“Distillation is a useful technique for creating smaller, task-specific models that capture a specific capability from a larger, more expensive, high-latency model.”