The Norvig points were particularly thought-provoking. How do you tell if google is returning the correct result.
The other is that starting a unit test without knowing what you are aiming for (e.g., the puzzle solver) you may not get there, and TDD won't help you.
I've seen major benefits from automated testing, and I'm convinced they're a net win (especially for maintenance), but I also think that being dogmatic about any particular testing methodology (TDD, etc.) is going to be counterproductive sometimes.
* I don't remember if it was from the same book, PAIP, AIMA, or something else - I've been bouncing around in several AI and Prolog-related texts lately.