The example he works with is writing a driver for Tokyo Tyrant in java. He begins by writing a high level test of what he eventually wants to be able to do. Instead of implementing it though he comments it out and then starts to play around with potential implementation code inside the body of the test. Once he's finished playing he then gradually refactors code from the test body into the implementation until he's got everything he needs and uncomments the test.
It may well be the case that by the time you've finished exploring that you've realised the test you've written isn't very relevant. This is fine, the purpose of the writing the initial test is to guide your exploration in a productive direction. Once you have some implementation code you can always go back and reframe the test according to your new understanding.
All my actual programming work here consists of adding printfs at strategic points to tease out exactly where I should be looking for the point data. Unit tests are completely useless for this process.
Come to think of it, even once I've got it figured out, it's very hard to see how to usefully use unit tests here...
http://paulgraham.com/arc0.html
I don't use Arc (or any kind of Lisp) but I understand what Paul Graham was trying to achieve.
If you know ahead of time exactly what the inputs and outputs should be then sure, writing unit tests first is a good practice.