The basic idea is to diff and merge ASTs rather than long strings. We have some more specific ideas about specific features and advantages, but since we have almost no code yet I won't get into them. We should (since this is actually for school, we have to) have a working prototype several weeks from now, so I'll talk more about it then.
We're not really fixing "version control" per se; rather, we're building a tool that you could use in tandem with existing systems like Git. So we'll help with diffs, interactive merges (maybe even automatic conflict resolutions) and some other stuff but let the actual storage and versioning stay with Git (or any other VCS).
I can't make any promises, but if we get something interesting working I'll be sure to post it here. We actually tried to do something like this for a hackathon once, and it sort of worked, but was really a big hack throughout. Hopefully we'll get it right this time.
Actually, ideally you would have content-aware DIFFs. So source code has one visualiser, images another, each binary format another.
That would make the experience much better.
Once you have that it should be much easier to check in more granularity, produce a more accurate map of dependencies and roll out refactorings that didn't work out.
Focus on expanding the concept, listen to a small group of critical people you can trust for feedback and you'll build the tool we're all supporting you to build :)
As an aside: to truly 'fix' files, you need to 'fix' version control. That's out of scope for Light Table, but it's something that could be addressed by a parallel project (if someone else fancies looking at it).
* i.e. complicating things so you can support Java/C# - few of those devs will ever use Light Table, and if the concept is a hit then you'll have Microsoft and Oracle waving big wads of cash at you to support their platforms.