They train it on their own non-public data, then summarize and encrypt the product’s new configuration. The model updates are sent again to the cloud, decrypted, averaged, and built-in into the centralized product. Iteration following iteration, the collaborative teaching carries on right up until the product is fully trained. Finance: https://cruzmtyce.blogoscience.com/41343557/openai-consulting-an-overview