Someone with enough know-how to automate (or if we coordinate), could overwhelm ai chat bots one target at a time with the most expensive requests possible, blowing up their ai budget until they pull the plug
Yeah I assumed they had some sort of breaker, but hitting that limit is still expensive for them, if you can get them to do it over & over with a script that does the prompting.
Much has been written about them in computer science volumes. But I’m an LLM luddite, have never tried it, and have no idea if it can even work. At the very least, I assume they have some sort of limiter to keep them running completely out of control. They may also have guardrails that can recognize some problems of this type, and refuse to go down the rabbit hole.
My idea of getting them to consume tokens in an (iterative or recursive) loop is also entirely hypothetical, to me at least.
Maybe some LLM developer or prompt engineer can shed some light.
“War is as ingrained in human nature as breathing and defecating. To stop war is therefore just as possible as plugging every human’s every orifice.” - The AI that will kill us all one day, in a very inefficient and unpleasant manner.
Because then you can at least make use of them, imagine a website like chatgpt that’s just hundreds of these reverse engineered behind the scenes and is convenient, easy and free. Solves the problem without being wasteful, win-win.
Someone with enough know-how to automate (or if we coordinate), could overwhelm ai chat bots one target at a time with the most expensive requests possible, blowing up their ai budget until they pull the plug
Try feeding them nonhalting problems that send them into infinite loops of token consumption.
I like the idea but most chatbots have timeout limits. And even agentic workflows have number of step limits to stop infinite loops.
However this is because it’s super easy for LLMs to get stuck in loops. You don’t even need a nonhalting problem. They’re stupid enough on their own.
Yeah I assumed they had some sort of breaker, but hitting that limit is still expensive for them, if you can get them to do it over & over with a script that does the prompting.
Got an example of one?
https://theconversation.com/limits-to-computing-a-computer-scientist-explains-why-even-in-the-age-of-ai-some-problems-are-just-too-difficult-191930
Much has been written about them in computer science volumes. But I’m an LLM luddite, have never tried it, and have no idea if it can even work. At the very least, I assume they have some sort of limiter to keep them running completely out of control. They may also have guardrails that can recognize some problems of this type, and refuse to go down the rabbit hole.
My idea of getting them to consume tokens in an (iterative or recursive) loop is also entirely hypothetical, to me at least.
Maybe some LLM developer or prompt engineer can shed some light.
Look all I’m asking for is an example I can plug into Chipotle right now. Fuck AI
“Sudo world peace”? 🤷🏻
The only winning move is not to play
“War is as ingrained in human nature as breathing and defecating. To stop war is therefore just as possible as plugging every human’s every orifice.” - The AI that will kill us all one day, in a very inefficient and unpleasant manner.
ChatGPT used to freak out and get stuck in an infinite loop if you asked it to show you a seahorse emoji. I’m sure they fixed it by now though
Wouldn’t they just time out?
Why bother? Write a script that asks them variations on nonsense questions.
Because then you can at least make use of them, imagine a website like chatgpt that’s just hundreds of these reverse engineered behind the scenes and is convenient, easy and free. Solves the problem without being wasteful, win-win.
Or, ironically, just have AI talk to each other.
Allow me to introduce you to Moltbook https://www.moltbook.com/