On 18 April 2024, Meta released Llama 3, a family of language models whose "weights", the numbers that make up the trained model, could be downloaded and used by developers, subject to a licence. Some people call such releases open-weight. They are not always "open source" in the strict sense, because training data and some restrictions are not shared.
The case for open models
- Access. Students, start-ups and researchers who cannot pay for large cloud services can run and study the models.
- Transparency. More people can inspect how the models behave.
- Local solutions. Developers can adapt a model to local languages and needs, which matters for African languages that are poorly served by big products.
The case for caution
- Misuse. Once released, a model cannot easily be recalled, and safeguards can be removed.
- Responsibility. It is harder to hold anyone accountable when a model is modified and spread widely.
Where we stand
There is no simple answer, and thoughtful experts disagree. For learners, the useful point is that AI is not only a product from a few companies. Understanding how models are built, trained and shared opens a path to becoming creators of technology, not only users.