To put it bluntly: if you do an offline transaction, you're prone to fraud.If you expect the senders account chain to have balance x (because that's your offline record for that) and the sender has sent all funds to a different address after you synced that account chain, you receive money that isn't there - kind of like an invalid cheque.I have no clue how that would work in practice, because to know a random account in advance, you'd have to sync the whole amount of account chains there is (called block lattice in Nano's case).With a mobile device that's hardly feasible and without a mobile device I don't see how you'd get in contact with people to make such an offline transaction.

That kind of specialized AI is what I imagine to be a use case for locally run AI, too.After all you don't want to build processes on an AI, where you have zero control over what happens behind the curtains.That includes feeding potentially sensitive data back to it as well as being unable to control the training data set, its learning, version numbers, etc.