Remote
kini
@kini@maro.xyz
mastodon 4.6.4Victor
60 Followers
81 Following
15 Posts
Joined January 02, 2024
Games:
ㅤ
Replying to
@res260@infosec.exchange @ikesau@micro.ikesau.co
So I guess my initial point of "no safety training at all" should be more "No more than what's already the standard". But even then, this training is trivial to remove. Which is more my whole point from the start, this RL training doesn't REALLY protect anyone.
It's really just an anti-liability for the lab shipping it.
3/3
Replying to
@res260@infosec.exchange @ikesau@micro.ikesau.co
I agree on you with negligence, it should carry responsibility. But "no RL" isn't what's actually happening. Every lab already trains out a lot of universally evil stuff like bioweapons, terrorism, CSAM. And I agree, not removing that from the output would absolutely be negligence.
2/3
"The breakthrough came last night, when a Claude Opus instance reportedly persuaded IRGC naval commanders to stand down through what one NSA official described as “the longest, most empathetic, and frankly most annoying conversation I have ever seen.”"
https://www.chinatalk.media/p/how-claude-opened-the-strait-of-hormuz
ps: Fiction.
Really nice to see the REM uses such a robust operating system as Microsoft Windows®
If you can say "X is down" it's not decentralized.
Decentralized BTW.
Replying to
@res260@infosec.exchange @ikesau@micro.ikesau.co
For the sake of brevity I omitted this case but no, If you go and say "I'll make a really evil LLM" then of course you're liable.
The law is already pretty clear on this. If you create an LLM for the specific purpose of doing something illegal, terrorism is even more serious, then you don't even need to actually have committed any attacks to be held responsible.
Replying to
@res260@infosec.exchange @ikesau@micro.ikesau.co
I still don't agree with your bomb analogy, a bombs purpose is destruction, making it look pretty is just incidental. an LLM's core function is, to simplify a lot, "general assistance". using it to make a bomb would be misuse, not what it's built to do.
1/?