about code and math; sleep is interrupted by thoughts about code and math; the day begins (often in the late afternoon) with thoughts about code and math. It takes a while to wear off.

More specifically, here are some thoughts that keep popping up since we first released consensus:

  • Why don't any of the models presented in research papers work as presented? We've had to modify almost all of them.
  • Why is our refactored code an order of magnitude slower than our initial naive code?
  • Why did using numarray for churning numbers on our dataset make things slower, not faster?
  • If AudioScrobbler's dataset is so large, why are their all-time top recommendations for me so unstable?

Anyway, I hope that posting these thoughts will get them off my mind, because I'm ready to think about other things for a while.