This talk has moved to Spring 2019 — There exist interpolation methods with optimal/near-optimal prediction risk under noise.