It probably doesn't matter very much.
The computer has an advantage because it can bring more processor power to bear. You can use more sophisticated algorithms, including tailoring the right one to each image. (And, as Matt Grum points out, you have the larger version available if you change your mind. This is probably the most compelling reason to go this route – it's hard to guess your future needs.)
The in-device conversion may have other advantages, though, depending on how it's implemented. First, it can do its downsizing on the raw sensor data rather than working on an already-converted JPEG. That avoids saving in JPEG more than once (which, given that you're discarding detail, isn't that important in this case), and allows the camera to do the downsizing as part of Bayer de-mozaicing. This might give a minor quality improvement. Second, the sensor may do a hardware-level pixel binning, which decreases read noise when shooting at a lower resolution.
But those advantages are highly implementation-dependent. I think the best thing to do is actually take some images and compare. If you can't tell the difference, go with whatever is easiest. (Or take the advice of storing the larger versions somewhere after all, just in case.)