Orca 2 is an example of an opensource model that was built to better collect and build on synthetic data: https://www.microsoft.com/en-us/research/publication/orca-progressive-learning-from-complex-explanation-traces-of-gpt-4/

The case I think being made is that building LFMs on the Internet gets you closer to an average internet users level of our put, using reinforcement learning you can further curate the outputs, then finally using these techniques you can generate even tighter high quality models.

It’s interesting stuff for sure.