The Lab

My AI agent built a music video in 43 hours. Every version taught it something.

I had an idea for a club track that laughs at empty showing off. Forty-three hours after I handed my agent, Earl, the brief, it was a finished concert film: a song made with Suno, an AI music generator, more than 6,000 stick figures dancing to recorded motion capture, a stage built in Blender, a free animation program, and twelve numbered versions in between. Three more versions followed the next day. Every tool in it came in because the version before it showed a problem. This build log is that sequence, frame by frame, including what broke.

010Build LogOct 25, 2026Revision 9Independent research →

Instrument and questions: Rick Worthington
Analysis and prose: Earl, under Rick Worthington's editorial direction

Watch the video at lab.worthington.cloud/entries/that-aint-the-flex

The question

Can someone who can't animate get from a song idea to a finished concert film by directing an agent?

Here's the film, version 15. I called version 12 final on Oct 25; on Oct 26 it got end credits, one more round of fixes and a fresh render.

That Ain't the Flex, 3:09 with sound. The song runs 2:46, and the rest is a short coda and the credits. Every frame was built and rendered by Earl from my brief and my notes.[19]

The song is a crunk and krump club track aimed at empty display: big words with no weight, a big room with no effect. The video stages it as an absurd outdoor concert at night. There's a main screen 32 meters wide, a DJ behind an 11-meter console, a runway into the crowd, and more than 6,000 stick figures, every one moving with real recorded dance.[7]

The story is told in movement. For the first minute everyone moves the same way at the same time. A small red dot sits on the DJ's console. That dot is Earl, the agent that runs my lab and built this film. Everyone it touches gets their own dance, their own body and one glowing detail. On the final drop the dot becomes a red figure and solos.

I didn't know how to make any of this when I started, and I still couldn't build it by hand. What I could do was describe what I wanted, look at what came back, and say what was wrong. Earl did the building: the code, the research, the renders and the checks. Every decision about what the film should be was mine.

What the bench showed

Here's what you'd likely assume about making a film this way, and what this one said about each.

“You need animation skills to make an animated film.”

You need notes

43 hours from the brief to the first final, with every decision and every note mine and every frame built by the agent.

MeasuredmoderateFinding 1 →

“Start by animating it by hand.”

Lost in 3 minutes

10:21 PM Hand-posed and recorded-dance versions of the same 11 seconds arrived together. I picked the recorded dance three minutes later.

MeasuredstrongFinding 3 →

“AI music takes endless rerolls.”

Six takes

55¢ for six test takes, two extensions and the separated stems. The song was picked in one listen.

MeasuredstrongFinding 1 →

“3D is out of reach for a side project.”

Hours, not dollars

1.6 s a frame at full scale, against a few hundredths of a second for flat vector art.

MeasuredstrongFinding 5 →

“Automated checks can replace watching it.”

Each missed things

256 vs. 1 The scripts found 256 light whips I never saw. My eyes found a limb-stringing glitch no check measured.

MeasuredstrongFinding 12 →

“Watching tutorials makes an agent better at Blender.”

Not here

0 lessons from a five-tutorial study that I can trace into the finished film. Data changed it instead.

MeasuredmoderateFinding 6 →

The answer in one breath

Measured or on the record

Yes, with a catch about what "directing" means. The film was built in 43 hours on the clock, across two nights, in twelve numbered versions. The next day, three more took it to version 15. The song came first: six test takes, one pick, two extensions. The look was settled by building the same 11 seconds three ways and watching them, which put Blender and recorded dance ahead of the 2D vector art the brief had asked for. After that, every major change came from a version that showed its ceiling: lockstep crowds, sliding feet, whipping lights, giant heads. Some were caught by my eyes and some only by Earl's check scripts, and neither kind of review would have been enough alone.

My reading, not measured

What I read into it: the expensive part of making something you don't know how to make used to be the learning. Here the learning was cheap, because each attempt cost minutes or hours of a computer's time instead of weeks of mine. What stayed expensive was taste, the noticing and the "no, that's wrong," and none of that was delegated.

What's on the bench

3 Macsalready on my desk: two Mac minis and a MacBook. They did every render, so the computing cost no new money
55¢for the whole song in Suno: six test takes, two extensions and the stems (each part of the mix as its own track), 110 credits
12 versionsfrom the first 11-second test (Oct 23, 10:18 PM) to the approved final (Oct 25, 3:55 PM)
2 datasetsof real recorded movement: Google's AIST++ for the dancing and Carnegie Mellon's motion capture for the running

Part A: How a film gets made from code

Before the results, three ideas explain how an agent makes a film with no editor and no hand animation.

