You're right that JS is pretty ill-suited for implementing production-level Deep Learning architectures that operate on a lot of data, I think it's still well suited for explaining the concepts in a clear way. A lot of the models have intuitive (and visual) interpretations, something that can be nicely visualized with JS.

As for Karpathy's implementation, you're right, he chose JS for that reason. His library is meant as a tool for experimenting and teaching, not to deploy these models in production.

Op: Send me an email if you're interested in this (email in profile). I actually started something similar just recently at http://www.wildml.com/.