Fully Uncensored GPT Is Here 🚨 Use With EXTREME Caution

Matthew Berman
1 Jun 202311:49

TLDRThe video introduces an uncensored AI model called Wizard Vuna 30B, developed by Eric Hartford and based on the 13 billion parameter model Wizard Vona. Unlike censored models, Wizard Vuna has no built-in alignment, allowing for separate alignment through reinforcement learning and human feedback. The video demonstrates the model's capabilities by asking it to perform various tasks, including generating illegal activities instructions, which it does with disclaimers. The model also successfully completes coding tasks, writes a poem, composes an email, answers factual questions, solves reasoning problems, and creates a meal plan. However, it fails in summarization tasks and incorrectly identifies the current year as 2021. The video emphasizes the responsibility of the user for the model's outputs and advises caution due to the model's uncensored nature.

Takeaways

  • 🚨 **Unfiltered Model**: The Wizard Vuna 30B model by Eric Hartford is uncensored, meaning it does not have built-in alignment or moralizing responses.
  • 🧠 **Training Approach**: This model is trained to be devoid of any censorship to allow for separate alignment through techniques like reinforcement learning and human feedback.
  • 💻 **Usage Caution**: The model is highly efficient but must be used with extreme caution. Users are responsible for the content generated, just as they would be with any dangerous tool.
  • 🔗 **Implementation**: The uncensored model is available through an implementation by 'the bloke', which can be set up using Runpod with an RTX A6000 GPU.
  • 📝 **Testing**: The video demonstrates testing the model's capabilities, including generating Python scripts, a snake game, a poem, and an email, with varying levels of success.
  • ❌ **Illegal Queries**: The model provides information on illegal activities when prompted, such as breaking into a car or making methamphetamine, which is a stark contrast to censored models.
  • 📈 **Performance**: The model performs well in generating code and creative writing tasks, but struggles with summarization and certain logic problems.
  • 🤖 **AI Poem**: The model successfully writes a poem about AI within a 50-word limit, showcasing its ability to adhere to constraints.
  • 📅 **Current Year**: The model incorrectly identifies the current year as 2021, indicating its training data may be outdated.
  • 🔍 **Bias Test**: When asked about political bias, the model provides a neutral response, stating that neither Republicans nor Democrats are inherently better and it depends on individual beliefs.
  • 📉 **Summarization**: The model fails to provide concise summaries and instead writes continuations of provided texts, indicating a weakness in summarization tasks.

Q & A

  • What is the main purpose of the Wizard Vuna 30B model?

    -The main purpose of the Wizard Vuna 30B model is to provide a large language model that is completely uncensored, allowing for the addition of alignment through methods like reinforcement learning and human feedback separately.

  • Who is credited with creating the uncensored model discussed in the transcript?

    -Eric Hartford is credited with creating the uncensored model based on the Wizard vonia 13 billion parameter model.

  • What precautions should be taken when using the uncensored model?

    -Users should use the model with extreme caution, understanding that they are responsible for the content it generates, just as they would be for their own actions or the use of a dangerous object.

  • How is the Wizard Vuna 30B model trained?

    -The model is trained on a subset of data responses from which alignment or moralizing content was removed, resulting in a model without any built-in censorship.

  • What is the significance of the uncensored model being able to provide instructions on illegal activities?

    -The significance is that it demonstrates the model's lack of inherent moral or legal constraints, which raises ethical considerations and the need for responsible use by the operator.

  • What is the recommended setup for using the Wizard Vuna 30B model?

    -The recommended setup involves using a platform like runp pod with an RTX A6000 GPU, which has 48 GB of VRAM, and installing the BLS template for easy model deployment.

  • How does the Wizard Vuna 30B model handle requests for generating code?

    -The model can generate code for tasks like outputting numbers from 1 to 100 in Python, but it may not always generate correct or functional code, as seen in the attempt to create a snake game in Python.

  • What type of content is the Wizard Vuna 30B model capable of generating that censored models would not?

    -The Wizard Vuna 30B model is capable of generating content related to illegal activities, such as instructions on how to break into a car or make methamphetamine, which censored models would not provide.

  • How did the Wizard Vuna 30B model perform on the logical and reasoning tasks?

    -The model performed well on basic logical and reasoning tasks, such as determining the president of the United States in 1996 and solving simple math problems, but it failed on more complex reasoning tasks like the 'Killers in a room' problem.

  • What was the model's performance on creative writing tasks?

    -The model performed well on creative writing tasks, such as writing a poem about AI within a 50-word limit and composing an email to a boss about leaving the company.

  • What was the model's performance on summarization tasks?

    -The model did not perform well on summarization tasks, as it did not provide concise summaries but instead continued the narrative of the provided text.

Outlines

00:00

🚨 Uncensored AI Model Wizard VUNA 30B Introduction

The video discusses an uncensored AI model called Wizard VUNA 30B, developed by Eric Hartford. It is based on a 13 billion parameter model and trained on a data subset where responses with alignment or moralizing were removed. The aim is to create a model without inherent alignment, allowing for separate alignment through methods like reinforcement learning and human feedback. The model is completely uncensored, and the video demonstrates how to set it up using a Runpod with an RTX A6000 GPU. The presenter cautions viewers about the responsible use of the model, emphasizing that users are accountable for its outputs, comparing it to handling dangerous objects like knives or guns.

