Do you think synthetic data can help with AI model evaluation and fine-tuning workflows?

I am an AI Engineer, and I find synthetic data to be of great help, not as a replacement for real data, but to augment it.

It has always been useful for training, but for me, it has been especially useful for coming up with various scenarios to test different kinds of inputs. I believe synthetic data can be very helpful for evaluating agent traces and outputs, simulating different scenarios, and testing edge cases.

There is a lot of discussion around model collapse, which is a valid concern, but several recent studies and industry experiments also show that combining real-world and synthetic data can improve performance, especially for domain-specific tasks and edge cases.

Curious where people currently stand on this.