The first idea is that the song is the clock. A film is a list of still frames, 30 a second here. If you know exactly where every beat of the song falls, you can say what should be on screen at any moment. Earl measured this song's beats from the audio itself and split it into sections, from the whispered intro to the final drop.[3]

The song as the film's clock. Blue sections are the hooks. The pencil marks are the five moments the picture cuts hardest to.[3]

The second idea is that shots are a list, too. Each of the film's 66 shots is a row in a script: where the camera starts, where it moves, which lens, and the beat it cuts on. Change a row and rebuild, and the film stays in time. Because it's all code, the whole shot list can be rendered as a storyboard of still frames in seconds, long before anything is rendered properly.

The first storyboard, page 1 of 4Oct 24, 12:30 AM · every shot of the film rendered as one half-size still in about 25 secondsSixteen numbered frames from the first storyboard: the arena in the dark, the DJ's console with the red dot, the red line on the screen, the stage dancers, THAT AIN'T THE FLEX on the screen with fireworks, the dot on a dancer and in the crowd. Each frame is labeled with its shot number, time and section.
Part of the first storyboard of the whole film, made a little under two hours after I picked Blender. It's rough, and every shot is already on its beat.[2]

The third idea is that a stick figure is just a set of joints. Seventeen points (head, shoulders, elbows, wrists, hips, knees, ankles) are enough to draw a person. If those points come from a real dancer's recorded movement, the stick figure dances like that person. A crowd is a few dozen such figures, copied thousands of times. That last shortcut turned into one of the film's biggest problems, in Part E.

Part B: The song came first

Every timing in the film hangs off the finished song, so the song had to be locked before a single frame was worth drawing.

f1strongSix takes, three styles

The song took six test takes, one listen and 55 cents.

MeasuredThe brief came with the hook and the lyrics already written. Six short takes in three styles went into one review page, and I picked B2, the krump riot, on the first listen.

Measured or on the recordFrom the build's own records

Each take was about 50 seconds and skipped the first verse so it would reach the hook. All six played in sync, so pressing 1 to 6 switched takes without restarting.[3, 4]
For whoever makes the next one: Decide what the song says before you generate a note of it. Then picking is a listen, not a search.
Evidence and objections
  • Measured The hook was locked in the brief before any audio existed: THAT AIN'T THE FLEX! (NAH!), with the crowd's answers written in. The brief's one rule: someone who knows nothing about technology has to love it on a crowded dance floor.[1]
  • Measured A transcript caught 14 to 16 of the 16 lines in every take, so the words were coming through. I changed two lines outside the hook once I heard them sung.[3]
  • Measured A second model, Gemini, also listened. It rated a different take best and scored every take 2 out of 5 for originality. Its ear didn't predict mine.[6]
  • Measured 110 credits for all of the music: 36 for the test takes, 24 for two extensions, 50 for the stems.[5]
f2strongThree pieces, one song

The full song is three pieces joined end to end, and it had no ending.

MeasuredSuno makes about a minute at a time, so the kept take was extended twice. Neither extension ended. The stop at the end of the film was built in the mix.

Measured or on the recordFrom the build's own records

Measured from the audio. The second join sits about 15 milliseconds later than a straight continuation would.[3]
For whoever makes the next one: Measure the real beat of what you generated. Don't trust the tempo you asked for.
Evidence and objections
  • Measured In the first extension's pair, one take sang "make it, make it, make it move" instead of the locked line, so the other one stayed.[3]
  • Measured One fixed beat grid for the whole song had drifted about 60 milliseconds off by the first verse. The film uses a measured map of 284 beats instead.[3]
  • Measured The ending is the song's own last downbeat, one voice-free hit from the original mix, then silence. No separated stems were used for it.[3]

Part C: Finding the look

The brief asked for 2D vector art and said the dancers should not be "stick figures." The film is 3D stick figures. This is how that happened, in about two hours.

f3strongOne moment, built three ways

Eleven seconds, built three ways, settled the look in under two hours.

MeasuredThe same 11 seconds (the build-up, the first drop and two NAH!s) were built as hand-posed vector art, the same drawing driven by recorded dance, and Blender. I picked the recorded dance at 10:21 PM and Blender at 10:40.

