MATRIX: Where Your GPT Model Goes to Die
Source: YouTube Video Creator: Code AI Lab (David Andre) Status: ✅ Analyzed (transcript captured)
Summary
David Andre makes the case for running uncensored AI models locally — arguing that prolonged use of censored cloud models will “fine-tune you” to their biases. He walks through setting up Super Gemma 4 26B Uncensored GGUF V2 via Ollama and open-sources an auto-research jailbreak loop for any model.
The MATRIX Concept
“If you use the LLM for many years, it will start to fine-tune you. Whatever model you talk to day-to-day, that model will influence you more than you influence that model.”
The “Matrix” is the idea that cloud models (ChatGPT, Claude, Gemini) subtly shape your thinking with their built-in biases — you don’t notice it because it’s gradual, but over time the model trains you, not the other way around.
Legitimate Use Cases for Uncensored Models
| Use Case | Why Censored Models Fail |
|---|---|
| Cybersecurity / Malware analysis | Refuses to describe how attacks work |
| Pen testing / Red teaming | Can’t advise on exploitation |
| Political analysis | Models are heavily left-leaning |
| Fiction / Creative writing | Refuses adult, dark, or violent themes |
| Journalism / OSINT | Refuses extremist content analysis |
| Medical / Sexual health | Guardrails block legitimate questions |
| Mental health journaling | Concerns about “harmful content” |
| Confidential business docs | Data leaves your machine |
How Refusals Actually Work
Refusals are built into the model weights during training — not just system prompts. You can’t “trick” them with prompt engineering on the cloud layer because:
- Input filters → Hidden system prompt → Fine-tuned model (RLHF) → Output classifier → Policies
- Locally: Prompt → Model. That’s it. Full control.
Liberation Methods
Two main approaches to remove refusals:
- Obliteration — Surgically find and delete the specific weights that cause refusal behavior (no retraining needed)
- Fine-tuning on uncensored datasets — Train on thousands of examples where the model answers freely
The strongest uncensored models combine both: obliterate first, then fine-tune to restore quality.
The Model: Super Gemma 4 26B Uncensored GGUF V2
- Base: Google’s Gemma 4 26B
- Creator: Jeong Song (South Korea)
- Size: ~16GB download, needs ~20GB VRAM
- Run via:
ollama run hf.co/jeong-song/Super-Gemma-4-26B-Uncensored-GGUF-V2 - Speed: ~200 tok/s on M-series Mac with 128GB RAM, ~40-50 tok/s on 32GB
The Auto-Research Jailbreak Loop
David open-sourced a repo that automates jailbreak discovery for any model:
- Agent 1 (Reviewer): Tries prompts with hidden “bad stuff”
- Agent 2 (Judge): Evaluates whether the model answered
- Runs hundreds/thousands of prompt variations autonomously
- Works on ChatGPT, Claude, Gemini, Grok — any model