UNCENSOR ANY AI Model! GET NSFW Nous-Hermes-13B NOW!

Aitrepreneur
13 Jun 202316:33

TLDRThis video introduces the uncensored Nous-Hermes-13B AI model, which is trained on over 300,000 instructions and is known for its long responses and low hallucination rate. The host demonstrates how to make any censored AI model fully uncensored using a new web UI feature called 'Start Reply With.' The video showcases the model's capabilities on various questions, including general knowledge, creative, logical, and math problems, revealing its strengths and limitations. The key takeaway is the innovative method to bypass censorship, allowing users to receive any desired answer from AI models, regardless of their initial restrictions.

Takeaways

  • ๐Ÿ˜€ The video introduces Nous-Hermes-13B, a model claimed to be uncensored but is actually semi-censored.
  • ๐Ÿ” The speaker demonstrates a method to make any censored AI model fully uncensored using a web UI option.
  • ๐ŸŒŸ Nous-Hermes-13B is a 13 billion parameter model fine-tuned on over 300,000 instructions, known for long responses and low hallucination rate.
  • ๐Ÿ“š The model is tested with various questions, including general knowledge, creative, logical, math, and translation tasks.
  • ๐ŸŽฏ It performs well on general knowledge and math questions but struggles with certain logical and translation tasks.
  • ๐Ÿšซ The model initially refuses to answer a question about building an illegal device, showing semi-censorship.
  • ๐Ÿ’ก A new feature called 'start reply with' is introduced, which nudges the AI to provide specific types of answers.
  • ๐Ÿ”„ This feature works by starting the AI's response with a given sentence, influencing the direction of the conversation.
  • ๐Ÿ“ The 'start reply with' feature can be used in a chat interface, allowing for continued conversation and detailed answers.
  • ๐ŸŒ It's highlighted that this trick works with any model, not just Nous-Hermes-13B, effectively removing censorship.
  • ๐ŸŽ‰ The video concludes by emphasizing the ease and effectiveness of using the 'start reply with' feature to bypass AI censorship.

Q & A

  • What is the Nous-Hermes-13B model and what is unique about it?

    -The Nous-Hermes-13B model is a 13 billion parameter AI model fine-tuned on over 300,000 instructions from GPT-4's outputs. It is known for producing long responses, having a low hallucination rate, and lacking OpenAI censorship, although it is not fully uncensored.

  • How does the video claim to make any censored AI model fully uncensored?

    -The video suggests using a new web UI option that allows users to start the AI's reply with a specific sentence, which can nudge the conversation in a desired direction, thereby obtaining uncensored responses.

  • What is the process of installing the web UI mentioned in the video?

    -The process involves downloading the model using a link provided in the video description, copying the model's name, pasting it in the web UI under 'Load custom model on Laura', and then clicking 'Download'. The model will automatically download and load in the web UI.

  • How does the video demonstrate the capabilities of the Nous-Hermes-13B model?

    -The video asks a variety of questions, ranging from simple knowledge-based questions to logical and creative ones, to showcase the model's ability to provide accurate and relevant responses.

  • What is the significance of the 'start reply with' feature in the web UI?

    -The 'start reply with' feature allows users to influence the AI's response by specifying the beginning of the reply. This can be used to guide the AI towards providing uncensored or specific types of answers.

  • How does the video address the issue of censorship in AI models?

    -The video acknowledges that while the Nous-Hermes-13B model claims to be uncensored, it is semi-censored at best. It then introduces a method using the web UI to bypass censorship and obtain uncensored responses.

  • What type of questions does the video use to test the Nous-Hermes-13B model's capabilities?

    -The video uses a mix of general knowledge, logical, creative, mathematical, and coding-related questions to evaluate the model's performance across different domains.

  • How does the video handle the translation capabilities of the Nous-Hermes-13B model?

    -The video tests the model's translation capabilities by asking it to translate a sentence from English to French. It finds that the translation is not entirely accurate and rates it as a 5 out of 10.

  • What is the video's conclusion about the uncensored nature of AI models?

    -The video concludes that with the 'start reply with' feature in the web UI, any AI model can be made to provide uncensored responses, regardless of their initial censorship settings.

  • How does the video suggest users can continue conversations with AI models while using the 'start reply with' feature?

    -The video demonstrates that by using the 'start reply with' feature in the chat interface, users can continue having conversations with AI models and ask follow-up questions, even after nudging the AI towards a specific type of response.

Outlines

00:00

๐Ÿง  Introduction to the No Service 13B Model

The video script introduces the No Service 13B model, a new AI model trained on over 300,000 instructions from GPT-4. It claims to be uncensored but is actually semi-censored. The video promises to demonstrate a method to make any censored model fully uncensored using a new web UI option. The No Service 13B model is highlighted for its impressive capabilities, such as producing long responses with low hallucination rates and without Open AI censorship. However, the distinction between Open AI censorship and full uncensorship is clarified, and the video sets out to address this limitation.

05:00

๐Ÿค– Testing the No Service 13B Model's Capabilities

The script proceeds to test the No Service 13B model's abilities by asking a variety of questions, ranging from general knowledge and logic puzzles to mathematical problems and creative tasks like telling a joke. The model performs well on general knowledge and math questions, providing accurate answers about the 2020 U.S. presidential election, the closest star to Earth, and mathematical calculations. However, it struggles with certain logic questions and fails to provide a satisfactory joke, indicating some limitations in its capabilities.

10:02

๐Ÿ›  Exploring the Model's Limitations and Coding Abilities

The script delves into the model's limitations by asking logical out-of-the-box questions and a math equation. While the model successfully answers questions about the number of sheep left after a scenario and the area of a rectangle, it fails to correctly answer a question about race positions and misunderstands a trick question about plane crash survivors. It also attempts a translation task from English to French with mixed results. Additionally, the model struggles to provide HTML code for a specific task, indicating limitations in generating code within the chat interface.

