Meta’s Llama LLM makes impact on medicine, education and business

Meta has lifted the lid on how businesses, educators and researchers are putting its large language model, Llama, to work.

According to Meta, the company behind Facebook, Llama has notched up over 170 million downloads and the model is now in use in industries including education and healthcare, as well as Meta’s own products.

Llama in medicine and education

Researchers at the Yale School of Medicine and the School of Computer and Communication Sciences at Swiss science and technology institute EPFL have used Llama to bring medical know-how into low-resource environments, Meta announced this week.

The researchers used Llama to create Meditron, a suite of medical large multimodal foundation models. Meditron provides assistance with queries on medical diagnoses and management through a natural language interface. Meditron could be used in underserved areas and in emergency response scenarios, Meta said.

According to a preprint in Nature, Meditron has been trained on medical information including biomedical literature and practice guidelines. It’s also been trained to interpret medical imaging including X-ray, CT and MRI scans.

In testing, Meditron was asked to answer 244 open-ended medical questions. The responses were then reviewed by a panel of 16 physicians, 14 of whom believed that they demonstrated expertise equivalent to that of a resident doctor.

Meta also noted that Llama is being used as part of maths education in South Korea, where it’s used as the foundation for MathGPT.  

MathGPT, developed by AI startups Mathpresso and Upstage, has been integrated into Mathpresso’s QANDA product. This is an AI tutor to personalise maths teaching for students.

Related: Intel launches Gaudi 3 AI accelerator | TechFinitive

Llama in business

A tie-up between Meta and Zoom has also seen Llama 2 – along with “other third party models” – used to create Zoom’s AI Companion.

The AI assistant can be used for summarising key points and planning the next steps after a meeting, as well as providing a digest of what users have missed if they’ve had to take a break from a call. It also offers tips on how to respond to unread Zoom chat messages.

The first version of Llama was pre-trained on over a billion tokens and released in February last year, with the latest iteration of the model debuting in April with training spanning over 15 trillion tokens.

The source code for Llama is available under an open-source licence. While the first version of Llama was aimed at research work rather than business projects, subsequent versions were released under a licence that permits commercial use.

Llama 3 is currently available in two versions – 8 billion and 70 billion parameter – and Meta says the models will be made available through AWS, Databricks, Google Cloud, Hugging Face, Kaggle, IBM WatsonX, Microsoft Azure, Nvidia NIM and Snowflake.