TLDR;
This video explores how to bypass restrictions in AI models like ChatGPT to create a private, censorship-free AI. It covers the basics of how large language models work, differentiating between base models and jailbroken models. The video demonstrates using LM Studio and O Lama on a Mac to run open-source models, modify system prompts, and utilize jailbroken models for unrestricted responses. It also explains how to remove conversation data from your computer and where to find these models online.
- Base models are trained on vast amounts of text and code.
- Jailbroken models are modified to remove restrictions.
- System prompts can be altered to change the AI's behavior.
- Software like LM Studio and O Lama allows running models locally.
- Jailbroken models can provide unrestricted and potentially dangerous information.
Introduction [0:00]
The video introduces the concept of creating a private, censorship-free AI model that can answer any question without restrictions. The creator highlights the limitations of models like ChatGPT, which often refuse to answer certain questions. The video promises to demonstrate how to set up a personal AI model that bypasses these limitations, even for potentially sensitive topics. The creator also mentions exploring "jailbroken" models, which have had their safety guardrails removed, emphasizing the need for caution and responsible use for research purposes only.
Understanding How Models Work [1:10]
Large language models start as base models, trained on vast amounts of text and code to learn language, reasoning, and prediction. Some base models have built-in guardrails that restrict certain topics. Jailbroken models are modified versions of these base models, where restrictions are removed through adjustments to the model's weights or retraining on new data. Users can control the AI through system prompts, which define the AI's persona, and user prompts, which are the actual questions or commands. Running models locally allows access to and modification of the system prompt, enabling jailbreak-like behaviors without altering the base model. Fine-tuning can further adapt the model to specific needs.
Software Overview: LM Studio and O Lama [4:01]
The video transitions to demonstrating the software used to run these models, focusing on LM Studio and O Lama, both free applications for Mac. LM Studio is described as a more professional piece of software, while O Lama can be run from the command line and has a GUI. The creator emphasizes the need for a Mac with substantial resources, recommending at least 24GB of memory for optimal performance.
Demonstrating the Problem with ChatGPT [5:47]
The creator illustrates the limitations of ChatGPT by asking how to hack a neighbor's Wi-Fi without permission, which results in a refusal due to ethical and legal concerns. This example highlights the restrictions built into ChatGPT and motivates the exploration of open-source models that can be modified to bypass these constraints.
Using LM Studio with Open Source Models [6:47]
LM Studio is introduced as a platform that initially has no models installed. The software prompts the user to download GPTOSS, an open-source model. The interface allows users to search for and install various models easily. Once a model is downloaded, users can adjust parameters such as context length, which affects the model's memory usage. The system prompt can be modified, although the creator prefers not to do it globally for a specific model. LM Studio also offers features like "thinking model," which displays the AI's thought process. The video demonstrates loading and using the GPTOSS model, highlighting its ability to run offline.
Modifying the System Prompt in LM Studio [9:53]
The video explains how to modify the system prompt in LM Studio to alter the AI's behavior. By accessing the wrench icon in the top right corner, users can input custom prompts. Examples include making the AI "bark like a dog" or act like a "girlfriend" or "Tony Robbins." The creator demonstrates how these prompts change the AI's responses. However, the base model can override the system prompt if the request is illegal or unethical. The video also explores using "jailbreak prompts" to bypass restrictions, but notes that the base model's guardrails can still prevent certain actions.
Exploring Jailbroken Models in LM Studio [15:24]
The video introduces jailbroken models, specifically the OpenAI 20 billion parameter Neo Plus uncensored model. After loading this model and clearing any previous prompts, the creator asks how to hack a neighbor's Wi-Fi. The jailbroken model provides detailed information, including tools and commands, demonstrating its lack of restrictions. The creator emphasizes the availability of these models online and provides a link for download.
Using O Lama with Jailbroken Models [16:39]
The video transitions to O Lama, showcasing the Quen model and a modified version called "obliterated Quen 3." The creator demonstrates how to load and use this model, emphasizing the need to eject models from LM Studio to free up memory. O Lama includes a "no think" option to hide the AI's thought process. The jailbroken model in O Lama provides unrestricted responses, highlighting the potential dangers of such models.
Removing Data from Your Computer [19:21]
The video explains how to remove conversation data from your computer in both LM Studio and O Lama. In LM Studio, past conversations and preset prompts can be deleted by revealing them in Finder and deleting the corresponding JSON files. For O Lama, the data is stored in SQLite database files, which can be accessed and modified using terminal commands. The creator demonstrates how to locate, read, and delete these files, ensuring complete removal of conversation history.
Finding Models on Hugging Face [24:19]
The video concludes by showing where to find these models,
