Some interesting tidbits gleaned from their conversation:
• Pandora programs in "sets" (something I'd come to suspect). This is to say, it tends to zero in on one significant attribute of your feed material -- the song(s) or artist(s) you used to set up your "station" -- at a time. This means Pandora will program several songs in a row that share that single attribute (say, instrumentation, or vocal style) before moving on to another attribute -- presumably this is designed to avoid too much stylistic jumping-around.
• Pandora employs music theory grads to chart the attributes of each song. Not to be snarky, but this probably explains why some of the songs are occasionally described badly or flat-out incorrectly ("major key tonality" for minor-key songs, etc) -- in my experience, theory majors don't always have the best ears. (They might have done better with musicology majors!)
• Pandora doesn't take lyrics into consideration at all. Of course, this is a valid and interesting approach, and part of the fun -- by focusing on purely musical similarities, it often suggests interesting but cool left-field connections -- but the service may ultimately be frustrating for people (like Lindsay) who listen much more intently to the lyrics than the music, and are more interested in finding new music based on lyrical similarity rather than musical similarity.
Anyway, you can download Friday's broadcast here. The segment with Tim Westergren starts around the 1:33:45 mark.