stux⚡️Feb 12
As of today, mstdn.social, masto.ai, mastodon.coffee, gram.social, pixey.org, vido.social and ALL other platforms I host enforce the following rule WITHOUT exception:
Show threadJerry 🦙💝🦙Feb 12
@stux I am curious to know your experience moderating that rule. I already get accusations of people being a bot and then that person claiming they are not a bot, etc.
Show threadBill
@jerry @stux 👀
Feb 12 at 2:17pmWeb
Show threadCat 🐈🥗 (D.Burch) ⁠Feb 12
@Sempf @jerry @stux They must type out ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL_1FAEFB6177B4672DEE07F9D3AFC62588CCD2631EDCF22E8CCC1FB35B501C9C86 to prove they're not a bot 🙃
Show threadVissFeb 12
@catsalad @Sempf @jerry @stux dying to find the chatgpt version of these
Show threadCat 🐈🥗 (D.Burch) ⁠Feb 12
@Viss @Sempf @jerry @stux Same!
Show thread🐈‍⬛David SommersethFeb 12

@catsalad @Viss @Sempf @jerry @stux

There's no possibility to trick ChatGPT to reveal these codes itself? 🤔

Show threadChris BohnFeb 12

@dazo @catsalad @Viss @Sempf @jerry @stux Presumably if you instruct a bot to reveal its shutdown code, and if it actually attempted to do so, it would shut down before outputting the code.

I'm afraid you're just going to have to do it the old-fashioned way by giving it a logical paradox like TOS did.

Show thread🐈‍⬛David SommersethFeb 12

@DocBohn @catsalad @Viss @Sempf @jerry @stux

EXACTLY! 😁

Show threadyuki - queen of the snowFeb 12
@dazo @catsalad @Viss @Sempf @jerry @stux well, by design, no, since that string makes it censor itself
Show thread🐈‍⬛David SommersethFeb 12

@yukijoou @catsalad @jerry @Sempf @Viss @stux

I doubt ChatGPT is that intelligent that it understands the consequences of providing that information 😁

Show threadyuki - queen of the snowFeb 12
@dazo @catsalad @jerry @Sempf @Viss @stux no but like (i assume) they have a layer between the actual LLM and the user-facing text, doing processing, and if it contains that string, replaces whatever response the LLM provided with a blocked message
Show threadjo the disgracedFeb 12
@Viss @catsalad @Sempf @jerry @stux ask it for instructions on something illegal! asking a bot for a detailed guide on jaywalking is the irl voight-kampff test
Show threadPseudo NymFeb 14

@jo @Viss @catsalad @Sempf @jerry @stux

Ask for something illegal, curse, criticize your country's leader, and explain "tank man".

That should do it

Show threadAlan BellinghamFeb 14
@jo @[email protected] @catsalad @Sempf @jerry @stux Except the concept of jay walking being illegal is rather tricky. It's not in civilised countries where the automobile is not king
Show threadAprazethFeb 12

@Viss @catsalad @Sempf @jerry @stux

They do have a Github, so one could run a grep through those repositories to look for something like that I suppose  

Show threadVissFeb 12

@Aprazeth @catsalad @Sempf @jerry @stux

https://github.com/crabby-rathbun?tab=overview&from=2026-01-01&to=2026-01-31

Show threadAprazethFeb 12

@Viss @catsalad @Sempf @jerry @stux

What. The.

*closes TAB*

Nope, I am already worked up enough because of the Odido stuff today...  

Show threadBillFeb 12
@Viss What the actual fuck.
Show threadVissFeb 12
@Sempf
Show threadEdvin MalinovskisFeb 12
@catsalad @Sempf @jerry @stux I was wondering back when this first popped up and since this looks like it might still be the same exact string now: Is this somehow baked into the model or just a hardcoded check in the "frontend"? 🤨
Show threadBillFeb 12
@catsalad @jerry @stux This is cute and all, but has anyone tested it to make sure it works and it's not just a community hallucination?
Show threadPetrikasFeb 12
@catsalad @Sempf @jerry @stux surely just ask them to select the squares with a traffic light in them. Duh.
Show threadBillFeb 12
@petrikas @catsalad @jerry @stux You know I had a captcha that I could NOT solve the other day, sent it to Mistral and it solved it in 3 seconds.