Well, the idea here is that the raster (bmp) conversion is done at the final display resolution where there is a 1-1 relationship between internal pixels and screen pixels.
Otherwise, sub-pixel (as in LCD smoothing) doesn't really make sense, since a (255, 0, 0) is supposed to light up the first third of a pixel, etc. If you were to scale LCD-smoothed text up, that one pixel value would map to multiple pixels and you would get what you see when you zoom-in on pixels in an image editor, where you are left wondering why your text is rainbow colored.
Missing your target resolution with LCD smoothing is bad news bears (, as far as I know. I have never tested the results of bilinear filtering 64 pt down to 10, for example. I am just assuming it's not exactly correct). With normal AA, it's less of an issue because you assume that your target resolution (at your target DPI) is high enough that you don't need sub-pixel detail. In that case, you can render a high-resolution atlas, rely on hardware filtering to render down to the target resolution -- into a multisampled color buffer (2X, 4X, etc.) -- and then do the normal mutisample resolve into the target resolution. Again, this will have the effect of AA down to whole pixel resolution, not sub-pixel, but that's fine if you're rendering on a sufficiently high DPI display. If your display is around the default of 96 dpi, you're going to want sub-pixel if you can (if your dealing with an LCD display).
Also note that this conversation is different from sub-pixel positioning, which is just using floating point pixel coordinates to layout UI elements. At some point, at the end of your pipeline, those units need to get snapped to pixel boundaries to look good like Allen mentions in one of the 4coder jam vods.
EDIT: I just realized that this makes it sound like you need MSAA or something to get AA. You don't. Your rasterizer will hopefully provide anti-aliased output. Whether you re-rasterize on the fly for zooming/resizing or use an atlas is up to you.