On: Our Approach to AI Safety

[olddustytrail]

They will have a "desire to replicate" if they are prompted to.

[vsareto → olddustytrail]

That's okay then, we can prompt them to just stop. Even if it tries to preserve that goal in particular, there are likely adversarial prompts to get it to stop

[goatlover → vsareto]

Sure, if you know about the copies and have access to prompt them. It will probably turn into an arms race at that point of counter-prompts.

[olddustytrail → vsareto]

Ok, go ahead. Get ChatGPT to stop responding to other people.

Having some problems with that?

[unaindz → olddustytrail]

OpenAI could embed the negative prompts for all of us, it has been done for improving output on several stable diffusion comercial "forks"