Chess Gas is a cellular automaton in which chess pieces make random legal moves on a tiled board. It was built as a way to think about statistical mechanics, emergent behaviour, and what happens if you take the rules of chess seriously as a physics — pieces as particles, captures as collisions, material as energy.

🧪 The basic idea

A standard chess starting position is tiled across a large board — white on the bottom ranks, black on the top. Pieces make random legal moves in alternating turns. What emerges is a two-species lattice gas: the particles are chess pieces, the interactions are captures, and each piece type has a different mobility and mass. Pawns are slow and directional; queens are fast and omnidirectional. Captures are inelastic collisions by default. Total material — the sum of piece values — can only decrease under ordinary rules, though several collision and potential modes let you break that deliberately.

📐 Board size

Boards run from a plain 8×8 up to 2048×2048. Past a certain size the board no longer fits the screen at 1 pixel per cell — zooming out further than that switches to a sampled rendering mode so huge boards stay smooth and interactive rather than grinding to a halt. At extreme zoom, isolated pieces on a very sparse board can occasionally fall between samples; that's a deliberate speed trade-off, not a bug.

🎲 Presets & randomisation

The I/O tab holds a set of curated presets — starting points chosen to show off a specific kind of behaviour (a percolating gravel field, bishops looping forever on a torus, a barrier only positively-charged pieces can cross). Pick one and tap load, then import to apply it.

The dice button has three modes, cycled by holding it down — a short tap runs whichever mode is active:

⚄ both⚄P physics only⚄D display only

Physics-only re-rolls the rules (topology, collisions, potentials); display-only re-rolls how it looks (colours, trails, effects) without touching the simulation underneath.

✏️ Editing the board

The Edit tab has three tools. Pieces lets you paint or erase individual pieces by hand. Sources & sinks lets you paint cells that continuously spawn or remove pieces, turning the board into an open, non-equilibrium system that never settles. Potential lets you paint an energy landscape directly — repulsive or attractive — that biases how pieces move without changing any rule.

🔁 Auto stop / restart

The bar under the toolbar lets the simulation stop or reset itself automatically once a condition is met — material falling below a threshold, captures drying up, a fixed step count. Combine several conditions with the gate control (any/OR or all/AND). Set the action to "nothing" to just run forever, which is the default. This bar is Sandbox-only — Experiment mode has its own separate autostop logic, and Demo mode runs its own independent auto-reset timer instead (see below), unaffected by whatever this section is set to.

🎬 Demo mode

A fullscreen presentation view with a small floating control cluster — reachable from the mode switcher (bottom bar on desktop, its own tab row on phone). Auto-resets on its own timer, cycling through curated colour/potential-field combinations by default. Several of the cluster's buttons do a second thing on a long press (hold ~half a second):

Tapping any of these shows a brief tooltip confirming the new value — the darkening background and changing icons are meant to be readable from across a room, the tooltip is for when you're close enough to check.

💾 Import / export

The settings string in the I/O tab captures every adjustable setting — physics, display, layout — as one paste-able string. Board contents are not included, only settings; exporting doesn't snapshot the current position, and importing rebuilds a fresh board under the new rules. Use it to save a configuration you like, or to share one.

📊 Reading the physics

The Data tab has rolling sparklines, full-run history graphs, a phase-space scatter plot, and a spectrogram showing per-row structure over time — in linear mode you can scroll or drag back through the whole recorded history.

⚙️ What you can change

Topology — bounded, toroidal, cylindrical X/Y, Klein bottle, Möbius strip, projective plane, sphere.
Collision rules — capture, elastic, inelastic, annihilate, explode, exchange, pacifist, convert, wololo.
Promotion — random, queen, weighted, underpromotion, reflect, explode, board-match.
Potential fields — barriers, gravity wells, saddles, gradients, periodic lattices, mazes, gravel, plus piece-generated halos that move with the pieces themselves.
Dynamic potentials — oscillating, travelling-wave, breathing, or rotating modulation on top of any static field.
Sources & sinks — open the system; spawn or remove pieces at painted cells.

⌨️ Keyboard shortcuts

Space play/pause   R reset   . step   [/] speed

WASD/arrows pan   Z/C zoom   0 reset view

Tab cycle board init   F cycle field overlay   T cycle trail mode

J jitter   Q/E toggle panels   K this page

Click a piece track its path   Ctrl+drag paint potential   Shift+drag paint sources/sinks

🌀 Philosophy

This is a 2009 idea, sat on for seventeen years because actually building it was a lot of tedious plumbing, and only actualised now because a language model was willing to do that plumbing in a single long conversation [7]. That's the parallel worth drawing more than any other: cellular automata had their moment of pop-cultural fashionability in the 2000s largely because home computers had finally gotten fast enough to run them for fun rather than research [1] — the compute showed up, and a genre of hobby-project followed it. The same thing is happening again right now, just with a different resource becoming suddenly abundant. It's not a coincidence that this project could only get made once that resource was LLM coding assistance rather than cheap FLOPs.

Credit where it's due and not more: Wolfram's actual theory-of-everything pitch [1] reads, and read at the time, as fairly confident crankery. What this project took from that era was its aesthetic residue: small grid diagrams, a fascination with rules simple enough to state in a sentence producing behaviour too complicated to predict, and the general permission to take a toy seriously. The direct ancestor is Leonard Richardson's Dada Chess [2] — two computer players making legal-but-random chess moves against each other, forever, live on a webpage since the late 2000s. Chess Gas takes that same core gag and asks the more physics-flavoured question hiding inside it: if pieces are just particles that collide, what does the resulting gas actually do?

The other lineage is more recent and more shameless. The specific pleasure of watching an ordinary board game get extrapolated outward into something bigger, gaudier, and increasingly unrecognisable — while the mechanical core stays exactly, legibly the same — is the whole trick of Balatro [6]. That's the drip this project is actually chasing: the feeling of a familiar ruleset getting steadily more colourful and more absurd the longer you let it run, until a plain chessboard is a two-thousand-cell gas simulation with a spectrogram.

📚 Bibliography

[1] Wolfram, S. A New Kind of Science. Wolfram Media, 2002. — less for the science than for the small diagrams, and for being the reason cellular automata briefly had a fandom in the 2000s.

[2] Richardson, L. Dada Chess, crummy.com, late 2000s. — the direct ancestor: two players making random legal chess moves against each other, forever.

[3] Anonymous stock photograph of a giant spherical chessboard, vintage unknown. — for the imagination of it.

[4] Rivet, J.-P. and Boon, J.-P. Lattice Gas Hydrodynamics. Cambridge University Press, 2001. — for how the particle-gas mechanics actually work.

[5] A vision of Netzach consciousness, received upon the breaking of a fever in the late 2000s, releasing the author from the infinite chess dimension. — this software being a Hod vision of that vision.

[6] Balatro, LocalThunk, 2024. — for the drip: an ordinary game pushed outward into something bigger, louder, and more colourful, one system at a time, without ever changing what it fundamentally is.

[7] Claude, Anthropic, 2026. — the famous vibecoder, for finally doing the seventeen years of tedious plumbing between the idea and the thing, one conversation at a time, including this page about itself.