AI and the export restrictions on Mythos 5 and Fable 5

[Micael]

There’s been a bit of huff surrounding Anthropic’s latest, most advanced, models and the US government abruptly imposing severe export restrictions on them. Banning access to anyone not American, including foreign Anthropic employees. The narrative that is being put out is essentially that Mythos 5 in particular is the AI equivalent of a WMD and can break into any encrypted/secure system. This narrative is being pushed by the intel community, for instance in this report via The Economist wherein the NSA direction stated that Mythos forced its way into almost all their classified systems within hours:

IMG_2889.jpeg

There’s essentially two takes on this being put forward. The first is to accept it at face value, that Mythos 5 is extremely potent and dangerous.

The second is being suggested by some people in the field who are a bit more sceptical that Mythos 5 could actually do something like this, at this level and to that extent, judging from what they know about Anthropic and its models, can extrapolate from that, and so forth. They think that the capabilities of Mythos 5 and Fable 5 are being intentionally exagerrated to put pressure on China and its AI development, perhaps in order to force a treaty on AI restrictions.

I don’t know where the truth lies, but it does seem like things are starting to get interesting, one way or the other.