For full details on the changes, see (as usual) the change history - that's available on tads.org on the Manuals page, along with the current documentation.
But there are also some serious pluses. The big one is that TADS now has a true network infrastructure. The obvious possibility this opens up is multi-user play, for building MUDs or multi-PC games. Now, this is just a possibility at this point, since there's no library support for any of that, but the network plumbing is in place to support it. I see a few other possibilities that are also pretty interesting. One is collaborative play - which is more than a possibility, actually, in that the updated Adv3 supports it out of the box. Collaborative play means that multiple people can simultaneously connect to a session - still playing a single-player game, but with each player able to see what the other players see and able to enter commands. The same idea as a couple people huddling around a computer playing together, but remotely via the Internet. Another idea I'd like to explore at some point is using the networking to let the author be involved while people are playing - either directly involved in the sense of actively manipulating the game in response to what the player is trying, or more simply, just gathering ongoing information on where players got stuck, verbs that didn't work, etc., that could go into improving the game in a rapid update cycle. Same idea as beta testing, but continuously after release, and without people having to manually gather logs and send them to the author. I also have this back-of-the-mind idea of using the network as a real-time human supplement to the parser, where commands that aren't recognized by the mechanical parser could be instantly flashed to a network of people who could step in and interpret the commands into something recognizable to the parser, and send it back, so that from the player's perspective the parser seems to understand whatever a human could.
The second big new feature area is dynamic compilation, and a bunch of related improvements in the reflection system. This pushes TADS pretty fully into dynamic language territory. So far it's not something that Adv3 uses at all, but I added it because I have a bunch of library facilities in mind that could take advantage of it. Adv3 or, more likely, the "Mercury" library I've mentioned will probably start taking advantage of it in the next round, and in the meantime, I think it opens up a lot of new possibilities for add-on libraries.