With its sights set on leading the language model race, Meta has launched ‘Llama 2’, a comprehensive set of large language models designed to surpass the performance of OpenAI’s dominant GPT-4. To foster innovation and inclusivity, Meta is making Llama 2 available for free, encouraging businesses to build tailored models upon this new AI foundation. The venture is further bolstered by a strategic partnership with Microsoft, allowing effortless access to Llama 2 models via Microsoft’s Azure cloud platform.
Ranging from 7 billion to 70 billion parameters, the Llama 2 models come pre-trained and fine-tuned, demonstrating superior performance in most benchmarks tested against existing open-source chat models. With an impressive 40% more tokens than the previous iteration, Llama 2 has a significantly improved capacity to read and recall extended prompts up to 4,000 tokens.
While Meta’s prominent AI research organization invests time and resources in the evolution of language models, it breaks from the trend of restrictive access by openly sharing its models. Unlike OpenAI’s for-profit access model, Meta’s choice to provide open-source access could shift the AI landscape, offering developers and enterprises more customization and cost-saving opportunities.
Notwithstanding the guarded secrecy dominating the AI sphere, Meta takes an open-book approach. It has published a research paper on Llama 2’s development, offering parameter weights and starter codes for various versions of the models. However, detailed documentation of training and human coaching methods remains under wraps.
Feedback from the open-source community is praising Llama 2 for its superior performance and adaptability. Although some skeptics suggest Llama may not yet outshine OpenAI’s models in areas like coding, the majority viewpoints towards a positive model evolution.
An interesting subplot is Meta’s deepening alliance with Microsoft, which might have a ripple effect on OpenAI. Microsoft, a partial owner of OpenAI, provides critical server support for OpenAI’s models. Now, developers can leverage Llama 2 from Azure’s AI model catalog and build upon it using Azure’s content filtering and safety tools.
While Llama 2’s association with Azure is significant, it is not exclusive. The model will also be accessible via Amazon Web Services, Hugging Face, and others, further increasing its reach.
Last week’s announcement of the new text-to-image model, ‘CM3leon’, which can generate images from text prompts and write captions for existing images, is another testament to Meta’s ambition to compete head-on with OpenAI and Google in the AI sphere. The tech titan’s collaboration with Microsoft to integrate productivity apps into Meta’s “metaverse” experiences further highlights its commitment to AI innovation.