I learned some things too. However, even if you have a tiny piece of software, like a game that you can reliably complete manually in 20 minutes, completely playing it through after every change is way too slow for "live coding". You should see your results instantly. Just like your test suite should run in seconds. If you have to wait minutes to compile, deploy and test your productivity will fall. A full integration/acceptance test suite can take longer, but it should be as automatic as possible and you can't use it for rapid feedback like this, but more for regression/acceptance testing.

Notch was able to take advantage of immediate feedback for the most part of his coding. The first hours spent on the rendering was tested by watching the world being rendered live. I've been doing the same developing a game in Clojure. When he got as far as the gameplay, he slowed down considerably, since he had to wander around in the game world to test each new feature. He for example made temporary shortcut passages, so he was able to test the boss monsters, without actually having to pass through the levels.

He is a person who can concentrate on delivering results and hack away at code for long stretches. This is the kind of code he presumably has done again and again for years. Who professional web developer can't hack together a small site as quickly? Where most people fail is attention span and drive! I know my unfinished projects speak of that :)

Most interesting part to me, of watching Notch code, was to see how he used his tools. And it was inspiring and motivating to see the progress. The actual code was very hackish but quality code is not important in a throw-away project anyway :)