AI models can now learn 50,000 new facts without breaking themselves

Researchers found a way for large AI models to learn many new facts without forgetting what they already know. This means companies can update their AI with new information much more cheaply than before.

Keeping large AI models up-to-date with current information has been expensive. Companies either retrain the entire model, which costs millions, or they try to 'edit' specific facts. Those edits often break other parts of the model. This new method suggests a way to continuously update models with tens of thousands of changes without those problems.