[tezzer]

I was working with a decoder for the LD-CELP voice compression algorithm. The machine it was running on was compute-bound and the playback stuttered a lot. I dug into the code, then went to the papers describing the algorithm, then went back into the code and modified a piece of it to be better tuned to the data I was processing. In the average case I probably made the code worse, but in my specific case I sped up the thing by 30x. That gave me confidence in a few things:

1) I can figure out issues in other people's implementations.

2) The best solution may not be the best solution for you, and my hunches on optimization might be correct

3) It's fun to look like a wizard to your peers