I imagine a system where you start with a basic model that just reads through the script (the same way a low-end tech support employee would be expected to), but you get forwarded to a better model if that doesn’t work[1].
Or if you jailbreak it into sending you to a human specialist, which would require SOTA knowledge on what works for that and what doesn’t, limiting the number of specialists you’d need and commensurately increasing quality. Sort of like this old XKCD comic, except real.
The problem with backdoors is that all it takes is someone to go blabbing about it and then everyone knows and you’re back to square one. What would be a game changer is if you had, say, a personal AI that could provide attestation for its user’s skill level, but that comes with privacy concerns… Yeah, I’m not convinced my solution is better, actually.
I imagine a system where you start with a basic model that just reads through the script (the same way a low-end tech support employee would be expected to), but you get forwarded to a better model if that doesn’t work[1].
Or if you jailbreak it into sending you to a human specialist, which would require SOTA knowledge on what works for that and what doesn’t, limiting the number of specialists you’d need and commensurately increasing quality. Sort of like this old XKCD comic, except real.
The problem with backdoors is that all it takes is someone to go blabbing about it and then everyone knows and you’re back to square one. What would be a game changer is if you had, say, a personal AI that could provide attestation for its user’s skill level, but that comes with privacy concerns… Yeah, I’m not convinced my solution is better, actually.