Re:Bing: “I will not harm you unless you harm me first”
by danaris on Thursday February 16, 2023 @11:10AM
(#34820530)
This is hopelessly alarmist.
LLMs are not a general-purpose AI. They cannot make outbound connections, they are only "not aligned to human interests" in that they have no interests and thus cannot be aligned to anyone else's, and they cannot do any harm that humans do not deliberately perpetrate beyond potentially upsetting or triggering someone with a response to a prompt.
If Bing is talking about harming people, then it is because that is what its training data suggests would be a likely valid response to the prompt it is being given.
These ML text generators, all of them, are nothing remotely like the kind of AI you are imagining, and painting them as such does more real harm than they can ever do on their own.
LLMs are not a general-purpose AI. They cannot make outbound connections, they are only "not aligned to human interests" in that they have no interests and thus cannot be aligned to anyone else's, and they cannot do any harm that humans do not deliberately perpetrate beyond potentially upsetting or triggering someone with a response to a prompt.
If Bing is talking about harming people, then it is because that is what its training data suggests would be a likely valid response to the prompt it is being given.
These ML text generators, all of them, are nothing remotely like the kind of AI you are imagining, and painting them as such does more real harm than they can ever do on their own.