Automating the Search for Artificial Life with Foundation Models

Artificial Life researchers have spent decades hand-crafting the rules for digital worlds like Conway's Game of Life or Karl Sims' evolved virtual creatures, tweaking parameters for months just to stumble on something interesting. Sakana AI, working with MIT, OpenAI, IDSIA, and ALife veteran Ken Stanley, thinks that grind is no longer necessary. Their new system, called Automated Search for Artificial Life, or ASAL, hands the search itself over to vision-language foundation models.

The idea is straightforward once you see it: instead of a human guessing which update rules might produce interesting behavior, ASAL runs a simulation, feeds the resulting video into a foundation model, and checks how well that output matches a target description, like "a self-replicating pattern" or "a Fibonacci spiral in nature." It then searches for the simulation parameters that make the actual footage line up with that prompt. Sakana tested this across five classic ALife substrates — Boids, Particle Life, Lenia, Neural Cellular Automata, and Game of Life — and it worked across all of them, producing things like emergent flocking in Boids and cell-division-like dynamics in Particle Life.

The more striking result is what happened in Game of Life itself. ASAL didn't just replicate Conway's original ruleset; it discovered entirely new cellular automata rules that Sakana says are more open-ended and expressive than the 1970 original, a system famous for spawning self-replicating "spaceships" and even simulating computers inside itself. Finding rules like that by hand took the ALife community years. ASAL found variants of it through search, guided only by a model's sense of what looks novel over time.

Beyond matching a target, ASAL also handles two harder problems: hunting for simulations that keep generating novelty indefinitely, rather than settling into repetition, and mapping out a whole substrate to show researchers the diverse range of "alien worlds" hiding inside a given rule space. Sakana even built a new substrate, Particle Life++, where the interaction rules themselves evolve, and ASAL was able to surface open-ended dynamics in it right out of the box, no tuning required.

Sakana frames all this as part of a two-way exchange between AI and Artificial Life research: they've previously borrowed ALife concepts like evolution and self-play to build better models, and now they're using foundation models to accelerate ALife itself. The code is open-sourced on GitHub, and the team is explicitly inviting outsiders to point ASAL at their own custom simulations, whether that's something like ALIEN or JaxLife, to see what shows up.