The error rate of 15% (I actually believe it's more like 30% in practice, which is 2x PacBio's) is actually excellent and comparable to PacBio's in that the error is primarily randomly distributed, and not systematic like current Illumina or IonTorrent systems (which have admittedly lower error rates, at 1-3%). The short of it is, absolute error rate is not the whole story.

When looking at what ends up being just a massive string of 4 characters, having 15% error on a particular read doesn't matter too much; you end up getting many reads (fragments of a full sequence) which overlap the same area of the genome. With enough reads (and having many, say 15-100x coverage as it's called, is not uncommon practice in sequencing), this error is obviated by consensus. The long read length is especially useful for this, as the lengths which overlap are much greater than with current NSG reads of 150-200bps.

Additionally, I think the goal of the MinION were more to display a sort of disposable (sub-$500) machine which could be used on-site and without sample-prep. A forensic kit, really (though probably more for creatures such as bacteria and fungus rather than humans—I recall that it would take 2-5 MinIONs to sequence a human genome to even a short depth, but I may be misremembering).

In short, this is a cool proof-of-concept for Oxford Nanopore, as their technology has been talked about for a long time with nothing to show until now. It will be interesting to see where it goes next, particularly if the reads lengths keep climbing and the error profile drops even more.