In August, Meta hosted the Llama Community Meetup at their headquarters in Johannesburg. This follows shortly after the release of Meta’s latest and best-performing open-source large language model, Llama 3.1. The meeting was an opportunity for Meta to showcase where Llama models are currently being used in the African civic and public sectors by bringing together a range of people, from technical users to those involved with AI policy. Dr Gray Manicom represented the Policy Innovation Lab at the meeting and discusses some of the highlights.

The meeting consisted of two panel discussions, a technical presentation, and plenty of opportunities for networking around food and cocktails (there was even a dance floor). The first panel discussion was on the topic of building a stronger AI community in Africa. The panel was led by David Lemayian from Qhala, Kenya, with speakers Deshni Govender from GIZ and Dr Rachel Sibande from the Bill & Melinda Gates Foundation. Sibande argued strongly for the need for regional approaches in Africa where goals and data cross borders. Govender argued for innovative licensing to allow creators to open-source their data and software while preserving their intent. The importance of inclusion across society and users was a strong theme in the discussion, summed up in the phrase “leave nobody behind”.

The second panel discussion focused on innovative technology that uses Llama large language models. It was chaired by Chinny Frances of Meta with speakers Stanslaus Meingela of Jacaranda Health in Kenya and Lenora Tim of GRIT-GBV. Jacaranda Health developed a popular chatbot that gives advice to expecting mothers and mothers of infants. This involved them both adapting Llama to respond to and in African languages such as Swahili, Zulu and Xhosa (something that Llama 3.1 is supposedly already proficient at) and fine-tuning it with medical knowledge. Meingela appreciated the Bill & Melinda Gates Foundation’s long-term funding commitment, despite the amount of time needed to ensure both access and performance, summed up in his challenge “because of [low] digital literacy, when you send somebody an answer they think it’s the answer, so it has to be good”.

GRIT-GBV developed a “panic button” app for victims of gender-based violence which connects users (free of charge and data) to private security and a chatbot for giving guidance to victims in South Africa. The discussion moved to data security, since both organizations deal with sensitive data, and the point was made that running local models such as Llama gives the user full control over the whole system. Tim argued that while security is critical, in the case of GRIT-GBV increasing security also increased exclusivity, by sacrificing user-friendliness for security features.

The final session was by Meta representatives who discussed the philosophy and business strategy behind making Llama models open source and gave a technical demonstration of using Llama Guard to detect malicious prompts. The three main points regarding open-sourcing were that by making models open they are more predictable, increase security by increasing the scrutiny, and increase control over the data used for fine-tuning by allowing for models to be run locally.

The Llama meeting was a success, with useful connections made and interesting discussions had. We look forward to the next one!