So, I've been playing with the OLPC a bit (I got a chance to play with a prototype in person at a party a while back) and I wonder: where's the science/mathematics? Authoring languages and EToys, which seem to be the focus presently (well, that and Tetris of course) are all well and good, but it would be really nice to see some more focus on the physical sciences, which leads naturally into topics like data analysis and statistics (collecting data is just the first step after all).
The minimal implementation would seem to be to start with something like the TILE/StatDocs projects that Duncan Temple Lang (Duncan, I hearby dub you DTL for the remainder of my posts. Sort of like DHH, the Ruby on Rails guy) and Deb Nolan have been doing for a while now (there used to be a StatDocs.org website, but it seems to have disappeared). They give you an organizational unit of a "Lab" or an "Experiment," with some data and ways of interacting with the data (plotting and so on). These'd probably be hosted on the school's server and brought across the network dynamically. Another interesting tack would be to let the server also COLLECT data from the kids' experiments to be combined together into larger datasets via surveys or actual data collection. An ad hoc version of this is pretty common in introductory statistics courses for example.
This leads to the question of what to use as the analytical environment. Being biased, I would tend towards using R + GTK + Gecko, with a UI designed for the more modal experience of the OLPC. Alternatively, Python could be used for the implementation, though I'm not convinced that Python would be a particularly good interactive data language (things like R/S's formula DSL are really quite powerful for example). I tend to dislike point-n-click on principal, especially in an environment that stressed authoring languages.
Beyond that, you want kids to be able to collect their own data. I remember having this thing for my Atari 800 when I was a kid that had a number of different probe options for things like temperature and pH and rainfall and such. The software it had was pretty primitive, but it was lots of fun (for me anyway) playing with the temperature probe and such. It seems like the USB ports could be used to build data acquisition systems pretty cheaply these days---you wouldn't need really high performance for the most part and that lets kids collect their own data, which will always be more compelling than some prepackaged dataset since they "own" the data. There's at least one, possibly more, Open Source hardware project for building lots of these sorts of sensors for DIY weather monitoring that can probably be leveraged for that sort of thing.