They have infinite amounts of training data, and probably lots of interested users who also like to push the limits of what the model is capable of and provide

They have infinite amounts of training data, and probably lots of interested users who also like to push the limits of what the model is capable of and provide all kinds of test cases and RLHF base data. They have millions of people training the AI for free basicallly, and they have engineers who pick and rate pieces of training data and use it together with other sources and manual training.