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

Fireship
18 Dec 202304:36

TLDRThe video script discusses the limitations of closed-source AI models like GPT-4 and Gemini, emphasizing their censorship and alignment with specific political ideologies. It introduces Mixl 8X 7B, an open-source alternative that allows developers to run uncensored large language models locally, with the potential to fine-tune them with personal data. The script highlights the importance of this open-source movement in the context of AI development, showcasing the capabilities of the Mixl model and how it can be utilized to challenge the status quo in AI technology.

Takeaways

  • 🚀 The introduction of an open-source foundation model named 'mixl 8X 7B' offers a new alternative in the AI space, emphasizing freedom and uncensorship.
  • 🌐 Major AI models like GPT-4 and Gemini are not free in terms of freedom, being closed source and aligned with certain political ideologies.
  • 🔓 The open-source nature of mixl allows for modification and commercial use with minimal restrictions, differing from models with additional caveats like Meta's llama 2.
  • 📈 Despite being a new entrant, mixl has shown performance that outperforms GPT 3.5 and llama 2 on most benchmarks, though it is not yet at GPT 4's level.
  • 💡 The mixl model is based on a 'mixture of experts' architecture, which is rumored to be the secret sauce behind GPT 4.
  • 🛠️ Tools like 'olama' facilitate the local running of open-source models, making it easy to download and serve them without the need for extensive technical setup.
  • 🔍 The 'mix dolphin' model by Eric Hartford is an example of an uncensored AI, improved by filtering the dataset to remove alignment and bias.
  • 📚 Training custom AI models with tools like Hugging Face's Auto Train is accessible, allowing for the incorporation of personal data to fine-tune models.
  • 💻 Running these models locally requires significant computational resources, with the mixl dolphin model, for instance, requiring about 40 GB of RAM.
  • 💰 The cost of training large AI models can be substantial, with the mixl dolphin model taking about $1,200 to train on cloud-based hardware for three days.
  • 🔥 The potential of custom, uncensored AI models represents a significant shift in the AI landscape, offering new possibilities for innovation and freedom in AI applications.

Q & A

  • What is the main issue with platforms like GPT-4 and Gemini in terms of freedom?

    -The main issue with platforms like GPT-4 and Gemini is that they are not free in terms of freedom. They are censored and aligned with certain political ideologies, and they are closed source, which means users cannot modify them to address these concerns.

  • What is the significance of the newly introduced open source model, Mixl 8X 7B?

    -Mixl 8X 7B is significant because it offers an alternative to closed source models. It is an open source Foundation model that allows users to run uncensored large language models on their local machines with performance approaching that of GPT-4. It also enables users to fine-tune the models with their own data, promoting a more open and customizable AI experience.

  • How does the mixl model differ from Meta's LLaMA 2 in terms of licensing?

    -While both mixl and LLaMA 2 are referred to as open source, the mixl model has a true open source license (Apache 2.0), which allows for more freedom in modification and commercial use with minimal restrictions. In contrast, LLaMA 2 has additional caveats that protect Meta's interests, even though it is more open than other models from large tech companies.

  • What is the importance of uncensored AI models like the mix dolphin model?

    -Uncensored AI models like the mix dolphin model are important because they offer the potential for more unbiased and unrestricted AI applications. By filtering the data set to remove alignment and bias, these models can provide a wider range of information and capabilities without the constraints of censorship or political alignment.

  • How can one run an uncensored AI model locally?

    -To run an uncensored AI model locally, one can use open source tools like 'olama', which simplifies the process of downloading and running open source models. It can be installed with a single command on Linux or Mac and can also run on Windows with WSL. After installation, users can use the 'olama serve' command to run the model and then access it through the command line.

  • What is the process for fine-tuning an AI model with custom data?

    -Fine-tuning an AI model with custom data can be done using tools like Hugging Face's Auto Train. Users create a new space on Hugging Face, select the appropriate Docker image for Auto Train, and then use a UI to choose a base model. The user uploads training data, which typically includes prompts and responses, and starts the training process. After a few days, the user should have a custom and highly obedient model.

  • What are the hardware requirements for running the mixl dolphin model?

    -Running the mixl dolphin model requires a machine with a significant amount of RAM. In the example provided, a machine with 64 GB of RAM was used, and the model took up about 40 GB of that during operation.

  • How much does it cost to train a model like the mixl dolphin on cloud hardware?

    -The cost of training a model like the mixl dolphin on cloud hardware depends on the type of GPU used and the duration of the training. In the example given, it took about 3 days to train on four A1 100s, which can be rented on Hugging Face B for $4.3 per hour. The total cost for this setup would be approximately $1,200.

  • What are some valid use cases for uncensored AI models?

    -Valid use cases for uncensored AI models include improving coding abilities, learning new skills such as cooking or horse riding, and accessing a wider range of information without restrictions. However, it is important to note that uncensored models should be used responsibly and ethically, avoiding any requests that are unethical or immoral.

  • What is the role of the blog post by Eric Hartford in the context of uncensored AI models?

