Why the leading marble sings
Every decision in RaceCut descends from one rule: whoever is winning is singing. It sounds like a gimmick. It is actually the entire product, and getting it to feel right ate more engineering than the renderer did.
If you’ve seen the marble videos on TikTok - four coloured marbles hurtling down a course while a cartoon voice belts the chorus, the voice changing every time someone overtakes - you’ve seen the mechanic. Viewers don’t consciously notice the rule. They notice that the video keeps paying off every couple of seconds, and that they picked a marble in the first second without being asked to.
A race is a free hook
Most short-form video lives or dies on the first second, and most creators solve that with text hooks (“wait for it”) or editing tricks. A race solves it structurally: the moment two named things are moving, the viewer has involuntarily chosen one. That choice is retention. The comments write themselves - team red, Bea was robbed, rig it for Rex next time - and every comment is the brief for tomorrow’s video.
What the singing adds is a second engine on top: the leaderboard becomes audible. You can look away from a race; you can’t not hear an overtake when the voice changes mid-word. That’s the trick. The hand-off isn’t a visual detail, it’s the payoff, and it fires every time the physics produces a lead change.
The hand-off has to land on the physics, exactly
Our first version swapped voices on a timer. It felt wrong in a way that took us embarrassingly long to name: the ear noticed that the swap and the overtake were up to half a second apart, and half a second is enough to break the illusion that the marble is singing. Now the pipeline works backwards from the simulation: we bake the race first, extract the exact timeline of who leads when, and send those segment boundaries to the voice engine so the vocal switches on the frame the lead changes. If a lead lasts less than about a second and a half we don’t switch at all - a voice that sings one syllable reads as a glitch, not a joke.
Rigged races are real races
Creators asked to pick the winner almost immediately (comment sections demand justice for yesterday’s loser). The obvious implementation is to fake it - animate the outcome you want. We refused, for a practical reason: faked physics looks faked. Marbles that decelerate on cue read as wrong even to twelve-year-olds, especially to twelve-year-olds, and the comments turn on you.
Instead, every race is a deterministic simulation from a seed. When you rig a race, RaceCut re-rolls seeds and re-simulates until it finds a run of honest physics where your pick happens to win, then renders that one. Nothing on screen ever disobeys gravity; you just curated which universe got filmed. (Fair races publish their seed - the same seed always reproduces the same race, which has settled at least one comment-section dispute we know of.)
The course is composed from the song, not decorated with it
The other half of the illusion is musical. If the race and the song ignore each other, you have a screen recording with background music - the thing we were trying not to build. So the song goes in first: onset detection finds the beats and the bright, loud sections, and the course generator places its drops, bombs and palette shifts on those moments. In Musical Path, the corridor itself is built so the ball’s bounces land on the beat - the level is the sheet music.
This is also why every render is different. A fresh seed produces a new race through a course composed for the same song - which matters enormously if you post daily. The format survives repetition because the physics supplies infinite variation inside a structure the song dictates.
What the audience taught us
- Names beat colours. Races with named marbles get argued about; races with anonymous colours get watched once. The studio pushes you to name your cast for a reason.
- 15–30 seconds wins. We render up to 60, but the chorus-length race is the one that loops, and loops are what the algorithm counts.
- The winner matters less than the rivalry. The accounts that grow run the same cast every day. Persistent characters turn a video into an episode - that’s where we’re taking the product next.
The mechanic looks like a toy. Underneath it is a physics engine doing search, a beat detector doing composition, and a voice pipeline doing frame-accurate hand-offs - all so that a marble can appear to want to win badly enough to sing about it.