News Summary

The Allen Institute for AI (Ai2) has introduced a groundbreaking large language model named FlexOlmo, which allows data owners to maintain control over their data even after it has been used to train AI models. This innovative model challenges the traditional AI industry practices where data, once utilized, becomes inseparable from the model. FlexOlmo allows contributors to train their data independently and merge it with a publicly shared anchor model, ensuring that the data itself is never fully relinquished. This model architecture, known as a ‘mixture of experts,’ enables the separation and control over specific data contributions, allowing for removal if necessary. The FlexOlmo model has shown superior performance in tests, outperforming individual models and achieving better results on common benchmarks. This new approach signifies a shift in how AI models can be trained while respecting data ownership and control.