Approximating Nothing

Everyone reaches for the sharper tool — higher-order Runge-Kutta, BFECC advection, a finer grid — on a belief so common we forget it's a belief: that more accurate is more. It isn't. Accuracy is faithfulness to a referent. A solver exists to not lie about a physics that already exists. That's honest engineering. But faithfulness assumes there's something you're supposed to be true to, and most people never chose that assumption — they inherited it.

Notice how little of the interesting work was ever faithful to anything. Turing's two reaction–diffusion equations model no real animal, yet grow leopard spots. Conway's four rules — cells only on or off — are Turing complete. Boids: three local rules and no bird, but a flock. Perlin noise corresponds to no real surface and sits under half the textures you've seen. Each is crude, local, "wrong," and each out-produces the faithful model.

Rich Sutton called a version of this the bitter lesson: general methods riding on computation beat the knowledge we hand-build in. But his lesson still keeps a scoreboard — a better move, a lower error rate. A cellular automaton has none. Ask how accurate the Game of Life is and the question dissolves: accurate to what? It isn't a bad model of something. It isn't a model. It's a small world that runs.

Which is the real craft: not tuning accuracy, but designing the system. You set the local rule, the coupling, the boundary, the thing that drives it — and let disorder organize itself into form. This is why entropy was never the enemy. Prigogine's dissipative structures are fed by dissipation, not defended against it; cut the flow and the pattern dies. The living region is the edge of chaos — not the crystal, which is order and dead, not the noise, which is disorder and dead, but the thin line where a simple rule holds the two in tension. The design problem is finding that line and staying on it.

So when I stopped chasing better advection, it wasn't disdain. Accuracy is one term in a large system, and spending compute there while the structure does the real work is just a bad gradient. The rest — the multi-scale fields, the patterns that assemble themselves out of a scattered grid — isn't me being faithful to a fluid. It's me writing a small world and letting it run. The most interesting move was never sharpening the knife. It was setting the referent down and making something that answers to nothing.