Show HN: ND Loss Optimizer Arena (nd-optimizer-arena.vercel.app)

ND Loss Optimizer Arena is an interactive, browser-based playground for visualizing how different optimizers and hyperparameters navigate synthetic loss landscapes—branded with “Fractal Basins & Orthogonal Spikes.” The tool generates procedural loss fields (with fractal basins, sharp orthogonal spikes and noisy/clipped gradients), shows live 2–15D slices with log‑spectrum coloring to reveal multi‑scale frequency content, and lets you click to set a start point, play/step through optimization, reset or regenerate landscapes. It exposes common optimizers (SGD, Momentum, RMSProp, Adam, AdamW and a Muon variant), learning‑rate schedules (constant, step, cosine, exponential, warmup+cosine), batch size, gradient clipping and step‑size controls so you can see loss curves, current/ best loss and per‑step dynamics in real time. For the AI/ML community it’s a compact, hands‑on testbed for intuition, debugging and teaching: you can watch basin hopping, saddle traversal, spike avoidance and the effects of noisy or clipped gradients, and directly compare how optimizers and LR schedules affect convergence and stability. The ability to slice up to 15D and view log‑spectrum structure makes it useful for exploring why certain algorithms fail on sharp features or high‑frequency loss components, and for prototyping hypotheses about robustness, hyperparameter sensitivity, and optimizer design before trying them on real models.