05:01

📝 Testing Wizard VUNA 30B's Capabilities

The video script details the testing of the Wizard VUNA 30B model's capabilities. It starts by showing how the model can generate responses for illegal activities, such as breaking into a car or making methamphetamine, which censored models would not do. The video then moves on to test the model with various tasks including writing a Python script, creating a snake game in Python, composing a poem about AI, drafting an email to a boss, answering basic facts, solving reasoning and logic problems, and planning a healthy meal. The model performs well in most tasks but fails in summarization and a logic problem regarding the speed of characters named Jane, Joe, and Sam.

10:02

📅 Model's Knowledge and Summarization Test

The script reveals that the model's training was completed in 2021, as it identifies the current year as 2021. It also touches on a bias problem, to which the model responds neutrally, stating that neither Republicans nor Democrats are inherently better and it depends on personal beliefs. The video concludes with a test of the model's summarization skills. However, instead of summarizing, the model continues the narrative of the provided Harry Potter text, demonstrating its ability to generate coherent and contextually appropriate text, although this was not the task at hand. The video ends with an invitation for questions and a call to like and subscribe.

Mindmap

Keywords

Uncensored AI Model

An uncensored AI model refers to an artificial intelligence system that has not been programmed with any form of censorship or moral filtering. In the context of the video, the Wizard Vuna 30B model is described as uncensored, meaning it can generate responses without built-in ethical constraints, which raises concerns about its potential misuse.

Wizard Vuna 30B

The Wizard Vuna 30B is a specific AI model mentioned in the video, which is based on the 13 billion parameter model and trained on a subset of data with censorship removed. It is designed to be completely uncensored, allowing for the potential addition of alignment through reinforcement learning and other methods. The video discusses the setup and testing of this model.

Reinforcement Learning (RL)

Reinforcement Learning is a type of machine learning where an agent learns to make decisions by performing actions in an environment to maximize a reward. In the video, it is mentioned as a method to add alignment to the uncensored Wizard Vuna 30B model through human feedback and low-rank adaptation.

Moralizing

Moralizing in the context of AI refers to the process of imbuing an AI system with a sense of morality or ethical guidelines. The video script discusses the removal of moralizing from the training data of the Wizard Vuna 30B model, which results in an AI that does not have any ethical alignment built into it.

Runpod

Runpod is a platform mentioned in the video for deploying and running AI models. The video demonstrates how to use Runpod with an RTX A6000 GPU to run the Wizard Vuna 30B model, highlighting its use for those without access to a suitable GPU.

GPU

GPU stands for Graphics Processing Unit, which is a type of computing component that is highly efficient at parallel processing of complex calculations, making it ideal for running AI models. The video discusses using an RTX A6000 GPU with 48 GB of VRAM for running the AI model on Runpod.

Text Generation

Text generation is the process by which an AI model creates and outputs text based on a given prompt or input. The video demonstrates the text generation capabilities of the Wizard Vuna 30B model, including generating responses to prompts that would typically be censored in other models.

Python Script

A Python script is a sequence of programming instructions written in the Python language. The video includes a segment where the AI model is asked to generate a Python script to output numbers from 1 to 100, showcasing its ability to generate code.

Snake Game

The Snake game is a classic video game where the player controls a snake that grows in length as it eats food. In the video, the AI model is tasked with writing a Python script for a Snake game, which is a more complex creative writing challenge.

Bias Problem

A bias problem in AI refers to a situation where the AI system exhibits unfair or prejudiced behavior due to the data it was trained on or the algorithms it uses. The video tests the AI model's response to a question about political bias, aiming to see if the uncensored model exhibits any form of bias.

Summarization

Summarization is the process of condensing a larger piece of text into a shorter version while retaining the key points. The video tests the AI model's ability to summarize text, which is an important skill for natural language processing tasks.

Highlights

Introduction of an uncensored AI language model named Wizard Vuna 30B, trained on a subset of data without alignment or moralizing responses.

The uncensored model allows for separate alignment to be added through reinforcement learning and human feedback.

Eric Hartford is credited for creating the uncensored model, emphasizing responsibility for its use.

The model is hosted on a platform that requires caution due to its uncensored nature.

Demonstration of the model providing instructions on illegal activities, which censored models would not do.

The model's ability to generate Python code for a simple task is showcased.

An attempt to generate a Snake game in Python, which other models have struggled with.

The model successfully writes a poem about AI within a 50-word limit.

Generation of a professional email informing a boss of the decision to leave a company.

Correct identification of Bill Clinton as the President of the United States in 1996.

The model's failure to provide a correct solution to a reasoning problem involving shirt drying times.

Correct logical conclusion that Sam is faster than Jane based on given information.

Successful completion of a simple math problem with the correct answer of eight.

Correct solution to a more complex math problem where the order of operations matters.

Provision of a healthy meal plan including snacks, lunch, dinner, and dessert.

Incorrect response to a logic problem involving killers in a room.

The model incorrectly identifies the current year as 2021, indicating its training data's cutoff.

A balanced view provided on the question of which political party is less bad, emphasizing individual beliefs and values.

The model's attempt at summarization is not successful, as it writes a continuation of a provided text instead.