As I've mentioned a couple times before, I participated in the 2007 Pittsburgh Game Jam. The games themselves are posted here.
In spite of the fact that the only constraints were technological (design for the XO platform, using Python and Pygame), and there was not a single mention of experimental gameplay or innovation in game design, a surprising number of games were highly experimental in nature (though sad to say, not mine). What follows is a rundown of, in my opinion, the most important games of the Jam and why students, teachers and professional developers alike should pay attention.
XO Maze (by Team Thailand), the winner of the jam, was a four-player cooperative maze game (if you've never heard of the "maze game" genre, think Pac-Man) with fog of war. Design lessons: Whenever a new innovation in gameplay is added to a genre (like fog of war in RTS games), someone should revisit all older games of all genres and see if the new mechanic will work with old games. In this case, it works beautifully. Another interesting thing: the developers originally wanted network play but it was broken, so they "settled" for having four-player on a single computer. The result was actually better than if they had network play -- it turns out, four players crowded around a single keyboard gives a uniquely satisfying experience when playing co-op. With all of the rush for modern consoles to have games that are all network-enabled, we must not forget the joy of sitting next to your friend and playing games on the same couch.
Caketown (by Team Argentina) is a puzzle game where you click on stuff to make other stuff happen; the object is to get the strangely cute zombie-animals to find each other, which lets them eat cake (don't ask, just see for yourself). Only two levels created during the Jam, but after playing them you can see the potential. Design lessons: This is a great experiment in UI. There aren't really any printed instructions or text, but the visual clues are still there to tell you what to do, and if you can't figure out the iconic signs you can figure things out experimentally by clicking on the stuff that obviously looks like it can be clicked on. If you're working on a game that requires the player to go through an hour-long tutorial just to figure out the basic rules of the game, try taking a look at this and seeing how the graphics and art (even environmental art) can do wonders to teach the gameplay.
Head Cat (by Team Brazil) is my personal favorite game of the Jam; the only reason it didn't win is that the judges were mostly 8 to 12 year olds, and the creators of this game admitted that they ignored the judging and just made the game that they wanted to make (it's clearly targeted at a slightly older audience). The gameplay is something between Lemmings and The Incredible Machine. The goal is to get your programmable robot to the cat; the cat then happily goes to sleep on the robot's head (anyone who has ever owned a cat will instantly understand this). Design lessons: This is a great example of what I unsuccessfully tried to teach my Capstone class last year -- namely, to attempt depth of mechanics instead of breadth (especially when working on a time-limited project, such as a Game Jam game or a semester-long class project). There are only a few player verbs in Head Cat: the drill (breaks blocks beneath you), the balloon (slows your rate of falling), the magnet (pulls you up if there's another magnet overhead), and the ability to detect running into walls and floors and reverse direction. Each of these was clearly easy to implement, and yet look at the variety of levels and challenges! The developers even had time to include a full level editor, and I imagine you could make 50 or 100 levels with just the mechanics that exist in the game, alone. Incidentally, it's also an example of how a cute theme can take a decent game and make it instantly memorable.
Honestly, all of the games were impressive for some reason or another. Fruitix (Team Ethiopia) is something like a retro-arcade version of Harvest Moon. Red Bird (Team Tibet) is a timing-based puzzle game, something like a Match-3 game where all panels are the same color and their position matters rather than their color. Star Catcher (Team Peru) had some very satisfying 2D physics and a full level editor (remember that these games were made in about 36 hours). Tumble Boy (Team Libya) did what I tried and failed to do: get scrolling to work on the XO at a reasonable framerate -- and it not only included a level editor, but made it accessible as a plaintext file that you could edit in notepad. Light Ball (Team Nigeria) was another impressive physics game, with the interesting mechanic that you're trying to prevent the sun from going out -- so you have to balance your limited ammunition with the limited time that the sun has left, and in the mean time the screen starts going dark (making gameplay more difficult). I guess what I'm saying is, download and play all the games; you'll be happy you did.
And play with the sound on. All of the games have sound; sometimes it's critical to gameplay, other times it's just amusing, but I can't think of any game that had actively horrible sound. Except perhaps my own, as I explained.