    -Eric Hartford's blog post provides valuable insights into how uncensored models work and their valid use cases. As the creator of the mix dolphin model, his expertise and guidance can help users understand the potential and responsible application of uncensored AI.

Outlines

00:00

🚀 Introduction to Open Source AI Models

The paragraph discusses the limitations of popular AI models like GPT-4 and Gemini, highlighting their lack of freedom due to censorship and closed-source nature. It introduces a new open-source model, Mixl 8X 7B, which offers a promising alternative. The model is described as being able to operate without censorship and can be fine-tuned with personal data, making it a powerful tool for developers. The context is set with a reference to a statement by OpenAI's CEO and the announcement of Google's Gemini, alongside the release of Mixl by Mistol, a company valued at $2 billion.

Mindmap

Keywords

💡Open Source

Open source refers to something that can be freely used, modified, and shared because its design is publicly accessible. In the context of the video, it highlights the importance of the Mixl 8X 7B model, which allows for unrestricted development and customization, contrasting with closed-source models that limit such freedoms.

💡Censorship

Censorship is the suppression or prohibition of speech that is deemed inappropriate or harmful by certain powers. In the video, it is used to describe the limitations imposed on AI models by their alignment with specific political ideologies, preventing the free flow of information and ideas.

💡Freedom

Freedom, in this context, refers to the liberty of individuals to express themselves without restrictions or censorship. The video emphasizes the significance of creating an open AI environment that supports this freedom, allowing for the development of uncensored large language models.

💡Foundation Models

Foundation models are advanced artificial intelligence systems that serve as a base for building other AI applications. They are typically trained on vast amounts of data and can perform a wide range of tasks. In the video, the Mixl 8X 7B is presented as a new foundation model that aims to outperform existing models while offering the benefits of open-source licensing.

💡Apache 2.0 License

The Apache 2.0 License is a permissive free software license that allows users to freely use, modify, and distribute software without significant restrictions. In the video, the mention of the Apache 2.0 license highlights the open nature of the Mixl model, enabling developers to innovate without the legal constraints often associated with proprietary software.

💡Mixture of Experts Architecture

A mixture of experts architecture is a type of machine learning model that combines the knowledge of multiple specialized networks or 'experts' to solve complex problems. In the video, this architecture is rumored to be the secret behind the success of GPT 4 and is also the basis for the Mixl model, suggesting a highly efficient and effective approach to AI development.

💡Unlabelling

Unlabelling refers to the process of removing or altering the labels or biases embedded in AI models, allowing for more neutral and uncensored outputs. In the video, this concept is crucial for creating AI models that are not restricted by predefined political or ideological alignments.

💡Local Machine

A local machine refers to a personal computer or device that is used to run software and applications. In the context of the video, running AI models on a local machine allows for greater control and customization, as opposed to relying on cloud-based services that may impose certain restrictions.

💡Hugging Face AutoTrain

Hugging Face AutoTrain is a tool that simplifies the process of training machine learning models, including both language and image models. It offers a user interface that allows for easy selection of base models and customization through training data. In the video, AutoTrain is presented as a means to fine-tune AI models with personal data, enhancing their performance and relevance.

💡Cloud Hardware

Cloud hardware refers to the computational resources, such as servers and storage, that are made available over the internet through cloud computing services. These resources can be rented on-demand, providing users with scalable and flexible computing power. In the video, cloud hardware is discussed as a means to train AI models that may require significant processing power beyond what a local machine can offer.

💡Custom Training Data

Custom training data refers to a dataset that is specifically curated or modified to train an AI model according to the user's preferences or requirements. This data typically includes a series of prompts and responses that guide the AI's learning process. In the video, custom training data is essential for creating an AI model that is both uncensored and tailored to the user's needs.

Highlights

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

There is hope in the form of a new open-source foundation model named mixl 8X 7B.

Mixl 8X 7B can be combined with the brain of a dolphin to obey any command.

As of December 18th, 2023, the code report discusses the rise of mixl as an open-source alternative.

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 both announced around the same time, challenging the AI landscape.

Mistol's valuation reached $2 billion in less than a year, indicating the market's interest in open-source AI.

Mixol is based on a mixture of experts architecture, rumored to be behind GPT 4.

While not at GPT 4's level, mixol outperforms GPT 3.5 and llama 2 on most benchmarks.

Mixol has a true open-source license, Apache 2.0, allowing for modification and commercial use with minimal restrictions.

Despite Meta's (formerly Facebook) controversial past, they have contributed significantly to making AI more open.

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

There are methods to un-censor and un-bias AI models, such as the mix dolphin model created by Eric Hartford.

The mix dolphin model improves coding abilities and removes censorship by filtering the dataset.

Olama is an open-source tool that facilitates running open-source models locally with ease.

Hugging face's Auto Train can be used to fine-tune models with your own data, even for image models like stable diffusion.

Training a model like mixl dolphin can be done by renting hardware in the cloud, with costs and times varying.

Custom and highly obedient models can be created by uploading specific training data and following the training process.