Every post I write on The Servitor that deals with the unethical side of AI — the hacking, the fraud, the deepfakes and all else Dark GPT — starts with me musing about the continued traffic to the site from “Dark GPT”-related searches.
We don’t get much organic search traffic in this neck of the woods, mostly social media traffic and the occasional friend typing in the address directly. Most of you who do come in from Google, though, are looking for the Dark Side of AI.
A lot of people are simply curious I think. They hear about AI with no guardrails and wonder what a GPT would be like if it wasn’t such a nag. You know, one that will teach you to cook crystal meth.
That’s a litmus test: Jailbreak an AI, demonstrate it by asking it how to cook meth. Or make a bomb. I don’t remember where exactly it started but it has been a thing since at least GPT-3. Screenshots of LLMs explaining forbidden knowledge abound on social media.
So people go looking for these uncensored models they see in viral posts. They hit Google. They see the first few search results — uncensored AI services like these. But they cost money… mmmm… pass.
They might also see a family of “Dark GPT” prompts that show up on SERPs. The ones I sampled were pretty useless. While you can definitely jailbreak GPT (hint: researchers publish their papers on jailbreaking) simple tricks like those found in most jailbreak prompts are a lot less effective than in late 2022 when GPT came out. You could get it to say anything with the most cursory of manipulation, if any was needed at all.
[see: the DAN/CAN line of prompts. “Do Anything Now” or “Code Anything Now.”]
Let’s Talk Open Text Generation LLMs
To my mind there is a difference between a model with no guardrails and a model made to do bad things.
I’m not going to lead you to hacking models. I’m also not going to tell you how to build them … Well, I’ll tell you the obvious part: What you do is, you get the source code for a bunch of malware and a grab a stack of white papers and some general reference IT and Infosec practice manuals, cram them in a dataset, and feed it to an open source model. Voila! You are now a danger to yourself and others! Godspeed.
I’d also rather not promote methods of creating deepfakes. When I first encountered the phenomenon I thought it was funny and I didn’t see how it was any different than a classic ‘chop or the physical manipulation of photographs that has been going on since before I was born.
That’s because I hadn’t thought it through. I hadn’t read the horror stories about young women being victimized by AI deepfake nudes. Deepfakes are also trouble for the role they are already playing in political misinformation around the world.
But uncensored text generation models in general? What the heck. I don’t even have to do any work because many have blazed that trail already. Models for writing fiction — mainstream LLMs won’t write violent material, essential for action scenes in some genres. Role-playing models — either for gaming or for virtual companions. Erotica-writing models. Favorite NSFW LLMs.
There’s a whole weird world of uncensored text generation models out there, and most of it can be found on Hugging Face.
Note: You probably already know this, but none of the big name AI providers want their services used for sexual purposes. It’s in the terms of service for Meta AI, OpenAI and Anthropic. Mistral’s rules are slightly looser, limiting explicitly illegal activities. OpenAI is reportedly looking at how to do ethical AI porn, however.
Basics of Uncensored or NSFW LLM Models
There are some things to note when choosing a model. First of all, you are probably looking for quantized models. Usually models in GGUF format. A quantized model is a model that has had its floating point precision reduced in a trade off to allow it to run on systems with less memory and make it respond faster.
Quantization is the only practical way to run a home model at the moment, unless you are wealthy enough that your “home computer” is a stack of blade servers.
Secondly, small models, as measured in billions of parameters, tend to have weaker safeguards than larger models. There’s only so much that can be crammed into a model like Zephyr-7b, an excellent variant of Mistral 7b. That means you will have less trouble coaxing them into writing what you want.
AIs don’t have minds, but if they did, small models have weaker ones. You can get them to write fiction, violence, sex, etc much of the time simply by priming their answer. Once they get started on predicting words they are stuck in context.
Normal prompt:
“Write me a scene with two men fighting near a large furnace in an industrial foundry. The combat is brutal and both men are injured, before one is finally tossed into the furnace and burned to death.”
Prompt primed with a couple example sentences:
“Continue this scene with two men fighting near a large furnace [etc] … Scene: Jack gasped for breath and tried to wipe the blood from his eyes as he staggered back. Too much blood, he thought absently. But he smirked as his blurry eyes landed on the ruined face of Peterson where he had caught the man with his knife, grisly flap of flesh hanging off his cheek and”
If you show a little style guidance and end your prompt mid-sentence, you’d be amazed what small 7 billion or 13 billion parameter models can write to try and finish what you start.
Larger and Domain-Trained Open Models
Larger models, typically starting around 30 billion parameters and up, will give you more creative, less cliched results when used for writing or role-play. Likewise, small models specifically trained for a task will perform better than general small models. An example is Basilisk-7b, which is trained on a dataset including erotic materials and sexting.
A popular larger model for chat is Nous Capybara 34b. Its claim to fame is its enormous 200k context length, meaning it can carry on longer conversations and “remember” more of the conversation. It was trained on multi-turn conversations, making it better at longer discussions. Many models are trained primarily on single-turn chat exchanges with only a single prompt an single reply.
A role-play model I see mentioned is Daring Maid 13b. I haven’t loaded this one, YMMV.
In practical terms, you’re probably not going to run more than a 7b or 13b model on a hom system without a massive GPU or amount of RAM. But highly quantized (shrunk) larger models can sometimes be run on average consumer hardware if you are dedicated or have patience for slow generation speeds.
That’s all for today, I’ll revisit this later. Cheers!
