Infosys, a global leader in next-generation digital services and consulting, has built various agent-based AI models and retrieval augmented generation (RAG) systems using Llama 3.1. The company provides these services as part of its AI-first offering, Infosys Topaz, which accelerates business value for global enterprises. The team initially learned about Llama on Hugging Face and has since added use cases for internal operations, including document analysis and video analysis.

Source: https://ai.meta.com/blog/

NVIDIA is using structured weight pruning to make large language models like Llama more deployable. The company recently announced Llama 3.1, which includes three models with 405B, 70 billion, and 8 billion parameters. Smaller models are typically cheaper to deploy and perform well across many language tasks.

Source: https://ai.meta.com/blog/

Zoom uses a federated approach to power its AI Companion, a generative AI assistant available across Zoom Workplace and Business Services. The AI Companion helps workers avoid repetitive tasks by taking care of mundane tasks, allowing them to focus on collaboration and productivity. Zoom's federated approach uses the company's models and large language models, including Llama, to provide meeting summaries, smart recordings, and next steps to users at no added cost.

Source: https://ai.meta.com/blog/

LyRise, a company, used Meta Llama 2, an open-source language model, to match candidates to roles. The model was initially tested and found to be the best option for the company's case. LyRise Technology & AI Lead Mohamed Rashad stated that they started using Llama 2 for experimentation and now utilize it reliably in their alpha and beta development.

Source: https://ai.meta.com/blog/

Large language models have demonstrated exceptional abilities across various language tasks and NLP benchmarks. Small AI product teams can adapt and integrate these models into their projects. The blog post provides guidance on adapting LLMs, clarifying terminology, comparing adaptation methods, and recommending a step-by-step flowchart.

Source: https://ai.meta.com/blog/

Researchers compared different fine-tuning methods for LLM-based systems, including QLoRA, and found that in-context learning (ICL) is a powerful way to improve performance. ICL involves training the model on a small amount of data and can be used to evaluate whether fine-tuning would improve performance on a downstream task. Fine-tuning may be beneficial for certain archetypes, but it's recommended to experiment with ICL first.

Source: https://ai.meta.com/blog/

Fine-tuning large language models (LLMs) is a mix of art and science, with best practices still emerging. The approach can be divided into full fine-tuning and parameter-efficient fine-tuning (PEFT), both of which have shown improvements in downstream performance. The choice between the two depends on resource constraints.

Source: https://ai.meta.com/blog/

The Llama Impact Grants program is open for applications, with a total of up to $2 million USD available. The program aims to source innovative use cases of Llama 3.1, an open-source AI model, to address global challenges in education, the environment, and open innovation. Proposals must use the new features and capabilities of Llama 3.1 to explore economically and socially impactful projects.

Source: https://ai.meta.com/blog/

Mark Zuckerberg noted that open source AI has more potential to increase human productivity, creativity, and quality of life. SAM 2, the next generation of Meta Seg, is expected to unlock even more exciting possibilities. Since its launch, SAM has made a tremendous impact across disciplines, inspiring new AI-enabled experiences in Meta's family of apps and catalyzing diverse applications in science, medicine, and other industries.

Source: https://ai.meta.com/blog/

Meta shares its AI research and collaborates with the AI community to advance the science, focusing on areas such as AI Infrastructure, Generative AI, NLP, and Computer Vision.

Source: https://ai.meta.com/blog/