There is no such thing as an autonomous LLM. Humans are responsible for their software systems.
LLMs Are Not a New Construct
It’s a computer program. A company is behind it. Humans are in the company. If it’s running, it’s because a human prompted it, or ran a script that launched a bunch of them, or ran a script that launched an agent that then launched a bunch of other agents. Doesn’t matter how far you recurse.
Even if some agent started copying its weights (or any weights of any other model, doesn’t have to be its own) and instructing its progeny to do more copying that’s … just a worm. Who launched the initial agent? Someone did. There’s a human at the end of the chain. And there’s a company that that human works for. It’s already illegal to hack other people’s computers. So we don’t need any new laws. We have plenty of computer crimes laws. Enforce them if you’re so worried about it. Make OpenAI pay a fine (a large one) for the HuggingFace attack, for example.
If you’re worried about China or some cybercriminals in a far off land using LLMs to attack us, then how would a law stop them? Any laws passed in the United States would only affect the entities in the United States, or those it can influence. This is why I find the recent talk of “slowing down” silly. There’s no slowing it down. Be responsible, absolutely, but don’t think that announcing a slowdown is going to do anything.
Same with bioweapons (which I think is a made up scare-tactic), which are already illegal to make (or even conspire to make). So if someone in the US is doing that … then arrest them for conspiracy to make bioweapons? And if that person is a terrorist in some far off land then use our massive surveillance apparatus and military to handle them.
But don’t pass some laws at home that give me access to only two model providers for “safety” while the rest of the world continues to use whatever LLMs they want. Remember, you can’t enforce “you can’t use open weight models” worldwide, and you can’t make China stop developing software. Pausing only hurts those who follow the law, and nothing for those who don’t (the exact group you’re trying to control via pausing). And if you’re worried about some rogue AI agent doing something accidentally, then start enforcing the laws against the companies and humans that launched them. There is no such thing as an autonomous LLM.