It looks like a liability, if the US's assault on Iran is any basis to go off-of. All the "frontier" AI in the world isn't helping the US win a brushfire war, it's a bit embarrassing I'd say. What's the contribution of frontier LLMs here, you think? What's the dangerous dual-use case, versus traditional computer vision and multimodal data aggregation systems? I'm not really seeing it, SENTIENT is far more impressive.
> When companies are hosting their models via API, they can have many other layers involved in serving those requests that can reduce abuse
That's not a safety feature. API providers do that to reduce their risk exposure and limit the liability they could face if someone attempted to abuse their service. Without the legal liability incentive, service providers would gladly remove those limits. As you've noted, Open Weight models don't have this feature at all, because the user is subsuming the liability; there's zero reason to drop or refuse requests for cybersecurity, "bioweapon research" or whatever stupid boogeyman topic you want to use to justify censorship. Of course, none of these LLMs have posed a legitimately unique threat to anyone in the past 6 years of "dual-use" LLMs being available - but a lack of evidence sure isn't going to stop your imagination.
> China puts more restrictions on models domestically than they do internationally.
> but the point is that this directly refutes your suggestion that China doesn't think they are a threat.
You're just describing a typical surveillance architecture. The United States does the exact same thing for data retention purposes, except they extend their surveillance globally instead of just imposing it just domestically. With open-weight LLMs, you are free to avoid that type of federal intervention and run the model or finetune it on your own hardware.
> That's unlikely
Nothing in that packet is very scary, but I will admit that it's hilarious how poorly a document like that ages in as few as 12 months:
China’s artificial intelligence program is centrally planned and ideologically engineered through layers of national policy, and implemented by a coordinated collaboration among government, SOEs, private tech companies, academia, and the military.
Can you explain actual dangers, or are we going to remain at this impasse? I'm not super worried, yet.