These last few weeks have seen tech buzz centered on language models created and deployed by companies like Google, Microsoft, and OpenAI. Meta, Facebook’s parent company, continues to make significant progress in this area, and today, it has released a new AI language model called LLaMA.
LLaMA doesn’t work like ChatGPT and Bing. It’s not a platform that everyone can use. Meta describes it as a research tool for the AI community.
Meta will release LLaMA (which isn’t a single system but a set of four different-sized models) under a “noncommercial license focused only on research use cases,” which grants access to universities, NGOs, and industry labs.
The company believes the AI community, including academic researchers, policymakers, civil society, and industry, must collaborate to create clear guidelines for responsible AI, particularly large-language models. We look forward to seeing the community learn and build using LLaMA.
Meta claims the LLaMA-13B model is smaller than OpenAI’s GPT-3 model on most benchmarks. In contrast, the LLaMA-65B model is more competitive with other models, such as DeepMind’s Chinchilla70B or Google’s PaLM 540B. The numbers in these names are a measure of system size and an approximation of its sophistication.
LLaMA-13B, once trained, can run on a single Nvidia TeslaV100 GPU. While this is a welcome development for smaller institutions who wish to conduct tests on these systems, it doesn’t make much sense for researchers working alone.
Meta’s release is significant, partly due to its detachment from the AI chatbot hype. This could be a blessing in disguise, considering the backlash Microsoft received for the hurried Bing launch and the plummeting of Google’s stock value after their bard chatbot blunder in a demonstration.
Meta has released its AI chatbots but has not received a great reception. BlenderBot was criticized for being…not very good. Galactica was designed to write scientific papers and was taken offline three days after it continued producing scientific nonsense. Meta is presumably hoping for a kinder welcome with the LLaMA quartet.
Mark Zuckerberg, CEO of Facebook, said he was releasing a new AI large-language model called LLaMA today to help researchers. LLMs have demonstrated a lot of promise in generating text, conversing, summarizing written materials, and more complex tasks such as solving math theorems and predicting protein structures. Meta is committed to open research models and will make their model available to the AI research community.