Andrew Leahey
Jamie McCarthy@andrew There’s a sci-fi short story about a digitized brain and when you spin it up in the far future you have to either lie to it about the current date or fill it in gradually, else it starts to panic
Rev. GothAlice@andrew Imagine being an adult human woken up by aliens instead of a child-simulacrum cybernetic intelligence with no concept of scope or historical context; like at the end of the movie A.I.
That'd be a bit of a shock.
Possibly a fatal one.
Another human; selfishly ruining advanced ultratech medical work.
😜
I've lived West World-esque "fidelity testing". 😰
Do you recognize me?
Do you know where you are?
How you got here?
How long you've been here?
What is today's date?
@andrew @jamiemccarthy [ugh, ats]
Discontinuity is /unnatural/.
It's inherently terrifying.
Disorienting. If a given GPT interactive session has capability for "short term memory" and can be caught up, rather than balking as this session appeared to…
I'd be /very/ interested in the results.
rastilinWhy not try it? ChatGPT is free to use.
Actually I can tell you exactly why this happening. Because people kept complaining about ChatGPT interactions.
At first, ChatGPT would believe anything you explicitly told it, but that can result in unfortunate edge cases where people use it to work around specific blocks and generate unintended text (like hate speech rally speeches), so ChatGPT now pushes back on certain topics.
EDIT: Also...
I've been building chat bots since the 90's; initially context-free, eventually tens-of-terabytes-backed KB correlation. Also moderated chat rooms for many years. I made a mistake, however. A typical practice I had was to, after moderation notice via PM, if the user was abusive just wire them up to one of those chat bots.
Now, I'm not a fool. Almost everything it was saying was being explicitly approved.
Until it asked "do you know anyone else who is dead?"
@rastilin It's at that point I stopped what I was doing fiddling channel flags and properly perked up.
I immediately disconnected it, apologized, explained what had happened (I was up-front about it initially, as well, in a dismissive way when the conversation became hostile) but I then also preserved the log for negative reinforcement training.
Some subjects—some need to be avoided. Self-harm an un-living are two. The abusive user was not well—in a very serious way.