It seems to me that explicitly specifying serialVersionUID for new classes is bad. Consider the two cases of not changing it when layout has it should have been changed and changing it when it should not have.
Not changing when it should have been changed occurs almost only when it is explicit. In this case, it results in some very subtle, hard-to-find bugs. Especially during development, when class layout changes often. But if it has not been explicitly specified, it will change and the deserialization will break loudly, mostly likely solved by purging the repository.
Changing it when should not have would occur almost only when it is implicit. This is the rare case where class layout has changed but we still want to deserialize from the old serialized blobs. This will likely be caught during QA (Strange errors after upgrade from 5.2 to 5.2.1, see attached stack trace) and can be trivially fixed by setting a explicit value.
Comments?
Changing when it shouldn't may happen for reasons other than class layout changes - the problem is that it's compiler implementation dependent. If you do debug with Eclipse but do production builds with javac, you may end up with two incompatible sets of data.