Why this exists in ML Deep learning optimizes non-convex objectives; your tooling needs to detect “stuck” vs “slow.”