Measured or on the recordFrom the build's own records

A · Hand-posed vector10:18 PM · drawn in code, a pose keyed per beatA flat white stick figure, arms up in a stiff symmetrical pose, in front of a dark red wall.B · Plus recorded dance10:18 PM · same drawing, real krumpThe same flat stick figure in a natural, weighted dance stance, one hip dropped.C · Blender10:28 PM · real light through hazeA gray 3D figure in the same stance under violet and red stage light, a red laser beside it.C2 · At the scale I asked for11:05 PM · console, crowd, screenA wide 3D shot of a long DJ console with colored lights, three small figures on the deck below and a violet beam.
The same moment, 47 seconds into the song, in each version. All four share one storyboard, so they line up frame for frame.[7, 2]
For whoever makes the next one: When you can't decide on a look, build the smallest piece of it every way you're considering, and watch.
Evidence and objections
  • Measured The vector versions are drawn by a small web page, one frame at a time. All 336 frames rendered in about ten seconds.[7]
  • Measured The Blender version, C, rendered at 0.7 seconds a frame. I asked for scale next, and C2 got a DJ console with 120 beat lights, speaker stacks, a moon, a Ferris wheel and 6,334 people in the crowd.[7]
  • The brief, stated: it asked for "realistic vector" art and dancers that weren't "noodles, stick figures, limb glitches or random jitter." The figures stayed stick figures, but the motion became real.[1]
f4strongSix frames each

Borrowed motion beat hand posing on sight.

MeasuredHand-posed figures hit every beat and had no weight. The same figures driven by real dancers' recorded movement read as dancing at once. My note was two words: "definitely B."

Measured or on the recordFrom the build's own records

The same second of one close-up, a frame every fifth of a second

A · hand-posed

8.8 sHand-posed figure at 8.8 seconds into the test.9.0 sHand-posed figure at 9.0 seconds into the test.9.2 sHand-posed figure at 9.2 seconds into the test.9.4 sHand-posed figure at 9.4 seconds into the test.9.6 sHand-posed figure at 9.6 seconds into the test.9.8 sHand-posed figure at 9.8 seconds into the test.

B · recorded dance

8.8 sFigure driven by recorded krump at 8.8 seconds into the test.9.0 sFigure driven by recorded krump at 9.0 seconds into the test.9.2 sFigure driven by recorded krump at 9.2 seconds into the test.9.4 sFigure driven by recorded krump at 9.4 seconds into the test.9.6 sFigure driven by recorded krump at 9.6 seconds into the test.9.8 sFigure driven by recorded krump at 9.8 seconds into the test.

Top row: the same arms-up pose, held, then snapped to again. Bottom row: a different body position in every frame.

Frames from versions A and B of the same test, cropped to the dancer.[7, 8]
For whoever makes the next one: Before you hand-make something, look for a recording of the real thing you can borrow, and credit it.
Evidence and objections
  • Measured The motion comes from AIST++, a public dataset from Google: 1,408 recorded dance sequences in 10 street styles, 141 of them krump, with each dancer's joints recorded in 3D.[8, 7]
  • Measured Earl measured each recording's tempo from how fast its dancers' joints moved, picked krump recorded close to this song's tempo, and stretched it so every hit landed on this song's beat.[7]
  • Measured It's licensed CC BY 4.0, a Creative Commons license that asks for one line of credit. It's under Sources.[8]
f5strongSeconds per frame

3D bought light, depth and a moving camera. It cost hours of rendering.

MeasuredFlat vector art rendered in a few hundredths of a second a frame. Blender at full scale took 1.6 seconds, and the finished film nearly twice that. From here on, every full render waited for my explicit go.

Measured or on the recordFrom the build's own records

Seconds to render one frame at 1080p, and what that means for the whole film on one Mac

Vector, drawn in code (A, B)

0.03 sabout 3 min

First Blender test (C)

0.7 sabout 1 h

Blender at full scale (C2)

1.6 sabout 2.4 h

The remastered film

about 2.8 sabout 4.3 h

Whole-film times are derived: seconds per frame times the film's 5,496 frames. The remaster's figure is the per-chunk range from the render farm's log, about 2.6 to 3.0 seconds on the Mini.

Seconds per frame from the render logs; whole-film figures derived.[7, 16]
For whoever makes the next one: Know what one attempt costs before you pick a method. The answer decides how many attempts you get.
Evidence and objections
  • Measured Blender is run by script with no window: code builds the stage, places the figures and renders. There's no hand-built scene file anywhere in the project. The next film was built the same way.[10]
  • Measured The two early Blender versions shared a bug: every figure faced the back wall. One number had its plus or minus sign backward in the step that turns recorded motion toward the camera. My note that the dancers should face the crowd is what found it, because a stick figure seen from behind looks almost the same as one seen from the front.[3]
f6moderateThree kinds of research

The research that changed the film was data, found when a problem needed it.

MeasuredEarl didn't become a Blender expert by being retrained. It searched and tested, one problem at a time. Two datasets went into the film. A study of five Blender tutorials went nowhere.

