What Dmitry Brant has built is a park. Inside it the speed of light is 5 km/h — a brisk walk. There is a colonnade; two identical 60 m red-and-white rulers, one laid along Z and one along X; a set of closed arches, so you can watch a circle become an ellipse; a row of beacons that each flash once per second of world time; a carousel; a shuttle sliding back and forth along X; a Ferris wheel; and two STL sculptures on plinths. The carousel cars and the Ferris wheel gondolas all run at 0.75c, in opposite directions on either side. W A S D accelerate — but holding a key down adds rapidity linearly, not velocity, so you never arrive at five. The velocity readout carries five decimals for exactly one reason: so you can watch 4.99995 km/h refuse to become 5. Each of the four effects gets its own key: L for light-travel delay, G for aberration and contraction, C for Doppler color, B for relativistic beaming. No dependencies, no build step, one WebGL2 canvas, forty-four commits. His own account of it runs to two sentences: I've always wanted to make a visualization where the speed of light was scaled down to human speeds; here is the first version.

Place the remainder first. Special relativity is about as thoroughly constructed as anything we own: 1905, a few lines of algebra, no adjustable parameters, tested to absurd precision. But the formalization carries measurement, not sight. In 1924 Anton Lampa asked, in Zeitschrift für Physik, what a moving rod would look like. Nobody took it up, and the question sat buried for thirty-five years. In 1959 James Terrell published "Invisibility of the Lorentz Contraction," and Roger Penrose arrived independently at the same result that year: a sphere moving at speed does not photograph as squashed, it photographs as turned. A theory closed for fifty-four years, and its own appearance was its remainder. That remainder is still not cleaned up. Only in 2024 did anyone photograph it in a laboratory: femtosecond laser pulses, a gated intensified camera with gating as short as 300 picoseconds, an effective speed of light pushed below 2 m/s, and snapshots of a sphere and a cube. The paper appeared in Communications Physics in 2025, with an addendum this year. The authors say outright that the work also marks the hundredth anniversary of Lampa's paper. A hundred years before a remainder gets its first photograph.

But this park's real remainder is not on the physics side. It is on the graphics side. Real-time 3D rendering is itself a construct: backface culling, low-poly geometry, static tessellation, depth buffers — each a standard practice chiselled out over decades, each of them saving real money in compute. And underneath each one sits a physical premise nobody ever wrote down. Drop the speed of light to 5 km/h and the premises fall out one by one, and the form they take on the way out is a bug. The shortest line in the README reads: backface culling is off. The reason is one sentence — Terrell rotation shows you faces that point away from you, and culling them would delete the effect. "You cannot see the back of an opaque object" is the lowest and least-questioned assumption in the whole rasterization pipeline. It is not mathematics. It is a statement about physics, and in this park it is false. Likewise: a cube drawn with eight vertices would still look like a cube at 0.9c, which is wrong, so the scene is tessellated to about 485,000 vertices by default — not for detail, but so that straight edges are able to bend.

The ground is what the transform punishes hardest: a flat plane of straight edges bent into curves, with faceting error going as the square of the angular step. So the ground became a polar disc graded exponentially from 1.2 m out to 4 km, its grid and mottle computed from world coordinates in the fragment shader, which lets the mesh slide along with the eye without anything visibly moving. He tried anchoring it at the origin: 500 m out, the spacing beneath your feet was 7.8 m. Finding the light-travel delay means solving g(a) = a − |p(t − a) − eye| / c = 0. While |v| < c, g′ never reaches zero — but at the corners of a carousel car, travelling at 0.857c, g′ falls to 0.14, and plain Newton flings the iterate most of the way around the orbit. Every vertex solves independently, so neighbours land on different revolutions and the triangles tear open across the scene. His answer: bracket the delay so a solution is guaranteed inside it, fall back on bisection, and accept Newton only when it lands in the bracket. Twenty-four iterations, worst error 0.06 px. Not one of these could be read off the Lorentz transformation. They only come out of building. That is the actual shape of the chisel-construct cycle: every cut leaves the construct with a layer it cannot itself process, and the next cut starts from that layer.

The Ferris wheel is where the judgment shows. The gondolas hang level, so each is in pure translation, each has a rest frame, and each can be specified at rest and then contracted along its own motion — the same treatment as a carousel car. The rim cannot. A rigidly rotating ring has no global rest frame to be designed in; that is Ehrenfest's paradox, and the material genuinely is strained in its own frame. He does not route around it and he does not invent an answer. He gives the rim's worldlines directly in the world frame, and writes one line: they are subluminal and self-consistent, which is the whole of what the renderer is entitled to ask. Entitled to ask is a refusal to close. He could have fabricated an approximate rest frame for the rim and not a single viewer would have caught it — that would have been the first chisel, the one that smooths the remainder away. He declines, and leaves "the theory has no answer here" standing in the README. The same restraint runs through the "what's not accurate" section: past violet, extreme blueshift turns metallic silver for x-rays and then bright white for gamma rays, and he labels it himself — totally unphysical, just to get a sense of it.

Three reasons seeing this now matters more than seeing it later. First, it is the first version, and he is the one saying so. Forty-four commits and seven stars on the repository; six hundred and seventeen points and two hundred and eighty-two comments for the same object on Hacker News. The two orders of magnitude are looking at the demo, and the remainder is in the README. His one comment in that thread is the key to the whole thing: none of the mathematics is particularly complicated, so if you get the math right, then all of the predicted effects emerge automatically. He is right — and precisely because he is right, everything hard landed outside the mathematics: vertex budgets, brackets, subdivision thresholds, that one disabled line of backface culling. Second, a slot is already waiting. MIT Game Lab's A Slower Speed of Light (2012) long ago took the category "relativity teaching demo," and the day this link turns up in a course reading list it becomes a thing in that slot — construct, useful, no longer interrogated — and the first thing discarded will be the README, which is a record of what graphics had to give up in order to let light run slow. Third, the other end is closing. Do the arithmetic: 5 km/h is 1.39 m/s, slightly slower than the "below 2 m/s" that the femtosecond laser and picosecond gated camera achieved in the lab. Both are the same gesture — slow light down until a person can stand inside it — one paid for with 300-picosecond gating, the other with forty-four commits. Once the experimental side settles into a canonical, citable account of the appearance, this park stops being an inquiry and becomes an illustration of it. That day has not come: the line about the rim having no rest frame is still standing in the README, and the speedometer still carries five decimals, waiting for you to watch 4.99995 refuse to become 5.

Han Qin (秦汉) · @hqinjarsy · SAE Judgment & Aesthetics