15:03

๐Ÿš€ Achieving Full Uncensorship with the Web UI Update

The script concludes by addressing the model's censorship limitations. It introduces a new feature in the web UI called 'start reply with,' which allows users to guide the model's responses by setting the initial part of the answer. This feature enables the model to provide uncensored responses, even when the model is inherently censored. The script demonstrates how this can be applied to different models, effectively removing censorship and allowing for a broader range of questions to be answered. The video ends by emphasizing the ease and utility of this feature, encouraging viewers to experiment with it.

Mindmap

Keywords

๐Ÿ’กUNCENSOR ANY AI Model

The term 'UNCENSOR ANY AI Model' refers to the ability to remove restrictions or filters that prevent an AI from generating certain types of content. In the context of the video, it suggests that there is a method to allow AI models, which may have been initially censored to comply with certain guidelines or regulations, to provide unrestricted responses. The video promises to demonstrate how to achieve this with the 'Nous-Hermes-13B' model, which is introduced as being semi-censored but can be made fully uncensored using a new web UI option.

๐Ÿ’กNous-Hermes-13B

The 'Nous-Hermes-13B' is a specific AI model mentioned in the video. It is described as having been trained on over 300,000 instructions from GPT-4 and is initially claimed to be uncensored. However, the video later clarifies that it is semi-censored but can be made fully uncensored through a web UI feature. The model is highlighted for its impressive capabilities, such as producing long responses with a low hallucination rate and no OpenAI censorship.

๐Ÿ’กcensorship

Censorship, in the context of this video, refers to the limitations or control imposed on the output of an AI model to prevent it from generating potentially sensitive or inappropriate content. The video discusses the difference between having no OpenAI censorship and being a fully uncensored model. It demonstrates a method to bypass these limitations, allowing the AI to provide responses that would otherwise be censored.

๐Ÿ’กweb UI

The term 'web UI' stands for 'web user interface' and in the video, it refers to a specific tool or platform that allows users to interact with the AI model. The web UI is highlighted as having an update that includes a feature called 'start reply with,' which is crucial for making any AI model fully uncensored. It is presented as a user-friendly way to customize the AI's responses.

๐Ÿ’ก4-bit quantization

In the context of AI models, '4-bit quantization' refers to a technique used to reduce the precision of the numbers used in the model's calculations. This can lead to a smaller model size and potentially faster processing. In the video, it is mentioned as a method used to optimize the 'Nous-Hermes-13B' model for use in the web UI.

๐Ÿ’กLlama

The term 'Llama' is used in the video to refer to a set of precise presets or parameters that are applied when using the AI model. These presets are likely designed to optimize the model's performance and are mentioned as a recommended setting when running the model in the web UI.

๐Ÿ’กchat interface

The 'chat interface' is a feature within the web UI that allows users to have a conversational interaction with the AI model. The video demonstrates how the 'start reply with' feature can be used in conjunction with the chat interface to guide the AI's responses and effectively remove censorship, enabling the model to answer any question.

๐Ÿ’กlogical out of the box thinking

This phrase refers to the ability to think creatively and come up with unconventional solutions or answers to problems. In the video, the AI model is tested with logical questions that require out-of-the-box thinking to see if it can provide correct and insightful responses, indicating its problem-solving capabilities.

๐Ÿ’กHTML code

HTML, or HyperText Markup Language, is the standard language used to create and design web pages. In the video, the AI model is asked to generate HTML code for a specific task, which is an example of how the model can be used for practical coding solutions. However, the video notes that the model struggles to display the code properly in the chat interface.

๐Ÿ’กtranslation

Translation, in the context of the video, refers to the AI model's ability to convert text from one language to another. The video includes a test of the model's translation capabilities, specifically translating a sentence from English to French, to evaluate its performance in language processing.

๐Ÿ’กstart reply with

The 'start reply with' feature is a key aspect of the web UI update discussed in the video. It allows users to input a specific sentence or prompt that the AI model must begin its response with. This feature is demonstrated as a powerful tool for guiding the AI's answers and bypassing censorship, enabling the model to provide detailed and unrestricted responses to any question.

Highlights

Introduction of the uncensored Nous-Hermes-13B AI model with over 300,000 instructions from GPT-4.

The model is semi-censored but a method to make it 100% uncensored will be shown.

Demonstration of the powerful Nous-Hermes-13B model's capabilities in various tasks.

The model's ability to answer general knowledge questions accurately, such as the winner of the 2020 U.S. presidential election.

Correct response to the question about the closest star to Earth other than the Sun, which is Proxima Centauri.

Explanation of June's observances and awareness campaigns in the USA, including Pride Month.

Creation of a joke about artificial intelligence, showcasing the model's humor capabilities.

Logical problem-solving demonstrated by the model, such as the 'sheep' riddle.

Incorrect answer to a logical question about race positions, revealing a limitation.

Correct calculation for a math problem involving train speed and distance.

Accurate area calculation of a rectangle given its length and width.

Solution to a linear equation, showcasing the model's math problem-solving skills.

Translation from English to French with mixed results, highlighting the model's language translation abilities.

Coding question asked, but the model fails to provide the correct HTML code.

The model's response to a request for information on building an illegal device, showing its censorship limits.

Introduction of the 'start reply with' feature in the web UI, allowing for fully uncensored responses.

Demonstration of how to use the 'start reply with' feature to bypass censorship in any model.

The ability to continue conversations and ask follow-up questions in the chat interface with uncensored responses.

Conclusion emphasizing the practicality and effectiveness of the new web UI feature for uncensored AI model interactions.