Table of Contents

What happens if you control the rules of survival rather than the survivors themselves? Sakana AI turned that question into a browser-based simulation anyone can run. The God Simulator isn’t a game — it’s a research tool that makes evolutionary dynamics tangible, and the insights it surfaces connect directly to Sakana AI’s core research thesis: that evolution and collective intelligence offer a fundamentally different path to capable AI systems than raw scale alone.

TL;DR

  • The God Simulator is built on Neural Cellular Automata (NCA) — each pixel-organism is a small neural network that can grow, attack, defend, and learn
  • Users set the rules (survival thresholds, mixing rules, resource density) rather than controlling individual agents
  • Key finding: harsh rules cause extinction; too-easy rules produce fragile booms; alternating conditions can produce stable borders and even spontaneous cooperation
  • This is a public-facing demonstration of Sakana AI’s “evolution-first” research philosophy
  • Free to use at sakana.ai

What It Is

Neural Cellular Automata

You may know Conway’s Game of Life — a grid world where cells follow simple rules (alive/dead based on neighbor count) and complex patterns emerge from those rules. Sakana AI’s simulator extends this: instead of fixed rules, each pixel-organism runs a small neural network. It perceives its neighbors, integrates that signal, and acts — growing, attacking, defending, or attempting cooperation.

Because the neural networks are subject to selection pressure, evolution is literal here. Individuals whose network structure produces better survival outcomes reproduce; others don’t. Across many generations, behaviors emerge that no one explicitly programmed.

graph LR
    A[Pixel organism<br>Neural Cellular Automaton] --> B[Sense neighbors]
    B --> C[Neural network inference]
    C --> D[Act: grow / attack / defend / cooperate]
    D --> E[Environmental feedback]
    E --> F[Fit structures survive<br>and propagate]
    F --> A

Your Role: The Rule-Setter

You don’t control individual organisms. You control the environment:

  • Survival threshold: How hard is it to stay alive?
  • Mixing rules: What happens when different species meet?
  • Resource density: How abundant is energy?

This puts you in the position of setting incentive structures — not playing, but governing.

What the Simulator Reveals

Sakana AI documented several counterintuitive outcomes:

Harsh rules → Extinction: Set the survival threshold too high and nothing survives. The grid goes silent within a few hundred timesteps regardless of initial conditions.

Easy rules → Fragile boom: When conditions are too permissive, populations explode — but the selection pressure is so low that individual neural networks never evolve robustness. Tighten the rules slightly and the whole system collapses immediately.

Alternating strict and permissive conditions (the interesting case): Start permissive to let populations establish, then increase pressure. This tends to produce “crystallization” — stable territorial borders form between species. Under certain parameter combinations, cooperation emerges: two competing species begin protecting each other because the cost of continued conflict exceeds the cost of coexistence. No one programmed this behavior. It emerged from the incentive structure.

This maps cleanly onto evolutionary game theory — specifically the evolution of cooperation literature (Axelrod, Hamilton, Nowak) — but made viscerally observable rather than mathematically abstract.

Why Engineers Should Care

The God Simulator is an intuition pump for a problem that shows up everywhere in system design: how do incentive structures shape emergent behavior?

This is directly relevant to:

  • RL curriculum design: An environment too easy produces an underprepared agent; too hard produces nothing. The alternating-pressure finding from the simulator mirrors best practices in curriculum learning.
  • Multi-agent system design: The spontaneous cooperation emergence under specific conditions is the same phenomenon studied in multi-agent reinforcement learning research.
  • Organizational design: The same dynamics apply to team incentive structures, competitive vs. collaborative dynamics between teams, and how resource scarcity shapes culture.

Sakana AI’s Broader Research Direction

Sakana AI was founded by David Ha (former Google Brain Research Director) and Llion Jones (one of the original Transformer paper authors). Their research philosophy is explicitly anti-scale: rather than chasing larger models and more compute, they pursue evolutionary and collective intelligence approaches.

Evolutionary Model Merge (2024): Automatically merging existing open-source models using evolutionary algorithms — no gradient-based training, relatively low compute, can produce models that outperform their parents on specific tasks. Now integrated into frameworks like mergekit.

