I think it's probably enough of an obstacle that it's more likely an AGI will be developed first. In that sense I do agree with Bostrom. However, I wouldn't say it's completely infeasible, rather that it will require considerable advances in pattern recognition technology, our understanding of the brain, and our technological ability to interface with the brain first. The idiosyncratic morphology and distributed/non-localized information storage make for a very difficult engineering problem, but I'm optimistic that it can be overcome in some way or another.

We've already had some (granted, very limited) success with decoding imagery from the visual cortex through "dumb" (non-AGI) machine learning algorithms, which makes deeper interaction seem at least possible. If we can make advances in the above-mentioned fields, I would guess the biggest limitation will be that we'll never have a standardized "plug'n'play" protocol for brains--interfaces will require specialized tuning for each individual and a learning period during which the algorithms can "figure out" how your brain is wired up.