For my money, the most realistic chat bots will be the customer service ones: you have a reason to "talk" to them, they have a fixed identity, the context of your conversation is clear, and you share a common goal.
Of course I don't think the current generation of AI is anywhere good enough to be more than a novelty for a few minutes of fun. Maybe in the next few years. NLP is improving quite a bit every year, as are many other areas of AI. Google's chatbot trained just on movie scripts was pretty impressive. Though I don't think it would scale cheaply to millions of users. IBM's Watson has the same issue.
As for customer service bots, I'd love to see them become more general. Rather than just trained on one company's customer service data. Like a bot trained on all of stackoverflow that could specialize in discussions and questions about programming.
I made an irc bot that searched reddit for answers to questions. It worked really well, surprisingly. Especially on certain classes of questions that were likely to have been discussed on reddit before.
As for the reddit irc bot, that's a genuinely cool and useful project, but that's querying a curated dataset rather than having a coherent conversation or identity, so I think the intentions are fairly different. On the other hand, if you had a bot that only queried the comments made by a longtime user, you could get a pretty life-like effect!