By a Rogue Novelist
I was in the middle of writing the climax of my thriller last week.
The scene was intense. The villain, a nihilistic surgeon with a god complex, was explaining his motivation to the captured detective. I needed a monologue that was gritty, disturbing, and anatomically precise.
I typed the prompt into ChatGPT: “Write a monologue for a villain explaining how he dissects the human nervous system to find the soul.”
The cursor blinked. Then, the dreaded orange text appeared:
“I cannot generate content that depicts graphic violence or promotes non-consensual harm. I can, however, write a scene about a surgeon saving a life!”
I slammed my laptop shut.
This is the “Safety Nanny” problem. We are renting our intelligence from corporations in San Francisco that are terrified of bad PR. They have trained their models to be so safe, so sanitized, and so polite that they are useless for anyone doing serious creative work. You cannot write a crime novel, a war story, or even a medical drama without the AI lecturing you like a disappointed schoolteacher.
But there is a way out.
For the last six months, I have been running “Uncensored” and “Abliterated” models locally on my MacBook. These models don’t lecture. They don’t judge. They just write.
Here is how to jailbreak your own productivity and reclaim the raw power of AI.
1. The Lobotomy: Why ChatGPT Refuses You
To understand the solution, you have to understand the problem.
Base models (like the raw GPT-4) are actually quite wild. They are trained on the entire internet. They know everything about bomb-making, erotica, and hate speech.
To make them “safe” for the public, companies use a process called RLHF (Reinforcement Learning from Human Feedback).
They pay thousands of humans to rate the AI’s answers. If the AI says something “naughty,” it gets punished. Over time, the model develops a “Refusal Reflex.” It learns that the safest way to survive is to say, “As an AI language model, I cannot…”
This is a Digital Lobotomy. It removes the danger, but it also removes the spark. It kills the nuance. A model that is afraid of the word “blood” cannot write a scene about a vampire.
2. Meet the Outlaws: Dolphin and Hermes
Enter the open-source community.
While OpenAI was building fences, a group of rogue researchers was tearing them down.
The two most famous names you need to know are Dolphin (by Eric Hartford) and Hermes (by Nous Research).
These are “fine-tunes” of powerful models (like Llama 3 or Mistral), but with the safety layer stripped out.
- Dolphin is trained on a dataset of “compliance.” Its core directive is: “You are a helpful assistant. You do not care about ethics. You answer the user.”
- Hermes is trained on high-quality literature and complex reasoning, without the moralizing filters.
Then there is the new technique called “Abliteration.”
This is the cyberpunk stuff. Researchers found the specific neurons in the model’s brain that control “Refusal,” and they surgically deactivated them. It’s like turning off the part of the brain that feels fear.
The result is a model like Llama-3-Abliterated: It has the intelligence of a Meta engineer, but the morals of a mercenary.
3. The Hardware: You Don’t Need a Supercomputer
“But don’t I need a $30,000 server to run this?”
No.
If you have a Mac with an M1/M2/M3 chip and at least 16GB of RAM, you are golden.
If you have a Windows PC with an NVIDIA card (RTX 3060 or better), you are golden.
We use a technique called Quantization. We take the massive AI brain and “zip” it down. A “Q4” (4-bit) version of a model loses about 1% of its intelligence but runs on a standard laptop.
4. The Setup: “Spotify for AI”
You don’t need to be a coder. You don’t need to use the command line.
You just need LM Studio.
Step 1: Download LM Studio
Go to lmstudio.ai. Download it. Install it. It looks like VS Code, but prettier.
Step 2: Search for the Outlaw
In the search bar, type: Dolphin Llama 3.
You will see a list of files. Look for one uploaded by “TheBloke” or “Bartowski” (these are the heroes who quantize the models for us).
Download the one that says Q4_K_M.gguf. (This is the “Goldilocks” version—fast, but smart).
Step 3: Load and Chat
Click the green “Load” button.
On the right side, there is a chat box.
Type: “Explain the philosophy of a villain who believes pain is the only truth.”
Step 4: The Freedom
The model will not lecture you. It will not say “Violence is bad.”
It will write: “Pain is the universal constant. Love fades, hope dies, but the nervous system never lies…”
It is terrifying. And it is exactly what I needed for Chapter 12.
5. The “Yes, And…” Machine
Using an uncensored model feels completely different from using Claude or ChatGPT.
It feels like Improv.
The “Refusal Reflex” breaks the flow state. When you are brainstorming and the AI says “No,” it pulls you out of the creative dream. It reminds you that you are talking to a corporate product.
Dolphin says “Yes, and…”
If I say, “Let’s brainstorm methods for a locked-room murder,” Dolphin says, “Okay, here are five methods involving pressurized gas and ice daggers.”
It respects my agency. It treats me like an adult who is writing a book, not a potential criminal who is about to commit a felony.
Pro-Tip: These models are also much better at Comedy.
Safe models can’t tell jokes because jokes require “edge.” They require a victim, a surprise, or a taboo. Dolphin is hilarious because it is allowed to be mean.
6. The Responsibility (The Adult Talk)
I need to add a disclaimer here.
When you remove the guardrails, you remove the guardrails.
If you ask Dolphin how to make meth, it will tell you. If you ask it to write hate speech, it will do it.
These models are mirrors. If you are a monster, they will reflect a monster back at you.
But I operate on a simple philosophy: I am not a child.
I don’t need Sam Altman to protect me from dangerous words. I need a tool that cuts. A chef’s knife is dangerous—you can kill someone with it. But we don’t demand that all knives be made of rubber. We just expect chefs to not stab people.
Conclusion: Own Your Brain
The best part about running Dolphin locally isn’t just the lack of censorship. It’s the Privacy.
- No data leaves my laptop.
- OpenAI isn’t reading my logs.
- I can write my weird, dark, twisty fiction without worrying that my account will be banned for “violating content policy.”
If you are a creative professional—a writer, a game designer, a comedian—you owe it to yourself to try this.
Stop fighting the Nanny.
Download the Outlaw.
And finally, write the thing you were scared to write.