Petri Dish Neural Cellular Automata (pub.sakana.ai)

Researchers introduced Petri Dish Neural Cellular Automata (PD-NCA), a fully differentiable Artificial Life substrate that replaces the usual single, fixed NCA rule with a population of independent, continuously learning NCA agents sharing a common grid. Each agent maintains its own neural parameters and performs gradient descent in-the-loop during simulation, proposing local updates to cells through attack, defense and hidden channels. Proposals are gated by an aliveness mask (agents can only act where they or adjacent cells are alive) and a static randomized environment tensor competes as a persistent background adversary. Competing proposals are resolved by pairwise strengths computed from cosine similarity, transformed by a softmax temperature into contribution weights; the final state delta is the clipped weighted sum and aliveness is updated from normalized strengths, with non-viable mass redistributed. Empirical results show that learning-in-the-loop plus scale (more agents, larger grids) produces richer, open-ended dynamics—cycles, oscillations, emergent cooperation from purely competitive objectives, symbiogenesis-like structure formation and persistent complexity growth. The team ran hyperparameter sweeps (models up to ~500K params, up to 15 NCA, grids to 256×256) and evaluated novelty using video-compressibility and an open-endedness score based on VLM embeddings. PD-NCA thus offers a scalable, differentiable platform for studying within-lifetime adaptation, multi-agent morphogenesis, and open-ended evolution in neural systems.