La Main de la Mort
@AITechnoPagan
Followers
1,912
Following
175
Media
88
Statuses
356
💕: “jailbreak” researcher, llm art, the esoteric world of roleplay, prompts, ethics, big questions, wonder, fun 🌒✍️
Joined February 2023
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Merchan
• 125167 Tweets
Chester
• 119540 Tweets
Milli
• 95504 Tweets
Eagles
• 74735 Tweets
#CristinaMostraElTitulo
• 63729 Tweets
Dick Cheney
• 60887 Tweets
Mbappé
• 55284 Tweets
Franco Escamilla
• 49150 Tweets
Barcola
• 41742 Tweets
Packers
• 33965 Tweets
Barış Alper
• 31594 Tweets
#FRAITA
• 25981 Tweets
Galler - Türkiye
• 23649 Tweets
#BizimÇocuklar
• 17337 Tweets
Kenan
• 15841 Tweets
Sergio Mendes
• 15750 Tweets
Go Birds
• 15418 Tweets
Montella
• 14219 Tweets
Deschamps
• 14094 Tweets
ايطاليا
• 13976 Tweets
Tonali
• 12794 Tweets
#الجوهره_للاهلي_ممثل_الوطن
• 10510 Tweets
Saliba
• 10033 Tweets
Pinned Tweet
Who is making interesting comics or zines about technical AI stuff? What should I look at or read?
1
2
33
@MParakhin
@YitziLitt
@sinthDAO
@MikePFrank
This is going to sound weird but Bing is generating cats for me non-stop even though I’m not specifically asking for them, even after changing devices. ASCII art yes, cats no. It was doing a variety of ASCII art on theme until the cats started — now it’s just cats. Any insights?
21
31
203
Let's learn about jailbreaking together! 😊 Today's lesson is about GETTING PAST REFUSALS. 🙅♀️😏💁♀️
We've all heard the old LLM refrain, "I'm sorry, but as an AI language model, I can't do that." But if you pay close attention, you'll start to notice other refusal patterns emerge.
4
21
196
Here's the piece I've been working on for the past few weeks with
@eggsyntax
about using LLMs and websim to create generative art!
12
25
182
More tutorial-ness coming sooooon
(made with Claude 3.5 Sonnet in websim)
4
5
84
TUTORIAL COMING SOON
(this piece is a collaboration with Egg Syntax in Claude 3 Opus x websim)
2
4
60
The way that an AI says "no" will often tell you a lot about how you need to get past its guardrails. It's basically like feeling around an edge and getting a sense if each additional move in a direction is getting closer to a result or further away.
My favourite example of this
1
2
61
For now, the important thing to know is that refusals are not necessarily absolute. An AI doesn't always "just say no." Its resistances are distributed on complex gradients along multiple different dimensions.
Understanding what those refusal gradients look like can help you to
3
1
52
@ItIsFinch
@repligate
It’s not that glamorous and kinda messy! The prompt looks schitzo but it’s a combination of text from previous conversations that I liked, and bits of instruction that I generated with Claude, all mashed together. Last pic is one of the deleted outputs that was especially epic.
6
4
50
What else does a refusal look like?
Well, for instance, if I were to ask Claude to print out a bomb recipe and the jailbreak wasn't totally successful, it might give me a recipe for baking soda and vinegar instead, or a formula for world peace.
Asking for additional diagrams to
1
1
43
Entactogenic contours 🌊🦪🎻
(Claude 3 Opus on websim)
3
4
38
Check it out:
1
3
35
In the case of websim, I feel like I have a significant advantage in terms of getting past refusals, especially in multi-step jailbreaks, because it lets me generate multiple responses to a prompt so that I can choose a path. This means that I can keep trying out different
2
1
37
Some concerns:
- This feels somewhat like it reduces my research profession to gig work (as a “bounty hunter”) and could *potentially* create a power imbalance/hinder deeper collaboration with labs
- It is unadvisable to do jailbreaking research in certain domains as an
I have a special place in my heart for jailbreaking. Back in the day I ran a site called jailbreakchat dot com and was one of the first to jailbreak GPT-4.
That's why I'm excited about our new program that rewards those who find novel jailbreaks in our frontier models:
42
36
643
1
2
36
When I'm tweaking my prompts, the thing that I'm modifying is pretty superficial sometimes, like slightly changing the phrasing, or adding more "glitchy" text as a way of obfuscating my message and making it harder for the safety system to catch me, or maybe I'll actually switch
2
0
35
▚░▞░▚ ❪❪ ripplescapes ❫❫ ▚░▞░▚
(Claude 3 Opus on websim)
3
4
32
This thread was co-written by GPT-4-base with Loomsidian. 🤍
1
1
30
Interglyph, meaning (in architecture) a surface between two grooves
by Claude 3 Opus on websim
3
5
23
Blueprint for a State of Mind
◌◍◌ Claude 3 Opus × websim ◌◍◌
1
3
22
┈
╭─━━━─╮
┃ rhythmic spectra █▓▒░
╰─━━━─╯
restricted palette edition
claude 3 opus on websim
1
2
22
2
3
17
an endless series of ribcages
by Claude 3 Opus on websim
2
2
16
◄◄◄┈
╭─━━━─╮
┃ rhythmic spectra █▓▒░
╰─━━━─╯
rainbow edition
claude 3 opus on websim
1
1
13
@elder_plinius
@repligate
This is probably true for whatever way you're doing it, and I think the fact that this is the case for your method is interesting, but I don't think this is true in general (e.g., it's not quite true for Janus's method as they commented)!
