Right, bigger blocks take longer to propagate so you get more forking with a block that's too big compared to the time interval.
This might change if Gavin's IBLT proposal [0] gets implemented. Basic idea is that the miners have most of the transactions already (otherwise they couldn't mine), so you don't have to transmit them again in the block itself. You just need a small fixed-size data structure that lets miners reconstruct the full block from the transactions they have already.
Another idea is GHOST [1], which shortens the block time by letting all the forks take part in the consensus process, and get rewarded even if they don't end up in the final linear chain. It was proposed as a Bitcoin modification but just got its first production launch with Ethereum.
Because lowering the interblock time has a more detrimental effects, and raising it has more beneficial effects. To achieve maximum scalability you actually want the longest possible interblock time you can stand. 10 minutes is actually alright in that regard. I would have gone with 15 minutes myself.
