CRDT's trick is metadata. Good old diff guesses the changes by solving the longest-common-subsequence problem. There is always some degree of confusion as changes accumulate. CRDTs can know the exact changes, or at least guess less.