Google's Lion optimizer wasn't hand-designed — it was discovered by a program search over possible update rules, and it turned out to look like sign-based momentum. arXiv:2607.01552 pushes that idea further: instead of searching a discrete space of symbolic programs, it embeds update rules into a continuous latent space and runs Bayesian optimization directly on that manifold. The takeaway this toy tries to make tangible: SGD, Momentum, Adam, and Lion aren't four unrelated algorithms — they're four points sitting inside one smooth, continuously-connected space of possible optimizers, and every other point in that space is a valid, runnable update rule too.
Drag anywhere inside the square and you get a genuine hybrid, not a cosmetic slider: the ball's momentum decay, its use of sign() vs. raw gradients, and its Adam-style adaptive normalization are all bilinearly blended from the four corners live, every animation frame. Watch pure SGD crawl and oscillate up Rosenbrock's curved valley, Momentum overshoot past it, Adam adapt per-coordinate, and Lion take clipped, decisive, constant-size steps — then drag toward the middle and watch a new hybrid rule emerge that behaves like none of them individually.
Simplified for interactivity: single shared momentum buffer with bias correction, no weight decay, gradient- and step-norm clipping for stability. Faithful to the core mechanics of each optimizer, not a production implementation.