Many submissions aren't actually made in the given time period. They usually will bring either a partially to fully complete project (pretty sure it's cheating but yeah).
I don't think this is true, actually. I've been to two major hackathons and had many friends at HackMIT this fall, and this really doesn't occur- in fact, the culture was the exact opposite, people were strongly discouraged from bringing in already existing projects and almost nobody did.
I think what he was saying is people already started researching on what ideas to implement prior to the event. Many people brought in their own Pi and hardware hacks so it isn't like they didn't figure out 1/3 of the hard work. But I wouldn't undermine their ability to piece things together in 18, 20 hours.
There aren't really topics for a lot of the hackathons (including HackMIT). For the "themed" hackathons you'll know the topics at registration (ie. weeks/months beforehand).
Multiple awards seemed to be sponsored or product unique, that is why I wonder on this point. Certain items I would imaging having a hands on beforehand would be quite leg up. Excuse my mixed metaphors ;).
That's really not the case. I'm a regular on the hackathon circuit and it's extremely rare & frowned upon for people to start building before the hackathon. Plus, if you doubt it, peruse their GitHub's and you can see clear building.
It depends on the project, but in my experience many (if not most) people end up making their repos public.
As for finding them, I'll often just look up the hackers on GitHub and check if they have new repos.
Though that's not always the case. I kept the repo for the hack I made this weekend (http://socialsecurity.io — antivirus for Twitter) private because I'm commercializing some of the machine learning.
Most of the teams I talked to, mine included, were using private repos. As much as I trust the community and my peers, it would be interesting to require public git projects during a hackathon. Not only would they be a fantastic learning tool, the potential data visualizations would provide a really cool postmortem.