this post was submitted on 19 Nov 2024
82 points (100.0% liked)
Technology
37747 readers
212 users here now
A nice place to discuss rumors, happenings, innovations, and challenges in the technology sphere. We also welcome discussions on the intersections of technology and society. If it’s technological news or discussion of technology, it probably belongs here.
Remember the overriding ethos on Beehaw: Be(e) Nice. Each user you encounter here is a person, and should be treated with kindness (even if they’re wrong, or use a Linux distro you don’t like). Personal attacks will not be tolerated.
Subcommunities on Beehaw:
This community's icon was made by Aaron Schneider, under the CC-BY-NC-SA 4.0 license.
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Would be really interesting to know what kind of conversation preceded that line. What does it take to push an LLM off the edge like that. Did the student pull a DAN or something?
https://gemini.google.com/share/6d141b742a13
Thanks. Seems like a really freaky situation. Must be something with the training data. My guess is, this LLM was trained with all the creepy hostility found on Twitter.
I chalk it up to either a working clock being weird every now and then or prompt engineers trolling.
None that I can see, it looks like they were pasting in questions from their school assignments. There is a link to the chat, and I included some more thoughts in my other comment
Oh, there it is. I just clicked the first link, they didn’t like my privacy settings, so I just said nope and turned around. Didn’t even notice the link to the actual chat.
Anyway, that creepy response really came out of nowhere. Or did it?
What if the training data really does contain hostile and messed up stuff like this? Probably does, because these LLMs have eaten everything the internet has to offer, which isn’t exactly a healthy diet for a developing neural network.
Usually LLMs for the public are sanitized and censored, to prevent lot of creepy stuff. But no system is perfect. Some random state can cause random answers that makes no sense, if triggered. Microsofts Ai attempts, Google's previous Ai's, ChatGPT and other LLMs all had their fair share of problems. They will probably add some more guard rails after this public disaster; until next problem happens. There are dedicated users who try to force this kind of stuff, just like hacker trying to hack websites (as an analogy).
With the sheer volume of training data required, I have a hard time believing that the data sanitation is high quality.
If I had to guess, it's largely filtered through scripts, and not thoroughly vetted by humans. So data sanitation might look for the removal of slurs and profanity, but wouldn't have a way to find misinformation or a request that the reader stops existing.
anything containing "die" ought to warrant a human skimming it over at least
I don't disagree, but it is a challenging problem. If you're filtering for "die" then you're going to find diet, indie, diesel, remedied, and just a whole mess of other words.
I'm in the camp where I believe they really should be reading all their inputs. You'll never know what you're feeding the machine otherwise.
However I have no illusions that they're not cutting corners to save money
huh? finding only the literal word "die" is a trivial regex, it's something vim users do all the time when editing text files lol
Sure, but underestimating the scope is how you wind up with a Scunthorpe problem
i feel like that's being forced in here, i'm literally just saying that they should scan through any text with the literal word "die" to make sure it's not obviously calling for murder. it's not a complex idea
They could just run the whole dataset through sentiment analysis and delete the parts that get categorized as negative, hostile or messed up.
Stuff like this should help with that. If the AI can evaluate the response before spitting it out, that could improve the quality a lot.