Well that seems a bit pedantic, but yes, you're right. The difficulty I was focusing on however is the randomness itself, not how far it has to be extended. Ostensibly once you've incorporated a non-deterministic seed (which is necessarily external) into your process, you can extend it to the pad itself. The pad itself will just be a stream of that data, like a non-deterministic stream cipher. I get the impression you're aware of this already though, so I won't belabor it.

> One time pad systems are used regularly for high-security embassy-to-State Department communications. Since there are also secure couriers between those locations, there's a way to distribute key material. With today's storage densities, carrying key material around is easy.

Fair enough, I'll concede they have practical uses in situations where there are couriers and extremely high security requirements. That's about all they're practical for, though :)

I might be just misunderstanding you though.

So no, you're not misunderstanding - I just communicated poorly!

Is it just the fact that it would take two trips for the courier? Or that someone would need to intercept both communications (pad, ciphertext)?

The internet is an untrusted medium, since packet switching requires packets to travel through routers that can edit them at will. (The fabled man in the middle attack) This is the whole point of the certificate issuer public-key infrastructure, where issuer public keys are included in your OS's installation files, a secure communication channel. (If you can't trust your OS, you of course can't trust any communication made with it to be secure anyway)