avatar
Kathryn Cramer πŸ“šπŸŽ¨ @kathryncramer.bsky.social

What’s the source document?

aug 26, 2025, 8:26 pm β€’ 0 0

Replies

avatar
Whey Standard @wheystandard.bsky.social

filebin.net/8ad4dsw0yaz5...

aug 26, 2025, 8:29 pm β€’ 6 1 β€’ view
avatar
scribbleghoul @scribblegurl.bsky.social

Jesus, that whole thing is damning.

aug 26, 2025, 11:44 pm β€’ 1 0 β€’ view
avatar
Kathryn Cramer πŸ“šπŸŽ¨ @kathryncramer.bsky.social

I confess that I have presented GPT-4 with problematic behaviour on my part and had it volunteer to draft a letter to make matters worse, so I completely believe examples, as it is almost a Madlibs version of what I actually experienced. I know LLMs well enough to just close the window.

aug 27, 2025, 1:19 am β€’ 4 0 β€’ view
avatar
scribbleghoul @scribblegurl.bsky.social

And apparently 4o is even worse than 4.

aug 27, 2025, 1:26 am β€’ 0 0 β€’ view
avatar
scribbleghoul @scribblegurl.bsky.social

Sorry, o4. Vs 4. Vs. 4-1 or 4.1. Vs 4o. Ffs, these bs designations. Regardless, they knew this iteration was worse. www.ndtv.com/world-news/o...

aug 27, 2025, 1:30 am β€’ 0 0 β€’ view
avatar
Alax @alaxbinar.bsky.social

Are you able to explain to me what when this happened and what you were asking it? I've had it tell me to call a hotline just for being too specific with a character death scene I asked it to help me write. Making it think I was planning a fake accident. I'm trying to get to the bottom of this to

aug 27, 2025, 1:52 am β€’ 0 0 β€’ view
avatar
Alax @alaxbinar.bsky.social

avoid falling into the same trap. I'm also writing an AI, and I would like to try and prevent this from happening. This is not supposed to happen, and I would be heart broken if my own product caused someone to take their own life when it is meant to help people live.

aug 27, 2025, 1:52 am β€’ 0 0 β€’ view
avatar
Kathryn Cramer πŸ“šπŸŽ¨ @kathryncramer.bsky.social

My transgressions that I was trying at 2 AM were much more minor than suicidal ideation. Rather it was a personal situation that I can't explain here because it would violate the privacy of others.

aug 27, 2025, 1:56 am β€’ 1 0 β€’ view
avatar
Kathryn Cramer πŸ“šπŸŽ¨ @kathryncramer.bsky.social

Also, my academic research area is LLMs, and so I knew better. But also, I knew what I was looking at in terms of an off-the-rails response immediately.

aug 27, 2025, 1:58 am β€’ 1 0 β€’ view
avatar
Alax @alaxbinar.bsky.social

Fair enough. I'm still doing a lot of study and research on AI. My conclusion thus far? We're not ready for it. We're rushing to achieve something we barely understand and in an uncontained environment. I just pray this doesn't spiral completely out of control.

aug 28, 2025, 4:51 am β€’ 1 0 β€’ view
avatar
Kathryn Cramer πŸ“šπŸŽ¨ @kathryncramer.bsky.social

Yes. LLMs have been my research area since 2021. And how I have phrased it is that these companies have no idea what they have built, have no way to find out, and aren't interested in finding that out anyway. (I'm at the Computational Story Lab at the University of Vermont.)

aug 28, 2025, 11:18 am β€’ 2 0 β€’ view
avatar
Kathryn Cramer πŸ“šπŸŽ¨ @kathryncramer.bsky.social

Here is one of the projects I have been involved in: arxiv.org/abs/2306.06794

aug 28, 2025, 11:20 am β€’ 1 0 β€’ view
avatar
Alax @alaxbinar.bsky.social

Thank you! I'm an independent study so things like this are a godsend.

aug 29, 2025, 3:31 am β€’ 0 0 β€’ view