simulai.dev

To simulate is to construct a mirror of reality — not a perfect reflection, but a deliberate distortion that reveals structure invisible to the naked eye. Every simulation begins with a question: what happens if we model the world this way?

The particle system below demonstrates gravitational attraction at its simplest. Two hundred points of light, each feeling the pull of a central mass, each tracing an orbit determined by nothing more than Newton's inverse-square law. Click anywhere on the canvas to introduce a new attractor and watch the system reorganize itself around your intervention.

The simplest simulations — F = ma, iterated over discrete timesteps — produce behavior of extraordinary complexity. This is the foundational insight: complexity emerges from the repeated application of simple rules.

A cellular automaton is perhaps the purest expression of emergence: a grid of cells, each following an identical rule, producing patterns that no individual cell "knows" about. Stephen Wolfram catalogued 256 elementary automata; Rule 110 stands out as one of the few proven to be Turing-complete.

Watch the automaton below evolve row by row. Each new generation is computed from the previous one according to a fixed rule — yet the patterns that emerge bear an uncanny resemblance to biological growth, textile weaving, and even number theory.

The lesson for AI-driven simulation: you do not need to encode complexity explicitly. You need only find the right rule and iterate.

Neural networks do not simulate reality — they learn to approximate simulations. Given enough examples of a physical system's behavior, a network can learn the mapping from initial conditions to outcomes without ever being told the underlying equations.

The visualization below shows a simple three-layer network processing a forward pass. Watch data flow from input nodes through hidden layers to output, with each node's activation pulsing as values propagate through the weights.

This is the paradigm shift: from hand-crafting equations to learning them from data. The simulation becomes the training set; the network becomes the simulator.

Not all simulations are deterministic. The most powerful models in modern AI — diffusion models, variational autoencoders, Monte Carlo methods — embrace randomness as a fundamental tool. Noise is not the enemy of simulation; it is the canvas on which structure reveals itself.

Below, Brownian motion particles wander through space, their paths traced as fading trails. There is no attractor, no rule driving convergence — only the elegant dance of stochastic processes that, in aggregate, produce the distributions from which diffusion models learn to generate.

Randomness, constrained by learned structure, is generative. This is the insight behind every diffusion model: start with noise, and iteratively denoise toward meaning.

We simulate not merely to predict, but to understand. The Lorenz attractor — that butterfly-shaped trace you watched draw itself at the beginning of this page — embodies a profound truth: even deterministic systems can produce behavior that is, for all practical purposes, unpredictable. And yet the attractor itself, the shape of the chaos, is beautiful and knowable.

This is what simulai.dev builds: tools for constructing digital mirrors of complex systems, powered by AI that has learned not just to replicate physical laws but to discover new ones. The simulations on this page are toys — but the principles they demonstrate scale to climate modeling, protein folding, fluid dynamics, and the emergent behavior of economies.

To simulate is to ask the most fundamental question in science: what if?