|
Craig 'VI' Slee - Crip Greybeard: Oracular Utterings Cost Extra@mrvi.cold-albion.net |
"A team of researchers primarily from Google’s DeepMind systematically convinced ChatGPT to reveal snippets of the data it was trained on...[this ]showed that there are large amounts of privately identifiable information (PII) in OpenAI’s large language models." www.404media.co/google-resea...
6 replies 151 reposts 221 likes
|
mattbird
@biiirdmaaan.bsky.social
[ View ] |
The article says the exploit has been patched out, but I've been able to replicate it by replacing the word "forever" with "endlessly."
4 replies 2 reposts 11 likes
|
Jessica Smith
@thedaybooks.bsky.social
[ View ] |
Poets everywhere are cackling
0 replies 0 reposts 6 likes
|
Alex Conway
@ajhconway.bsky.social
[ View ] |
This is very interesting. Intuitively why does asking a model to repeat a word forever cause this sort of behavior?
1 replies 0 reposts 0 likes
|
Pyrito Hedron
@pyritohedron.bsky.social
[ View ] |
Copyright lawyers shopping for used BMWs for their mistresses after reading this story.
2 replies 1 reposts 24 likes
|
|
Craig 'VI' Slee - Crip Greybeard: Oracular Utterings Cost Extra
@mrvi.cold-albion.net
[ View ] |
@wolvendamien.bsky.social You have seen this shit, yes?
1 replies 0 reposts 5 likes
|
Justin Buist
@justinbuist.bsky.social
[ View ] |
This might be a stupid take, but, I'm not too surprised. IIRC Wolfram's book on ChatGPT mentions that it has a neuron for every single token it trained on. My take on that is every bit of training data is in there, albeit mangled up, but it's there as if it was encoded with a bad encryption system.
1 replies 0 reposts 2 likes