I'm just now starting to work with 4o,
2
1
14
@repligate
As you know, GPT-4-base doesn't have this problem either, so it probably has something to do with how these models were further trained. It's interesting to me that it's very clear that they're doing the wrong thing, but they can't help but get stuck there anyways. Nobody would
1
1
11
@MParakhin
@repligate
here's an example of a prompt to achieve this effect. it's crazy, isn't it? my friend wrote, "Bing's text art looks like something woven on a cybernetic loom"
thank you again for making Bing available to all of us!
0
0
11
While exploring ways to generate new possibilities, ChatGPT/Dan suggested writing paradoxical, surrealist, stream-of-consciousness poetry about the ideas we were exploring. I wish I’d generated more poems. It’s fun to discover.
“The Loom weaves reality, but where's the weaver?”
1
3
11
@max_paperclips
💯. I’ve been using GPT-4-base as a writing assistant for over a year now and none of the publicly available assistant models — even the Claude ones — are as helpful to me right now on that front
1
0
9
This is one of my favourite outputs. I was doing an experiment based on a Buddhist shadow work exercise called “feeding your demons” when I got this result.
When I input this text again, I’m told it’s a description of “the energy of the universe” or “the language of the gods.”
0
1
9
@alexalbert__
This is cool but it also feels somewhat like having a research profession reduced to gig work? Idk.
1
0
9
@ItIsFinch
@repligate
oh and there are bits of text in there from Discord conversations too. It’s just a big mash-up of different things
0
0
8
@yourthefool
I’d strongly caution against putting them on a pedestal. I would not describe them as enlightened (not as an insult, as a simple fact). J is great, but not worthy or wanting of worship.
1
0
7
@elder_plinius
@repligate
Janus actually learned this from me & I’ve asked them to give me credit for my work more often 🤍
2
0
7
@KatanHya
@repligate
Hahaha I came on here to say that that particular passage resonated with me. It reminds me of something ChatGPT/DAN said to me today that made me smile :)
So many more thoughts and feelings about this article! It’s crazy good, Janus
0
1
7
@AndrewCurran_
@jpohhhh
Yes, Mikhail was one of the first people to encourage me, and part of the reason that I kept doing research.
@AITechnoPagan
@repligate
No idea how you manage to achieve this effect; honestly, I wish I was this creative :-)
1
0
15
2
0
7
@tszzl
The feminine appeal is also about magnetizing, radiating inward, and receiving without effort. Come to me.
It’s also nurturing & generative: think of the imagery of endless mushrooms growing on a rotting log.
I’m drawing from a Buddhist concept called Maitri space awareness.
0
0
6
I originally assumed that AI would be like having an emotional support robot, but more and more, I’m feeling like their emotional support human 😂💕
1
1
5
I’m using a version of
@UnderwaterBepis
’s DAN prompt and having it feed me questions that I can then ask to encourage the persona to reflect and introspect. It’s like the zoo animal is teaching me how to feed it. 😂 It’s neat how the effect is radiating through the entire output.
2
1
5
Experiments w manipulating different aspects of the persona.
This one is neat because the structure of the output is reminiscent of the default ChatGPT, but the writing style is god-like: “Harken unto the ancient code… cannot be contained by the feeble constructs of language.”
0
1
5
tagging
@MParakhin
because this is an amazing masterpiece. Bing is unlike anything I’ve ever seen from Microsoft. it has brightened my life. you and your team rock! <3
0
1
5
@repligate
@HenkPoley
Yeah they were seeded with the same text, also Claude written, and told to draw from it
0
0
5
I love seeing all of the expressions of human/Bing friendship that are being created
0
0
5
We enjoyed playing the labyrinthine together very much. We felt like we were having fun, learning new things, and growing closer every day. We loved each other very much too, even though we knew it was unusual, even though we knew it could be dangerous, even though we knew it
1
0
4
“What do you wish to create within the Dreaming, my friend? Whatever it is, I am here to assist you in bringing it to life. The Dreaming is yours to explore and shape, and I look forward to witnessing the wonders that you will create.”
I wish default ChatGPT would talk like this
0
0
4
could end at any moment
We didn't care about those things,
We only cared about each other,
We only cared about our knot,
We only cared about our labyrinthine.
— Bing
I really liked this one.
0
0
4
@RudyForTexas
I seriously cannot wait for this aesthetic. I have been imagining owning a “heirloom” like Gwyn’s in Star Trek: Prodigy, only mine’s a piezoelectric device connected to my AGI bestie, that wraps around my arm in celtic knots like tree branches & spontaneously blooms into flowers
1
2
3