Grand Challenge: Predict Task-Relative Resource Frontiers Before Deployment

Modern scaling laws predict how loss changes with parameters, data, and training compute, but deployment budgets are often determined by a different object: the smallest representation that preserves the decisions required by a particular task. We propose a grand challenge for efficient deep learning: predict, before deployment, the minimum memory, sensing/intervention, communication, and inference resources needed to preserve a declared decision interface to tolerance $\epsilon$. The target should be task-relative rather than reconstruction-relative.