Language learning models are all about identifying patterns in how humans use words and copying them. Thing is that’s also how people tend to do things a lot of the time. If you give the LLM enough tertiary data it may be capable of ‘accidentally’ (read: randomly) outputting things you don’t want people to see.
I can’t wait until people find out that you don’t even need to train it on secrets, for it to “leak” secrets.
How so?
Language learning models are all about identifying patterns in how humans use words and copying them. Thing is that’s also how people tend to do things a lot of the time. If you give the LLM enough tertiary data it may be capable of ‘accidentally’ (read: randomly) outputting things you don’t want people to see.
But how would you know when you have this data?
It may prompt people to recognizing things they had glossed over before.