It would be very interesting to know how Google measures on-the-job performance since you obviously can't talk about correlation without numbers backing it up. I seem to recall having read that on-the-job-performance is not as easily measurable than Norvig seems to imply but I guess that if they successfully managed to reach this Holy Grail they probably won't share the numbers, even less the methods.
"Our interviews are more to do with practical problem solving, not with puzzles and tricks. Our interview scores actually correlate very well with on-the-job performance: we are doing quite well at hiring the right people, we believe, and we work hard at analyzing the process. Peter Seibel asked me if there was anything counterintuitive about the process and I said that people who got one low score but were hired anyway did well on-the-job. To me, that means the interview process is doing very well, not that it is broken. It means that we don't let one bad interview blackball a candidate. We'll keep interviewing, keep hiring, and keep analyzing the results to improve the process." http://www.mv-voice.com/square/index.php?i=3&d=&t=16...