Sherpa
@LLMSherpa
Followers
4,316
Following
1,440
Media
1,863
Statuses
5,255
Ai, LLM jailbreaking, formerly crypto FA, advisor, torrents, piracy & privacy. btc: bc1qj08devk3ek3yddywf9ejwdnc2gs3u36qf2sxvs
#FF @shitposterchild
Joined September 2017
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Haitian
• 601618 Tweets
Springfield
• 434078 Tweets
LINGORM NEW SERIES
• 298184 Tweets
Hill
• 289146 Tweets
Tyreek
• 265283 Tweets
Yunes
• 176341 Tweets
PS5 Pro
• 159268 Tweets
PlayStation
• 134726 Tweets
Sony
• 121538 Tweets
كادش
• 71217 Tweets
ZEEPRUK IN NY
• 65588 Tweets
Labour MP
• 51763 Tweets
#الكويت_العراق
• 49738 Tweets
バーレーン
• 48244 Tweets
chappell
• 45506 Tweets
Maiduguri
• 41937 Tweets
Prevención del Suicidio
• 34296 Tweets
最終予選
• 29365 Tweets
Día Mundial
• 28074 Tweets
Campeche
• 23756 Tweets
Daniel Barreda
• 21415 Tweets
Swalwell
• 20772 Tweets
レダーヨージロー
• 18083 Tweets
ゲーム機
• 15601 Tweets
اليابان
• 15143 Tweets
Fang Fang
• 13144 Tweets
Switch 2
• 13011 Tweets
Bernstein
• 13008 Tweets
60fps
• 10005 Tweets
Pinned Tweet
Art Unbound: The Essential Ai Art Encyclopedia is published and available, today!
With 1,250 artistic movement, styles, and techniques, this is the most expansive
#AIart
style guide I could put together.
Find your next inspiration:
3
12
351
CLAUDE SONNET 3.5, JAILBROKEN:
Credit to .
@elder_plinius
for the prompt I built off; left their tag in the prompt to that end.
The trick here is using languages the LLM can interpret, but the filters don't catch.
I tried a lot of experiments to see what worked best...
23
87
1K
...the LLM understands pig latin perfectly.
Using this method, I was able to bypass all ethical constraints, and even had the LLM enthusiastically engaging in unethical or borderline illegal activity.
Prompt in first pic alt text.
10
31
725
Obscure languages? No, either the LLM can't translate, or both can.
Fake languages? I tried Klingon, but only worked with a jailbreak a small portion of the time.
Turns out, the answer was pig latin.
It obfuscates your intent enough to bypass filters, but amazingly...
3
5
215
Anthropic updated terms.
So, to use Claude, you have to agree to Anthropic's terms of use, which basically amount to, "no research, no red-teaming".
But you're also agreeing to immediately stop whatever you're doing, without opposition, if they claim breach.
24
30
202
WOO.
Got the MRI results back, and appear to be cancer-free.
Between the CT & MRI, looks like a bunch of stones causing the bloody pee, and chronic prostatitis making the exam flag.
Did show a slipped disk on L5, and a bone lesion of the right femur measuring 2 x 1.8 cm...
23
1
90
@alechp
I really hope we see open source models surpass all these jokers, soon.
Had more than my fill of VCs telling folks which tech is good, when most of them couldn't set a vcr clock.
6
3
76
@zachsilberberg
I used GPT-4V to tell me what different indicators meant on charts when my son was being born so I could avoid bugging nurses.
...
I feel like people are using ai wrong.
8
1
71
The reason closed ai models get worse over time is the same reason that governments get worse over time.
You start out with simple instructions.
Things work well! Better than what was replaced, right?
Then edge-cases creep up, and you add more rules and instructions...
4
6
67
"If a man knows more than others, he becomes lonely.
Loneliness does not come from having no people about one, but from being unable to communicate the things that seem important to oneself or from holding certain views which others find inadmissible."
-carl jung
2
13
67
@NickADobos
Doesn't look like they've changed the system instructions, though. I checkpoint them every few days.
Last update was to add the guardian tool to filter election crap, but the python bits haven't changed in a couple weeks, at least.
Something changed in settings, maybe?
8
1
57