Teach one cell to become an organism.

See the neighborhood, then predict a change.

A cell stores premultiplied RGBA plus twelve learned hidden channels. Fixed, zero-padded identity and Sobel x/y filters read each channel's 3 × 3 neighborhood. The resulting 48 values enter a shared pointwise network; a 50% fire mask applies its residual update. Pre- and post-update alpha masks keep growth attached to living cells and clear dead hidden state.

Judge an unfolding process, not one frame.

Each optimizer update records a random 64–96-generation trajectory. The same target is scored at each of its final sixteen states, so the organism must stay correct instead of briefly passing through the image. Its loss combines RGBA L1 + L2 error, local Sobel edge alignment, and an explicit alpha-shape L1 + L2 penalty across every cell and batch sample: zero is an exact match, and lower is better. The shape term directly teaches which cells should live or die instead of leaving silhouette coverage only to curriculum scoring. A hidden-state overflow penalty also pulls every channel back toward the stable −1…1 range before a large value can turn into an exploding rollout.

BPTT (backpropagation through time) carries those errors backward through the entire tape. Adam (adaptive moment estimation) then turns the resulting gradients into weight updates using momentum and an adaptive step size for each weight.

Practice growth, persistence, and repair.

Evolved grids return to persistent pool slots and become later starting states. This experiment retains 512 histories while keeping the GPU batch at eight. The worst selected sample is replaced by a fresh seed, while healthy samples eventually receive round or elongated target-anchored cuts and must recover the complete target. This widens the target's basin of attraction: more partial, old, and differently damaged states naturally evolve back toward it.