A city of six shapes, and a film about robots

The GUI era, and why it had to die

Day one ran two things in parallel, and the tension between them turned out to be the whole project. One was screen automation — taking over the editor and clicking through it. The other was a scripting bridge over the engine’s own remote-execution protocol, written the same morning.

The screen approach broke on its own terms within hours: it stops responding, the view does not move, the camera is unfathomable. But the fatal objection was not fragility.

We are stuck with what the agent can do. We never really explore what’s possible — all options come out of that same number of combinations.

The cyberpunk street, in one very long day

Around forty scripts, in an order you can read straight off the filenames: build the street, fix the street, rebuild it curved, make it cyberpunk, detail the buildings, a second detail pass, props, street furniture — bins, newspaper stands, stop signs, zebra crossings, hydrants. Then neon signage generated from scratch, then night, then blue hour, then sky, then three grading passes to land on “just about sunset, street lights just coming on”. Then rain textures, then a Niagara particle system. Then navigation, a walker, a camera move, and six render passes.

The geometry underneath all of it is six primitive shapes. No models, no asset packs. A whole noir street out of boxes, cylinders and light.

The method that produced the detail has a name now, though it did not then — comparison against reality, deliberately: the wires are straight, and in reality wires are tangled, they have smaller wires wrapped around them, they are never this straight, they sag in the middle.

Drag through the six render passes. The set never changes across them — what changes is where the camera stands, who is in the shot, and how hard it is raining:

Pass 1 — The empty street, wide

The finished set with nobody in it. The camera sits well back and takes in the whole block — magenta on the left, the diner sign on the right, everything already reflecting in the wet road.

Don’t jump the gun

The morning after the street, the work stopped: no scene, nothing built, because none of the specifications existed yet. What came out of that was a twenty-thousand-word operating contract, and it is the single most valuable artifact in the whole body of work — more than the street, more than the film.

The operator model. It describes the human as “merely a taste factor”, and takes that literally: no editing skill, no engine knowledge. You can run a command, paste an error, record your voice, judge an image, and pick one of three. Every interaction is designed inside that envelope. The corollary is the useful half — you can reliably detect that something is wrong but rarely say why, so producing the signal is your job and translating it into a rule is the machine’s.

The render ladder. A one-minute contact sheet for a whole video is the primary review surface. Cheap regeneration is what makes “none of these, do it again” a rule you actually keep rather than one you abandon under time pressure.

The closed rejection menu. Nine permitted complaints — too fast, too busy, can’t read it, doesn’t sound like me, looks generic, I don’t believe it, nothing happened, saw it coming, lost me at a timestamp. A closed vocabulary is what makes taste into data.

Audio is the master clock. Voice first, all timing derived from it — which converts pacing from an editing skill nobody has into a performance decision anyone can make at a microphone.

The slop feeling does not come from using AI. It comes from automating judgement.

Money behaves like a gas

The first episode takes a 2000 paper on kinetic exchange and builds it. Two thousand agents each start with exactly one hundred. Two are picked at random, their holdings pooled, the pool split at random. Nobody is smarter, nobody cheats, and the total is conserved — asserted in the code, not assumed.

It converges to the Boltzmann–Gibbs distribution: the same law that describes how energy distributes among molecules in a gas. Inequality rises from nothing but fair trades.

And then the break, which is the part that makes it honest rather than a party trick. The exponential fits roughly the bottom 90% of real income data. The top 1% follows a power law that random exchange cannot produce at any runtime. The physics explains the unremarkable part of the distribution and fails exactly where the argument actually is.

Humanoids 2050

A detour that became the only finished film. It takes a bank’s humanoid-robot projection and gives every number in it a scene: the hands alone at $9,315 against a human hand’s 27 degrees of freedom; sensors at 37% of the bill of materials; six models unveiled in 2022 against fifty-one in 2024; China at 5,688 patents in five years against America’s 1,483.

Three structural inventions hold it together. Every chapter opens on a question card, which doubles as a valid start point for a short-form cut. A two-tone system where warm is always cost, China, acceleration and the uncomfortable side of any comparison, and cool is cheap, America and the conservative number — the room itself is cool, so warm reads as intrusion. And one time axis: the entire video is a single element tree rendered as a pure function of time, so nothing mounts or unmounts at a scene boundary and editing one duration re-times everything downstream.

Every scene, in order. Watch the two-tone rule hold across all of them — cost, China and acceleration are always warm; the machine, America and the conservative number are always cool:

The voice problem, which is not solved

Several synthesis models were tried and rejected in the same words each time: shrill, mechanical, sounds like it came from a machine. A self-recorded take was rejected too, and then the honest objection arrived — the delivery is monotonous, not the timbre.

The response was not a better model. It was measurement. A delivery plan fits every line to its scene window and records, per line: the hard limit, the designed duration, the word count, the measured duration, the resulting words per minute, whether it fits, a speed multiplier, padding, and a shaped rewrite. Alongside it, a cue sheet with continuous room tone across the full three minutes, ticks while a counter runs, and an impact when a reveal lands.

Nine voices were auditioned as an options system. Nine builds shipped, including one in Nandit’s own voice.

The open ledger

Several directories the contract refers to are still empty. Three documents it cites do not exist. The scene-kit module is a zero-byte file. The style decision is still open, which is precisely why the only renderer is named scratch. And the Unreal craft work and the channel system have never actually met — the first episode is drawn with a plotting library, not the engine.