The Chinese company also released the means to correct this though.

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.

there's a video exactly like this (app changes text mid response) of people asking gemini if google is evil.

There is a difference between not telling someone how to make Sarin/age-gating certain topics and censoring works of fiction or references to politicians and anyone who claims there is not is so full of shit they need to have their stomach pumped.