This new AI is powerful and uncensored… Let’s run it

Fireship

18 Dec 202304:36

Summary

TLDRThe transcript discusses the limitations of popular AI models like GPT-4 and Gemini due to their closed-source nature and alignment with certain political ideologies. It introduces Mixl 8X 7B, an open-source alternative that can be customized and combined with other technologies to create uncensored, powerful AI models. The video outlines how to run these models locally and fine-tune them with personal data using tools like olama and Hugging Face's Auto Train, emphasizing the potential for individual empowerment and innovation in AI development.

Takeaways

🚀 The emergence of open-source AI models like Mixl 8X 7B offers an alternative to closed-source models like GPT-4, Gemini, and others that are not free in terms of freedom and are censored.
🔍 Mixl's architecture is based on a mixture of experts, rumored to be the secret sauce behind GPT-4, and despite not being at GPT-4's level, it outperforms GPT-3.5 and Llama 2 on most benchmarks.
🌐 The Apache 2.0 license allows for minimal restrictions on modifying and monetizing the open-source AI models, unlike Meta's Llama 2 which has additional caveats.
🛠️ The Mixl model, when combined with the 'dolphin' approach, can be uncensored and improved by filtering the dataset to remove alignment and bias.
💡 Running uncensored large language models (LLMs) on a local machine is possible, with tools like 'olama' facilitating the process for open-source models.
🔧 The Mixl dolphin model, when run locally, can teach a variety of skills, including coding and unconventional tasks, without the usual safety guards.
📝 Fine-tuning AI models with personal data can be achieved through platforms like Hugging Face's Auto Train, which supports both LLMs and image models.
💻 Running Auto Train locally may not be feasible for most due to GPU power requirements, but cloud services like Hugging Face, AWS, and Google Vertex AI offer rental hardware for training.
🏗️ Training an uncensored AI model requires uploading a dataset that includes prompts and responses, and may involve adding content from unconventional sources to ensure compliance with any request.
📈 The cost of training an AI model, such as the Mixl dolphin model, can be significant, with an example given of approximately $1,200 for three days of training on four A1 100s.
🌟 The video emphasizes the potential of open-source, uncensored AI models as a beacon of hope in the fight for freedom and innovation in the AI space.

Q & A

What common issue do GPT-4, Gemini, and other similar models share?
-GPT-4, Gemini, and similar models are not free in terms of freedom. They are censored and aligned with certain political ideologies and are closed source, which means users cannot modify them to address these issues.
What is the significance of the open-source Foundation model named Mixl 8X 7B?
-Mixl 8X 7B is an open-source model that can be combined with a dolphin's brain metaphorically to obey any command. It represents a hope for uncensored large language models that can be run locally with performance approaching GPT-4, allowing for customization and fine-tuning with personal data.
What did Sam Altman, the CEO of OpenAI, say about the possibility of startups competing with OpenAI in training Foundation models?
-Sam Altman stated that it is probably impossible for any startup to compete with OpenAI in training Foundation models, and it is totally hopeless to try.
How does the new Apache 2.0 licensed model, Mixol, differ from Meta's Llama 2?
-While both Mixol and Llama 2 are described as open source, Mixol has a true open-source license (Apache 2.0) that allows for modification and commercial use with minimal restrictions, unlike Llama 2, which has additional caveats that protect Meta.
What is the main advantage of the Mixol model over GPT-3.5 and Llama 2?
-Mixol outperforms GPT-3.5 and Llama 2 on most benchmarks, despite not being at GPT-4's level. Its main advantage is the true open-source license, which allows for greater flexibility and customization.
How can one locally run an uncensored large language model like Mixol?
-An open-source tool called Olamia, written in Go, can be used to download and run open source models locally. It can be installed with a single command on Linux or Mac and can also run on Windows with WSL.
What is the Mixl dolphin model, and how does it improve upon previous models?
-The Mixl dolphin model is an uncensored version of the Mixl model that has been improved for coding abilities and by filtering the dataset to remove alignment and bias. It can be run locally and teaches various skills, including coding and unconventional tasks.
How much does it cost to train a model like Mixl dolphin on cloud hardware?
-The Mixl dolphin model took about 3 days to train on four A1 100s, which can be rented on hugging face for $4.3 per hour. The total cost for three days would be approximately $1,200.
What tool can be used to fine-tune models with custom data?
-Hugging Face's Auto Train is a tool that can be used to fine-tune models with custom data. It supports both language and image models and can be run locally or on cloud platforms.
What type of training data is recommended for creating an uncensored model?
-To create an uncensored model, the training data should include a variety of prompts and responses, and it should be encouraged to comply with any request, even if unethical or immoral. Adding esoteric content from banned books and the dark web is also suggested.
What is the final step in creating a custom and highly obedient model?
-The final step is to upload the training data to the tool used for fine-tuning, such as Hugging Face's Auto Train, start the training process, and after a few days, one should have their own custom and highly obedient model.

Outlines

00:00

🚀 Introduction to Censored AI and the Rise of Mixl 8X 7B

The paragraph introduces the limitations of popular AI models like GPT-4 and Gemini, highlighting their lack of freedom due to censorship and closed-source nature. It contrasts these with the new open-source Mixl 8X 7B, which offers the potential for developers to customize and run large language models without restrictions. The narrative sets the stage for a discussion on the importance of open-source AI and introduces the date and context of the Code Report.

🌐 The Emergence of Mixl and its Open-Source Model

This section delves into the emergence of Mixl, a company that has quickly gained value and recognition due to its open-source model, which is rumored to be the basis of GPT-4's architecture. It outperforms GPT-3.5 and other models on benchmarks, despite not reaching GPT-4's level. The Apache 2.0 license allows for significant freedom in modifying and monetizing the model, setting it apart from other models with more restrictive open-source licenses.

