An interesting question is whether AI can be manipulated effectively. In other words, can I teach a state of the art useful model that believes 2+2=5 Sure you

An interesting question is whether AI can be manipulated effectively. In other words, can I teach a state of the art useful model that believes 2+2=5 Sure you can get it to state 2+2=5 if asked, but there are a lot of ways to ask about 2+2 and various implications of a system of knowledge that believes 2+2=5. If having the LLM believe untruths borks the models, then that's a good thing. I think a similar question is about encryption. Is the nature of encryption defensive or offensive. I don't know enough to say its directly related to p=np but I imagine the question is if there is a way to encrypt something such that the physical laws of nature make cracking the encryption impractical.