[HarHarVeryFunny]

I suspect this new Anthropic requirement is coming from their ongoing negotiations with the government to re-enable Fable. Whatever "safety/security" measures the government requires of Anthropic will no doubt also be applied to other US "AI providers", perhaps based on assessed model capability. Apparently OpenAI already had an ID check in place before this.

What's going to be interesting is how the US government regulates US-based access to Chinese AI models, whether served domestically (e.g. GLM-5.2 from Amazon Bedrock, DeepInfra, etc), or from overseas.

I suppose if one fear is US domestic terrorists/hackers using AI, then restricting access to all models, regardless of country of origin, might make sense, but as far as restricting technology exports over national security concerns it wouldn't make much sense to restrict access to foreign models!

Crudely applied restrictions are likely to get worse before they potentially get better as the hype wears down and AI risk gets better assessed, but government control tends to be a one-way ratcheting up, so who knows. It's not inconceivable that the government may try to restrict use of local models too, which they could do by making it illegal to make open weights (maybe for selected models) available for download.

[anon373839]

> suspect this new Anthropic requirement is coming from their ongoing negotiations with the government

I would have put “negotiations” in scare quotes. Who reported the Fable jailbreak to the government? Anthropic’s biggest shareholder. Who is now sitting at the table designing the “benchmarks” that will dictate what other model builders will be allowed to deploy? Anthropic. Who is also in talks with the government about a bailout? Also Anthropic. I just can’t take any of this at face value because that’s preposterous.

[HarHarVeryFunny]

Agreed, Anthropic have been lobbying in favor of government regulation for a while now, and there is no indication that they are not getting exactly what they wanted. Maybe it helps to push any liability onto the government rather than themselves, and certainly let's them try to hand wash from any bad outcomes.

I recall a Dario interview from a few years ago talking about security in terms of protecting their model weights and as I remember even back then they had hired ex. government security officials ... I would imagine they know exactly what to tell the government to make the "negotiation" go the way they want to. Unrelated to the current conversation, but one detail of that Dario interview I recall was mention that there is an assumption that in any organization over a given size trying to protect tech from foreign governments, there 100% will be spies on your staff, and you need to plan accordingly.

[rescbr]

> making it illegal to make open weights (maybe for selected models) available for download.

So giving another win to China?

[HarHarVeryFunny]

Are you assuming that China is/will allow it's own citizens unfettered access to AI ?!