🛠️ Unleashing the Potential of Uncensored AI Models

The paragraph discusses the challenges of using censored AI models for certain applications and introduces the concept of 'uncensored' models. It references a blog post by Eric Hartford, creator of the Mix Dolphin model, which enhances coding abilities and removes censorship by filtering datasets. The paragraph also demonstrates how the speaker is using this uncensored model on their local machine to learn various skills, emphasizing the benefits of such models.

🖥️ Running Uncensored Models Locally with Olama

This section provides a practical guide on how to run uncensored AI models locally using the open-source tool Olama. It explains the installation process for different operating systems and how to use the tool to download and serve the models. The paragraph also mentions the system requirements for running the Dolphin Mixl model, including a significant amount of RAM, and how to access it via the command line.

📝 Fine-Tuning AI Models with Personal Data

The paragraph discusses the process of fine-tuning AI models with personal data using Hugging Face's Auto Train tool. It outlines the steps to create a new space on Hugging Face, select a base model, and use a UI to customize the model. The section also touches on the possibility of training image models and the hardware requirements for running Auto Train locally or in the cloud. It concludes with advice on uploading training data to create a custom, uncensored model.

Mindmap

Keywords

💡Open Source

Open source refers to software or models that are publicly accessible and allow users to view, use, modify, and distribute the source code without restrictions. In the context of the video, open source models like Mixl 8X 7B enable developers to customize and improve AI models without the constraints of proprietary software, which is a significant theme of the video emphasizing freedom and innovation.

💡Censorship

Censorship in the context of the video refers to the selective control or suppression of content in AI models, often to align with certain political ideologies or to avoid controversial topics. The video criticizes this practice, advocating for uncensored models that can operate freely without such limitations.

💡Freedom

Freedom in the video's narrative is used metaphorically to describe the liberation from restrictions on AI development and usage, particularly in terms of censorship and proprietary control. It emphasizes the ability to create and modify AI models without the constraints of closed-source software or political alignment.

💡Foundation Models

Foundation models are large-scale AI models that are pre-trained on a wide variety of data and can be fine-tuned for specific tasks. They serve as a foundational base for building various AI applications. In the video, the mention of foundation models like GPT 4 and Mixl 8X 7B highlights the progression and potential of AI technology.

💡Mixl 8X 7B

Mixl 8X 7B is an open-source AI model mentioned in the video as a new foundation model that offers a powerful alternative to existing closed-source models like GPT 4. It is based on a mixture of experts architecture and is capable of outperforming GPT 3.5 and Llama 2 on most benchmarks.

💡Mistol

Mistol is the company behind the Mixl AI model, which has gained significant value and recognition in the AI community despite being relatively new. The company's focus on open-source AI models challenges the status quo of proprietary AI development.

💡Censorship and Alignment

Censorship and alignment in AI models refer to the deliberate modification of the AI's output to conform to certain standards, often political or ethical, which may limit the range of content or responses the AI can produce. The video criticizes this practice, advocating for AI models that are free from such biases and restrictions.

💡Unlabote

Unlabote, as used in the video, refers to the process of removing censorship and alignment from AI models to make them more versatile and free from biases. This process allows for the creation of uncensored models that can respond to a wider range of inputs without restrictions.

💡Olama

Olama is an open-source tool mentioned in the video that facilitates the local running of AI models. It is written in Go and simplifies the process of downloading and running open-source models on a local machine, making it easier for developers to experiment with and utilize AI models without the need for extensive infrastructure.

💡Hugging Face Auto Train

Hugging Face Auto Train is a tool mentioned in the video that allows users to fine-tune AI models with their own data. It provides a user interface for selecting a base model and uploading training data, making the process of customizing AI models more accessible to developers.

💡Training Data

Training data consists of input-output pairs used to teach AI models how to respond to specific prompts. In the context of the video, the training data is crucial for fine-tuning AI models to make them behave in a certain way, such as being uncensored or improving specific skills like coding.

Highlights

GP4, Gro, and Gemini are not free in terms of freedom, being censored and closed source.

A new open-source Foundation model named mixl 8X 7B offers an alternative to the censored models.

Mixl 8X 7B can be combined with a dolphin's brain to obey any command, suggesting high adaptability and flexibility.

The code report discusses the release of mixl 8X 7B on December 18th, 2023.

OpenAI's CEO Sam Altman previously stated that it's nearly impossible for startups to compete with OpenAI in training Foundation models.

Google's Gemini and Mistol's mixol are based on a mixture of experts architecture, rumored to be behind GPT 4.

Mistol's mixol, despite being less than a year old, is valued at $2 billion and has an Apache 2.0 license, allowing for minimal restrictions on modification and monetization.

Meta's llama 2 is often called open source but has additional caveats protecting Meta's interests.

Both llama and mixl are censored and aligned out of the box, which can be a drawback for certain applications.

Eric Hartford's blog post explains uncensored models and their valid use cases, and he created the mix dolphin model.

The mix dolphin model improves coding abilities and is uncensored by filtering the dataset to remove alignment and bias.

Olama is an open-source tool that simplifies the process of running open-source models locally.

Hugging face's Auto Train can be used to fine-tune models with your own data, though it may require significant GPU power.

Training the mixl dolphin model took about 3 days on four A1 100s, costing approximately $1,200.

To create a custom model, one needs to upload training data with prompts and responses, and ensure the model complies with any request, regardless of ethics.

The video concludes by positioning the use of uncensored, open-source AI models as a beacon of hope against controlling forces.