a collection of thousands of prompts attempting to control the idiot genie of ChatGPT (github.com)

submitted 1 year ago by [email protected] to c/[email protected]

7 comments fedilink hide all child comments

all 9 comments

sorted by: hot top new old

[-] [email protected] 21 points 1 year ago* (last edited 1 year ago)

it's like little toy tugboats trying to steer the titanic around icebergs

[-] [email protected] 6 points 1 year ago

So, you know LLM has a context window right? That means context overflow is a valid LLM attack strategy if you are somehow able to send more stuff than what is allowed.

[-] [email protected] 7 points 1 year ago

and your point is?

[-] [email protected] 6 points 1 year ago* (last edited 1 year ago)

Nothing? I'm pointing out another way to "attack" some LLM?

[-] [email protected] 9 points 1 year ago

Probably would have been easier when the context window wasn't 128k.

Though what the point would be should someone actually achieve that eludes me a bit.

[-] [email protected] 10 points 1 year ago

I don't know man. Maybe a chat bot for help center that forgot they shouldn't give refund or non existent promotion?

[-] [email protected] 1 points 1 year ago

I have LLMs translate novels, I can tell when they run over the window because they forget characters' gender half way through, things like that. Lots of legal documents are even longer, maybe.

this post was submitted on 05 Jul 2024

47 points (100.0% liked)

TechTakes

2057 readers

426 users here now

Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.

This is not debate club. Unless it’s amusing debate.

For actually-good tech, you want our NotAwfulTech community

founded 2 years ago

MODERATORS

[email protected]