this post was submitted on 19 Feb 2024
517 points (98.9% liked)
Technology
60123 readers
2774 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
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
Generally, what's the best/most efficient way to make LLMs go off the rail? I mean without just typing lots of gibberish and making it too obvious. As an example: I've seen people formatting their prompts with java code for like 2 lines and replies instantly went nuts.
I use a few dozen novels in a single text file and randomize which lines the script pulls. It then replaces the text three times with a random pull. What you end up with are four responses in plain English. Which is the real one? You could filter out responses edited after "the great exodus", but I have been doing this to my comments a few times per year during my twelve years on reddit.
The truth is that even if I don't get them all, I get enough that it makes it far easier for the group that bought the data to just filter my username out rather than figure out what's junk and what isn't.