Performance -- even if the CSS is the same size as the image, a single CSS file can return all the icons with a single HTTP request, vs having to make separate requests for each image.
Even with CSS Sprites, there are two requests - one for the CSS, another for the sprite image file.
Unless you use data: URIs and base64-encode the images, which can result in a smaller CSS file than "pure" CSS while still eliminating extra HTTP requests. That depends on the complexity of the image and of the gradient set (or what have you) that would be needed to replace it. Optimization is always situational.
Oh great, I once felt the need of it, but did not it actually existed. I feel though that pure CSS solution would be better, since modifying the colors etc would be easier with pure CSS.
The main advantage of CSS is that you can change the color and appearance of components without using a graphics package, and they might scale better. But that's about it.
It was a technical exercise rather than a proof of concept. The icons themselves are very simple and limited. I wouldn't recommend using CSS in this way for production work (other than perhaps for very simple arrows).
I'm not sure some simple arrows can be achieved with unicode characters...nor would you want a screenreader to pay attention to what is purely presentational.
To see something like this in action, this HTML Mail.app clone demo (that I'm pretty sure hit the HN front page several months ago) uses data URIs for all its icons:
http://ukijs.org/mail/index.html
In theory, but looking across browsers shows that you'd never be able to use this in production without falling back to graphics, which adds to production, maintenance and degrades any performance benefits for a good portion of users.
It's a nice theoretical challenge, and may provide some benefits if you are building (mobile) device specific systems, but as a broad technique there is little production value at the moment.
Even with CSS Sprites, there are two requests - one for the CSS, another for the sprite image file.