The comparison to cryptography export restrictions is absolutely valid, but is that really favorable?

I'd argue that 70's cryptography export bans in hindsight look completely misguided, futile, burdensome and pointless in the end (which is why most of it was lifted/reverted over the last decades).

I don't see how AI-models are much different; it's certainly a better comparison than the fuzes, because we're both not at war right now and the underlying principle is already out of the bag.

Cryptography export restrictions absolutely worked until the internet became commonplace, and then they became futile and were removed.

Just like every other export restriction on technology: once the actual cat is out of the actual bag, they are often relaxed.

The "underlying principles" here are hundreds of billions of dollars in R&D - which is what is required to compete with the frontier models.

> not at war right now

We weren't technically at war during the Cold War, either.