The content is a bootstrapping tool. Once the language model gets critical mass it gets further training data from its interactions with users, which are priva

The content is a bootstrapping tool. Once the language model gets critical mass it gets further training data from its interactions with users, which are private to the language model's developer. It's like how Google Search bootstrapped off the content of the web, but the real reason you can't replicate Search today is all the information about who is searching for what, which links they actually click on, and how long they spend on the page.