While Phi is a good example of this technique, Phi as a model is very anemic. It was recently part of a CTF hosted by Microsoft, where other models were also included- I assume MS was looking to test performance of Phi against the competition... Phi performed the worst. Its outputs easier to predict, quicker to construct injection attacks and jailbreak. All models utilized the same defenses. As I have also trained and fine-tuned models using synthetic data, I have seen this approach increase determinism and increase predictability. Some might see this as a good thing- but I think it depends. On one hand this opens the model to several adversarial attacks such as jailbreaking, extraction, etc, on the other hand some consumers may prefer less random outputs.