Go try now, if you haven't. Google Suggest shows the feasibility of using type ahead with very large collections of terms, like tags in a folksonomy.
Now, one of the drawbacks of using ad hoc tags in is the lack of vocabulary control - people use different tags to mean the same thing. This is fine for organizing personal information architectures, but the lack of consistency, while reducing the cognitive cost of classification, actually increases effort in finding information.
To deal with the issue, there needs to be a feedback loop. has the most popular tags float to the top, and others use type size to show more popular tags. There's an argument for that kind of subtle feedback. However, to really bridge between levels of classification, to move from a distributed folksonomy to a controlled vocabulary and then to a formal thesaurus, we need more than implicit incentive in using a particular tag. Using type ahead to show other tags is one way of doing that, as James Spahr . But I've always wondered about how scalable this approach would be with a massive tagset. With Google Suggest, instead of wondering how type ahead would scale, I'm wondering how we can implement a similar scale system for tags...