Have you ever dreamed of having a powerful AI assistant that can write essays, poems, code, lyrics, and more for you? Or maybe you are a researcher who wants to explore the cutting-edge capabilities of natural language processing (NLP) systems? If so, you might be interested in Meta's LLaMA (Large Language Model Meta AI), a state-of-the-art foundational large language model recently released to the public.
LLaMA is a neural network that can generate text based on a given input. It works by taking a sequence of words as an input and predicting a next word to recursively generate text. For example, if you give it the prompt "The meaning of life is", it might complete it with "to find your purpose and passion".
LLaMA is not the first large language model to do this. You might have heard of OpenAI's GPT-3 or Microsoft's Bing, which are also capable of generating impressive texts on various topics. However, LLaMA has some advantages over these models:
LLaMA is smaller and faster. While GPT-3 has 175 billion parameters (which are numbers that represent what the model knows), LLaMA has only 65 billion parameters at its largest size. This means that LLaMA requires less computing power and resources to train and run.
LLaMA is more multilingual. While GPT-3 and Bing are mostly trained on English texts, LLaMA is trained on texts from 20 languages with Latin and Cyrillic alphabets. This makes it more versatile and inclusive for different users and use cases.
LLaMA is more accessible. While GPT-3 and Bing are only available through paid APIs or limited demos, LLaMA is open source and free for non-commercial use by researchers and entities affiliated with government, civil society, and academia. This means that anyone who meets these criteria can download and use LLaMA for their own projects.
How did LLaMA leak online?
Despite Meta's intention to make LLaMA available only to qualified researchers under certain conditions, someone managed to leak the model weights (which are the trained knowledge stored in the neural network) on BitTorrent on March 2. Since then, there has been an explosion of development surrounding LLaMA.
For example, a software developer created a tool called "llama.cpp" that can run LLaMA locally on a Mac laptop. Soon thereafter, people worked out how to run LLaMa on Windows as well. Then someone showed it running on a Pixel 6 phone, and next came a Raspberry Pi (albeit running very slowly).
This means that anyone who can access BitTorrent can potentially download and run LLaMa on their own devices without paying any fees or following any rules set by Meta.
How can I apply for access to the official version of LlaMa?
If you are interested in using the official version of LLAMA for your research or non-commercial purposes, you need to fill out an application form here: https://ai.facebook.com/research/llama-application/
You will need to provide some information about yourself or your organization, such as:
- Your name
- Your email address
- Your affiliation
- Your country
- Your intended use case
- Your expected data sources
- Your expected model size
You will also need to agree to some terms and conditions, such as:
- You will not use LLAMA for commercial purposes
- You will not share LLAMA with others without permission
- You will follow responsible AI practices
- You will acknowledge Meta as the source of LLAMA
Meta will review your application and notify you if you are approved or rejected.
How can I access the leaked version of LLAMA?
If you want to access the leaked version of LLAMA, you need to have some technical skills. The models are available for download on BitTorrent. You will need some solid compute hardware to run the larger models which most people don't have.