The AI Scientist: Fully automating the scientific research cycle — idea generation, literature search, experiment design, analysis, and paper writing. In 2025, AI Scientist v2 produced the first fully AI-generated paper to pass rigorous human peer review, published in Nature.

Darwin Gödel Machine: A self-modifying AI that rewrites its own code to improve performance, inspired by Schmidhuber’s theoretical work.

The God Simulator is the public-facing, accessible version of this research direction: it makes the abstract claim that “incentive structures determine system behavior” into something you can feel in your hands.

References

🇺🇸 English

Imagine you're God — but not the kind who moves pieces on a board. You set the rules of the universe and then step back. That's exactly what Sakana AI built, and it's one of the more thought-provoking things to come out of AI research this year.

The God Simulator is a browser-based evolutionary sandbox. You don't control any organism directly. You control the environment they live in — how hard survival is, how dense resources are, what happens when two different species collide. Then you watch what emerges.

The organisms themselves are tiny neural networks. Think of Conway's Game of Life, where simple rules produce complex patterns — but instead of hard-coded alive or dead logic, each creature in this simulator is running its own little neural network. It senses its neighbors, processes that information, and decides: grow, attack, defend, cooperate. And because bad neural network structures die off while good ones reproduce, you get literal evolution happening in real time. Behaviors nobody programmed appear because the incentive structure made them useful to survive.

And this is where it gets genuinely interesting. The simulator keeps surfacing counterintuitive results.

Set survival conditions too brutal — almost nothing survives. The grid goes quiet within a few hundred timesteps no matter how good the starting population looks. Too easy, and you get the opposite problem: populations explode, but they never develop any robustness because there's no pressure to. The moment you tighten conditions even slightly, the whole ecosystem collapses — a fragile boom that can't handle stress.

The sweet spot is alternating pressure. Start permissive, let populations establish, then crank up the difficulty. What you see under those conditions is crystallization — stable territorial borders forming between competing species. And in certain parameter combinations, something even stranger: two species that were fighting each other start protecting each other. Not because anyone told them to. Because at some point, the cost of continued conflict exceeded the cost of coexistence, and evolution found that solution on its own. That's the evolution of cooperation — a rich area of game theory with decades of mathematical literature — but now you can see it happen with your own eyes rather than just trust the equations.

For engineers, this should feel immediately familiar. The dynamics here are the same ones you fight in reinforcement learning curriculum design: too easy a training environment and your agent never learns to handle hard cases; too hard and it learns nothing at all. The alternating-pressure finding maps directly onto what practitioners call curriculum learning — progressively harder training environments. The simulator makes this intuition visceral rather than abstract.

It's also a clean model for thinking about multi-agent systems, or even organizational design. How do incentive structures between teams shape whether they compete or collaborate? Resource scarcity, clear survival thresholds, mixing rules between groups — the same levers exist in both the simulator and the real world.

Sakana AI isn't a random lab doing this. It was founded by David Ha, former Research Director at Google Brain, and Llion Jones, one of the original authors of the Transformer paper. Their core thesis is explicitly anti-scale: rather than just building bigger models with more compute, they're pursuing evolution and collective intelligence as an alternative path. The God Simulator is the public-facing tip of that iceberg.

The rest of the iceberg includes things like Evolutionary Model Merge — using evolutionary algorithms to automatically combine existing open-source models without any gradient-based training, often producing results that outperform either parent model. And The AI Scientist, their fully automated research system that in 2025 produced the first fully AI-generated paper to pass rigorous peer review and get published in Nature. And the Darwin Gödel Machine — a self-modifying AI that rewrites its own code to improve performance.

The God Simulator is their way of making the abstract argument concrete: incentive structures determine system behavior. Not the agents, not the components — the rules they live under.

Three things to take away from this. First: emergence is real and it's surprising — behaviors that feel like they'd need to be explicitly designed show up naturally when the selection pressure is right. Second: the middle path matters enormously — in evolutionary systems and in training AI, neither too easy nor too brutal produces robust outcomes; alternating pressure does. And third: Sakana AI is making a serious bet that the future of capable AI isn't just more scale — it's different paradigms entirely, and the God Simulator is one piece of evidence they're building toward that future rather than just talking about it.

