Sakana AI's God Simulator: Watching Evolution Happen in Real Time

What if you didn't play as one of the survivors, but as the thing that writes the rules of survival itself? That's the question Sakana AI turned into a browser-based simulation that anyone can run. They call it the God Simulator, and here's the important part — it's not a game. It's a research tool. It takes the abstract idea of evolution and makes it something you can watch happen, in real time, with your own hands on the dials.

Let's start with what's actually running under the hood, because this is the clever bit. You might know Conway's Game of Life — that grid world where cells live or die based on how many neighbors they have, and out of those dead-simple rules, these astonishingly complex patterns emerge. Sakana AI takes that idea and cranks it up. Instead of every cell following the same fixed rule, each pixel-organism is running its own tiny neural network.

So here's the loop each little organism goes through. It senses its neighbors. It runs that signal through its neural network. And then it acts — it might grow, it might attack, it might defend, or it might try to cooperate. The environment reacts to what it did, and here's where evolution kicks in literally: the organisms whose network structure led to better survival get to reproduce, and the ones that made bad calls don't. Run that across many, many generations, and you get behaviors that nobody sat down and programmed. They just emerge.

Now, your job in all this. You are not steering individual organisms around like little chess pieces. You're setting the environment. Three main levers. First, the survival threshold — basically, how hard is it just to stay alive? Second, the mixing rules — what happens when two different species bump into each other? And third, resource density — how much energy is floating around out there? So you're not playing the game. You're governing it. You're writing the incentives and then watching what the population does in response.

And what it does is genuinely surprising. Sakana documented a few outcomes that are worth sitting with.

First one: crank the rules too harsh, and you get extinction. Set that survival threshold too high and nothing makes it. The grid just goes quiet within a few hundred timesteps, and it doesn't matter how you seeded it — everybody dies.

Okay, so go the other way, make it easy. What you get is a fragile boom. Populations explode, the screen fills up — but here's the catch. Because life is so easy, there's basically no selection pressure. The neural networks never have to get good at anything, so they never evolve any real robustness. And the moment you tighten the rules even a little, the whole thing just collapses instantly. It was a house of cards the entire time.

But here's the case that's actually fascinating — alternating conditions. You start off permissive, let the populations get established and find their footing, and then you turn up the pressure. And what tends to happen is something Sakana calls crystallization. Stable territorial borders form between the species. They sort themselves into defined regions. And under certain parameter combinations, you get something remarkable: cooperation emerges. Two species that were competing start actively protecting each other — because at some point the cost of endless fighting became higher than the cost of just living side by side. Nobody coded that. There is no "cooperate with your rival" instruction anywhere. It fell out of the incentive structure on its own.

And if that rings a bell, it should — this is evolutionary game theory, the whole evolution-of-cooperation literature, the Axelrod and Hamilton and Nowak stuff. But instead of it living in equations on a page, you're watching it play out as colored pixels on a grid. It goes from abstract to visceral.

So why should an engineer care about a pixel simulation? Because it's what I'd call an intuition pump for a problem that is absolutely everywhere in system design: how do the incentives you set shape the behavior that emerges? You don't control the behavior directly — you almost never do — you control the conditions, and the behavior falls out.

Think about reinforcement learning curriculum design. An environment that's too easy gives you an agent that's underprepared for the real thing. Too hard, and the agent learns nothing at all. That alternating-pressure sweet spot from the simulator? That's basically curriculum learning best practice, staring right back at you. Same story with multi-agent systems — that spontaneous cooperation is exactly the phenomenon people study in multi-agent RL. And honestly, it scales all the way up to organizational design. Team incentives, whether teams compete or collaborate, how scarcity shapes a culture — same dynamics, different substrate.

Now, why does Sakana AI care about this, and why does it matter who they are? The company was founded by David Ha, former research director at Google Brain, and Llion Jones, one of the original authors on the Transformer paper — the "Attention Is All You Need" paper. And their whole research philosophy is deliberately anti-scale. Instead of chasing ever-bigger models and ever-more compute, they're betting on evolution and collective intelligence as a different road to capable AI.

You can see that thread through their other work. There's Evolutionary Model Merge, from 2024 — using evolutionary algorithms to automatically merge existing open-source models together, no gradient training, relatively cheap compute, and it can produce models that beat their own parents on specific tasks. That one's now baked into tools like mergekit. There's The AI Scientist, which automates the entire research cycle — generating ideas, searching literature, designing experiments, analyzing results, writing the paper. And in 2025, its version two produced the first fully AI-generated paper to pass real, rigorous human peer review, published in Nature. And there's the Darwin Gödel Machine, an AI that rewrites its own code to improve itself, inspired by Schmidhuber's theoretical work.

The God Simulator is the friendly, public face of all of that. It takes their core claim — that incentive structures determine system behavior — and turns it into something you can literally feel in your hands.

So let me leave you with three things to hold onto. One: the mechanism. Neural cellular automata means every organism is a little evolving neural network, so evolution here isn't a metaphor, it's the actual process running on screen. Two: the central lesson about pressure. Too harsh kills everything, too easy builds something fragile that shatters the moment reality shows up, and it's the well-timed alternation of pressure that produces stability — and sometimes cooperation nobody designed. And three: this isn't just a toy. The exact same dynamics govern how you train agents, how you build multi-agent systems, and how you design incentives for teams of actual humans.

You set the rules. The behavior takes care of itself. And it's free to try, over at sakana dot ai — go turn some dials and watch what your world decides to become.