"The G in GPU is for Graphics damnit " (ut21.github.io)

A researcher’s blogpost documents a GPU-accelerated PyTorch implementation and performance analysis of a Physarum (slime‑mold) growth simulator that blends an agent-based “data” layer with a continuum “trail” field (pheromones). Each particle has position, heading and three sensors (front/left/right); agents sample the trail with bilinear interpolation (torch.nn.functional.grid_sample), steer based on sensor readings, deposit pheromone, and the trail undergoes a 3×3 mean-filter diffusion plus multiplicative decay. The author ran large-agent simulations on CUDA, experimented with aesthetic initializations (including seeding the field from grayscale images) and showcased uses for procedural terrain/cave generation and a deterministic “encryption” scheme where knowing parameters reconstructs an image. The post also frames the simulator as a differentiable dynamical system and explores quantitative sensitivity: mapping boolean average sensitivity to continuous directional derivatives (sum of norms of Jacobian columns) and noting that stepwise Jacobian products govern perturbation growth — measured by Lyapunov exponents. Key implications for AI/ML and graphics: this is a GPU-friendly, differentiable procedural model amenable to optimization, learning or inverse-design; it demonstrates how systems work (and DSLs like Triton) matter for performance, and suggests rigorous stability/sensitivity analyses can assess pseudo‑randomness and robustness for generative or simulation-based ML pipelines.