🇹🇼 中文

如果你可以重寫生命的規則，世界會怎麼演化？Sakana AI 把這個問題做成了一個可以在瀏覽器裡玩的模擬器——他們叫它「God Simulator」，神明模擬器。

這個工具的核心技術叫神經細胞自動機，英文縮寫 NCA。你可能聽過 Conway 的「生命遊戲」——一個格子世界，每個細胞根據鄰居的狀態決定生死，複雜的圖案從最簡單的規則中浮現。Sakana AI 把這個概念往前推了一大步：格子裡的每個「像素生物」不是遵從固定規則，而是一個小型神經網路。它可以感知鄰居、學習、攻擊、防禦、合作。所謂的演化，在這裡不是隱喻——是字面意義上的：在環境壓力下，適者的神經網路結構被保留並傳播下去。

重點在於你扮演的角色：你不是控制個別生物，而是設定規則本身。生存門檻多嚴苛？不同物種相遇時誰佔優勢？資源有多豐沛？你是演化壓力的制定者，有點像一個解析度很低的上帝。

Sakana AI 公布這個工具時，分享了幾個讓人意外的觀察。

把生存門檻設太高，任何物種都站不住腳，整個格子世界幾百個時間步後就歸於空寂。這個很直覺。

但把條件設太鬆，會出現另一種問題：種群爆炸式成長，卻因為沒有任何選擇壓力，個體根本沒機會演化出健壯性。一旦稍微調緊規則，立刻崩潰。繁榮是假象。

最有趣的情境是「交替嚴寬」。先設寬鬆條件讓物種站穩，再慢慢加壓。這個過程會讓不同物種形成穩定的勢力邊界，甚至在某些參數組合下出現合作行為——兩個原本競爭的物種開始相互庇護，因為繼續消耗彼此的代價，比合作更高。這跟賽局理論裡「合作演化」的研究結果完全吻合。

對工程師和研究者來說，這個模擬器的價值不只是視覺享受——它是一個直覺泵，幫你建立「激勵機制如何塑造系統行為」的直覺。強化學習的課程設計、多智能體系統的競合平衡，背後的動力學跟這個模擬器展示的東西，是同一個問題的不同尺度。

要理解為什麼 Sakana AI 做這個，得先知道他們是誰。這間公司由前 Google Brain 研究員 David Ha 和 Transformer 論文原始作者之一 Llion Jones 共同創立，總部在東京。他們的研究路線跟主流大模型廠商明顯不同——不追求暴力擴張參數規模，而是走演化與集體智慧的路線。

這個哲學在他們幾個主要項目裡都看得到。Evolutionary Model Merge 不用梯度下降訓練，而是透過演化算法把現有開源模型合併，成本極低但可以在特定任務上表現出色。AI Scientist 是個全自動科學研究系統，從想法生成到論文撰寫全程 AI 主導，AI Scientist v2 在 2025 年產出了一篇通過同行評審、刊登於 Nature 的論文。Darwin Gödel Machine 更激進，可以直接重寫自己的程式碼來提升性能。

相較於主流大模型廠商——更大的參數、更多的資料、更高的訓練成本——Sakana AI 選的是另一條路：生物啟發、演化驅動，訓練成本相對低，但走的是完全不同的創新路徑。God Simulator 就是這條路上的一個公眾溝通工具，把原本很抽象的研究理念，變成任何人都可以親手體驗的互動介面。

整理一下今天的核心要點。

第一，生存規則的嚴苛程度決定系統能否演化出真正的健壯性——太嚴是滅絕，太鬆是虛假繁榮，交替施壓才能催生真正的適應能力與合作行為。

第二，這個動力學不只適用於生態模擬，它直接反映了 AI 訓練環境設計、多智能體系統，乃至組織激勵設計的底層邏輯。

第三，Sakana AI 用這個模擬器傳達的，是他們整條研究路線的核心信念：演化出來的複雜性，不需要頂層設計。

這個模擬器免費開放，在 sakana.ai 就可以玩。值得花半小時感受一下，當你掌控規則的時候，系統究竟有多脆弱——或者多頑強。

Tags

#ai#sakana-ai#evolutionary-computation#neural-cellular-automata#ai-scientist#research

Related Articles