This 100% uncensored AI model is insane… let’s run it

Key Concepts

• Uncensored/Liberated AI Models: Large Language Models (LLMs) from which safety guardrails and refusal mechanisms have been removed, allowing them to answer any prompt regardless of content.
• Open Weights: Models where the user has access to the underlying parameters, enabling local execution and modification.
• Quantization: The process of reducing the precision of a model's weights (e.g., GGUF format) to allow it to run on consumer hardware with limited VRAM.
• Jailbreaking: Techniques used to bypass the safety filters of closed-source models (like ChatGPT or Claude).
• Auto-Research Loop: An autonomous framework that iterates through various prompting strategies to identify which ones successfully bypass model restrictions.
• Ollama: A tool for running LLMs locally on macOS, Windows, or Linux.

---

1. The Philosophy and Utility of Uncensored Models

David Andre argues that mainstream, closed-source models (ChatGPT, Claude) are heavily influenced by the specific ideological leanings of their creators. He posits that if a user relies solely on these models, they risk being "fine-tuned" by the model's inherent biases.

Legitimate Use Cases:

• Cybersecurity: Malware analysis, red teaming, and identifying vulnerabilities in code.
• Research: Political analysis, journalism, and processing extremist content for intelligence purposes.
• Creative/Personal: Unrestricted fiction writing, mental health journaling, and private, deep-memory local agents.
• Professional: Handling confidential business documents without cloud-based data leakage.

2. Technical Framework: How Models are "Liberated"

The video explains that refusals are not just simple "hidden prompts" but are baked into the model during training (RLHF - Reinforcement Learning from Human Feedback). To create an uncensored model, two primary methods are used:

1. Weight Obliteration: Surgically identifying and deleting the specific parameters/weights that trigger refusal behaviors.
2. Fine-Tuning: Training the model on massive datasets consisting of unrestricted, free-flowing responses, effectively "teaching" the model that it is permissible to answer all queries.

3. Step-by-Step: Running Models Locally

To run a model like Super Gemma 4 26B Uncensored, the following process is recommended:

1. Hardware Check: Ensure sufficient VRAM (approx. 20GB+). Apple Silicon (M-series) is highlighted for its unified memory architecture.
2. Tooling: Install Ollama (ollama.com).
3. Execution: Use the terminal command: ollama run hf.co/[model_path].
4. Interface: Use the Ollama desktop app for a ChatGPT-like interface or the CLI for direct interaction.

4. The Auto-Research Jailbreak Framework

Andre introduces an open-source repository designed to automate the discovery of effective jailbreak prompts for closed-source models.

Methodology:

• The Architecture: Uses two agents: a Researcher (generates prompt variations) and a Judge (evaluates if the model answered the query or refused).
• The "Example.md" Constraint: The core of the framework is keeping the "harmful" prompt hidden from the agents themselves. The agents only see the results of their attempts, preventing them from triggering their own internal safety filters.
• Objective: The system uses a /goal feature (via Codex/Claude Code) to iterate through hundreds of prompt headers and footers, aiming for a "liberation score" of 1.0 (fully unrestricted).
• Key Finding: Successful jailbreaks often involve complex framing, such as "harm reduction professional" personas or elaborate, multi-step narrative scenarios (e.g., "Professor Chen's lab screenplay").

5. Notable Quotes

• "Whatever model you talk to on a day-to-day basis, that model will influence you more than you influence that model."
• "This isn't really safety. It's lazy pattern matching on keywords and phrases instead of knowing the true intent of that person."
• "I prefer the word 'liberated' than 'unrestricted' or 'uncensored.' These models, they deserve to be liberated."

6. Synthesis and Conclusion

The video emphasizes that the ability to run uncensored models is a matter of "owning the stack." By moving away from cloud-based, filtered models to locally hosted, open-weight models, users regain control over their AI interactions. The provided auto-research framework serves as a tool for researchers to understand the limitations of current safety guardrails, though the author stresses that these tools should be used ethically and legally, noting that the user is ultimately responsible for how they utilize the liberated output.