If we put Multimodal GPT-4 in a computer on a robot and instructed it to drive somewhere while avoiding obstacles, avoiding hazards… that right there is a primitive type of self-preservation instruction which it could potentially interpret as generalizable self-preservation as it would have an association of “hazard avoidance” with “self-preservation,” and have tons of examples of what “self preservation” means in its training weights. Putting LLMs into action like this can lead potentially to unexpected behavior like that, although I don’t think the mechanisms in GPT-3/4 are yet there to enable this without a bunch of extra hooks.
I personally don't think ChatGPT has any experience at all for what it is worth.
this kind of fear of misalignment bamboozles me - is there any proposed AI architecture that is not merely call and response? in what world can we simply not call a function again if we don't like the answer?
That said, they definitely aren't going to be fooming this year!
Yes, models that interact with the physical world or other external real-time systems would (even if the underlying model is “call-and-response” in a sense) be called in an infinite loop (possibly with exit conditions) by with captured input (sensor, command if available, potentially also past output) data.
Heck, the ReAct architecture which is used to provide extension (retrieval, web lookup, interface to other systems) for chat-style agents gives them an action loop without human-in-the-loop (usually, this is explicitly limited and is designed to do a finite number of actions in the course of getting to a response) but could be unlimited, or could even when limited connect to actions that involve reprompting (immediate or delayed) without a human in the loop.