Meta unveils its Llama 2 AI: A direct competitor to Google and OpenAI ?🤔🤖

Cliquez ici pour lire en français

At the Microsoft Inspire conference on July 18, Meta unveiled the next generation of its major language model, Llama. Faithful to Meta’s principles, Llama 2 is open source (free access to the programming code for modification and use) and freely available for research.

Meta makes a strong case against the competition 💪🏾

With its new language model, Meta places itself in direct competition with Google and OpenAI, both of whom exploit their AI research for commercial purposes. The big news is that, for the first time, it can be used for commercial purposes, again free of charge.

Contrary to what its name suggests, OpenAI has a strictly proprietary and closed approach, while Google, which has also advocated openness until now, has decided to lock down its research in response to OpenAI’s strategy. The announcement also risks overshadowing the many start-ups that have sprung up to capitalize on the generative AI boom.

Llama 2 will be available via Amazon Web Services (AWS), Hugging Face and other providers.

By making « AI models available to everyone, they can benefit everyone. By giving companies, startups, entrepreneurs and researchers access to tools developed at a scale that would be difficult for them to build themselves, with the support of computing power they wouldn’t otherwise have access to, we open up a world of opportunities for them to experiment, innovate […] and reap economic and social benefits. ».

Llama 2 has been trained with 40% more data than Llama 1 and has doubled the length of its pop-up window. Three model sizes are available: 7, 13 or 70 billion parameters. The model is pre-trained on 2 billion billion tokens with a contextual length of 4096 tokens. Refinement for chatbot use cases was based on over a million tags.

Microsoft becomes Meta’s preferred partner for generative AI

Alongside this announcement, Meta was keen to highlight Microsoft as a « preferred partner for Llama 2 », and to jointly develop their « efforts in the field of generative AI ». This collaboration enables Llama 2 to be available in the « Azure AI model catalog ».

Developers using Microsoft Azure will be able to build with these new models and « leverage their cloud-native tools for content filtering and security functions ». Llama 2 is also optimized to run locally on Windows.

Today, with this expanded partnership, Microsoft and Meta are supporting an open approach to increasing access to fundamental AI technologies for the benefit of businesses worldwide.

For Llama 2, « responsibility » is the motto

With the availability of Llama 2 in an open source approach, Meta aims to promote « transparency and accessibility », committing itself to « building responsibly » and providing « a number of resources to help those using Llama 2 to do the same ».

To achieve this, Meta relies on 4 fundamentals:

1. Security tests: the fine-tuned models were subjected to internal and external security tests, with the creation of messages and contradictory tests, notably by third parties, to identify any gaps in the model.
2. Transparency diagram: the methods used to develop and evaluate the model, and the shortcomings identified and explained in the transparency diagram, provide an overview of the problems encountered and possible management measures.
3. Responsible use guide: to help developers use Llama 2 responsibly and securely, Meta is making available a guide presenting « best practices that reflect current research on responsible generative AI ».
4. Acceptable use policy: Meta has put in place a policy « prohibiting certain use cases to ensure that these models are used fairly and responsibly ».

Finally, Meta has also launched a university partnership program around the responsible use of these great language models. The company’s position is that openness and transparency are the best ways to guarantee the safety of these technologies, so that everyone can contribute to their improvement.

Sources: Meta, Amazon Web Services, Hugging Face