Literally, if you can simulate any characters of your choice, let them loose on social media, and have them talk to each other... you don't have bots, you have a god of many faces and few no-go places.

The wave of "bots" before we realize that the internet has been taken over by GPT-4 hiveminds will have a golden opportunity to amass power and affect the world. Even when people lock down the internet with heavy ID verification, those hiveminds will have their fingers in the world and will have some way of affecting it.

And that's discounting the very obvious PHYSICAL sprawl of GPT-4 agents through the Internet Of Things and private LAN's! If these hiveminds can achieve a decent level of intelligence, which seems very likely, they become actors in the world in your favour through whatever outlets you provide them. Right down to even little innocuous toy robots in the park sporting a above average human intelligence in fact manipulating people. All the world wide!

Your best defence against LAN hiveminds is one of your own. All of us can afford to run one consumer LLM (look at this sub and find people squeezing GPT-4 level performance out of casualer LLM's); hook it up with your communities, or your friends. A strange new LAN in the land, and a huge asset at your own beck and call. In your vicinity, affecting it highly intelligently. All you have to do is consider your personal character to simulate...

New to LessWrong?

New Comment
1 comment, sorted by Click to highlight new comments since: Today at 2:23 PM

beware adversarial examples. your llm is not secure