Another alternative to avoid guessing is to use randomized 64-bit integer keys. You still risk collisions over sharding/replication, but only if you truly have a lot of data. You potentially lose some index performance but it shouldn't be any worse than with guids. If you really need the full size of a guid, just use them for the key. I don't get the rest of his argument for hiding internal surrogate keys.
A GUID has much superior collision resistance than a bare random integer.
The GUID includes a part that is a random (like the int), a time component (prevents collision for hashes generate at different microsecond) and network information (prevent collision across host names).
