---
title: "Simulating the World"
description: "Language predicts a storm without resolving a single molecule. First-principles simulation resolves every molecule but cannot reach a single living cell. The way to simulate the world is a new tokenization of physics."
canonical_url: "https://engineering.flexcompute.com/articles/simulating-the-world/"
markdown_url: "https://engineering.flexcompute.com/articles/simulating-the-world.md"
image: "https://engineering.flexcompute.com/images/og/simulating-the-world.png"
publish_date: "2026-08-24"
kind: "Essay"
series: "The Uncomputed World"
series_order: 2
authors:
  - "Zongfu Yu"
tags: []
---

To simulate the world is to know what will happen before it happens. First-principles simulation starts from the laws of physics, written as partial differential equations. They advance every physical quantity at every point in space, one step at a time.

The simplest simulations are routine, with one physics and few scales: light crossing a lens, air over a wing, currents in an antenna.

Engines, chips, fusion reactors are much harder. They couple more physics (multi-physics): on a chip, light heats the silicon, the heat expands the structure, and the expansion shifts the light that heated it. And they span more scales (multi-scale): the smallest scale fixes the resolution, the largest fixes the extent.

Then there is life, the ultimate multi-scale, multi-physics problem. A living organism requires coordination across many kinds of physics and across many scales of length and time; a chip is also built from layers of physics, but life is, in a sense, the hardest system to simulate. A single cell, ten thousand molecule-widths across, is a trillion points; one second of its life is a thousand trillion femtosecond steps. The total exceeds any computer ever built, for one cell alone. Can the world be simulated at all?

## Language is a reduced-order model of physics

There is a hint that it might be done, spoken by everyone every day: language.

Words represent the world at every level:

- **Physics: "wind."** Trillions of air molecules, moving together in one direction.
- **Chemistry: "rust."** Years of electron transfer, gathering as a reddish crust on iron.
- **Biology: "fever."** Billions of immune cells at work, heating the whole body.
- **Psychology: "afraid."** A brain's whole threat response, in a single feeling.
- **Sociology: "leader."** A web of alliances, converging on one person.

A word token is coined wherever physics produces a stable pattern humans care about. But tokens alone predict nothing. Prediction is the job of the token dynamics: the rule for how tokens drive one another.

"The sky is turning black and the wind is howling; a storm is about to hit." "Storm" follows "howling wind" because that is what happens next in the sky. The storm is predicted without resolving a single molecule, without spatial points, without time steps. The sentence is a simulation: the current state in, the next state out, no equations inside. That is what a reduced-order model does: keep a few variables in place of millions, enough to predict what matters.

A large language model (LLM) learns the token dynamics of language by training on trillions of tokens. It runs the whole model on a machine: the crudest simulator of the world, with all of language's advantage — and all of its handicap.

## The physical world model

As a simulator of the world, the LLM is handicapped first in its tokens: they are coarse. One word, "storm," stands for a million different storms and holds no fine detail: not when, not where, not how strong. But this part is curable: train the same machinery on pixels and measurements instead of text, and the tokens sharpen. Such efforts are underway; the field calls them world models.

Even with sharpened tokens, a deeper flaw remains: the dynamics still come from what has already happened. Asked how a machine never built will behave, the model answers just as fluently, with no sign that it is guessing. Pixel or word, the model still knows only where the world has been, never what the world can do. What the world can do is written in the laws of physics.

The guessing will not be trained away. It will be overcome by simulation that starts from the laws.

First-principles simulation is itself a tokenization. For the light crossing a lens, the token is the electric field at one point in space, and the token dynamics is Maxwell's equations, advancing each point from its neighbors.

But it is handicapped at the other end: language's tokens are too coarse; the laws' tokens are too fine. Fine tokens come in huge numbers: a trillion points for one cell, and the cell is already out of reach. Language, for all its guessing, covers a whole storm with one token.

The way to simulate the world is a new tokenization of physics: make each token hold more, without losing the law. Dynamics can be distilled from the equations, and every answer can be checked against them. Such a simulator runs far faster than the equations and stays far truer than language.

That is how we achieve our ambition of simulating the emergence of life from first principles. And perhaps we will find forms of life that have never been observed.