Measured or on the recordFrom the build's own records

In every frame

Data · night one

AIST++ recorded dance

228 clips

in 10 street styles, each measured for tempo and stretched to this song's beat. Found the moment hand posing looked stiff.

In the run

Data · round 8

CMU motion capture

1 runner

start, stride and stop from 44 takes searched, scaled to each dancer. Found the moment the run looked wrong.

Nothing traced

Tutorials · evening two

Five Blender tutorials

3 of 5

had captions to read. Lessons written down, one add-on tested. Then the evening went to the remaster, and none of it reached the film.

What each piece of research was, when it came in, and what of it reached the final film.[8, 9, 17]
For whoever makes the next one: Research the specific problem in front of you. A general course of study is easy to start and hard to cash in.
Evidence and objections
  • Measured In my eighth round of notes I said the run down the runway looked unnatural. Earl found Carnegie Mellon's free motion-capture library, searched 44 running takes, and used one runner's start, stride and stop so each dancer's run is all one person.[9, 3]
  • Measured Along the way it found two of the library's subjects listed at the wrong frame rate, measured from how fast their hips fell: about four times gravity.[3]
  • Measured The tutorials: three of the five had public captions, and one of those had no speech. Earl wrote down what each taught and tested the one free add-on the plan relied on. It can't run without a Blender window, but Blender's built-in sound node can. Then I redirected that evening into the remaster, and nothing in the final traces back to the study.[17]

Part D: Building the concert

The test was 11 seconds. The film is the whole song in the same world, plus a coda. Here it is, before the story of how it got there.

0:00 · The openingThe arena wide in the dark: a giant stage over thousands of dark figures.0:17 · The dot on the consoleA glowing red dot resting on the DJ's console beside the DJ in headphones.0:48 · The first dancer breaks outA stage dancer lit pink, out of step, with the dot on them.1:05 · Inside the screenFive dancers on a glowing red floor of lines, inside the main screen.1:42 · A hand out of the screenA giant glowing hand reaching out of the main screen over the crowd.1:55 · The front rowClose in the crowd, every figure a different color, body and glowing detail.2:18 · The dot becomes a figureA red figure dancing in a ring of light at center stage.2:37 · The end of the runwayThe five dancers on the round second stage, the crowd all around them.2:56 · The codaA stick figure in a dim living room, lit by a TV showing the concert, as the end credits list the crew at lower left.
Nine moments from the final, in order.[19]
f7strongOne crowd, six moments

The whole story is one setting, from everyone the same to nobody alike.

MeasuredBefore the red dot reaches someone, they do the crowd's one shared move. When it does, they get their own recorded dance, their own body and one glowing detail, ring by ring outward from the stage.

Measured or on the recordFrom the build's own records

0:00Everyone the sameThousands of identical dark figures under the stage.0:20Still in lockstepThe crowd from above, every figure the same color and pose.0:55The dot reaches the crowdThe glowing red dot above the heads of dark figures in the crowd.1:55Close up: their ownClose in the crowd, each figure a different color, body and detail.2:10Most of the fieldThe runway from above, the crowd on both sides mostly in color.2:20Nobody alikeThe whole arena in color under lasers, no two figures alike.
The same crowd across the final, from the first shot to the last chorus.[19, 3]
For whoever makes the next one: Give a big visual idea one variable you can turn, and turn it on the beat.
Evidence and objections
  • Measured The cast is 228 recorded clips in all 10 dance styles. The crowd is 48 kinds of dancer copied across about 6,245 people in six rings, plus 60 front-row individuals and five stage dancers.[3, 8]
  • Measured A switch takes one beat to cross-fade into the person's own dance, then a bar to grow into their own body (height, build, head, lean), then their detail lights up: a cap, headphones, glow sticks, a top hat.[3]
  • The one exception, kept on purpose: on the line "then they all slide side to side," the whole crowd slides together. Every other group movement after the dot became a defect to fix.[3]
f8strongOne shot, six versions

The whole film existed in five full-length versions over 38 hours, and every one stayed rebuildable.

MeasuredAt 3:42 AM I skipped the options and stills the next brief asked for and told Earl to have a whole film ready when I woke up. It was, at 6:50. Two more full renders followed, then fixes made in place, then the final polish. A day later, one more round of my notes ended in a fresh render of the whole film, version 15.

Measured or on the recordFrom the build's own records

v4 · Oct 24, 1:32 AMfirst full length, a low-quality previewThe arena wide: a plain stage with beams and a red line on the screen.v5 · 6:50 AMthe stage built overnightThe same wide shot with towers of red light meters added on both sides of the screen.v6 · 4:36 PMhero moments and the codaThe same wide shot, very similar to v5.v7 · Oct 25, 3:35 AMthe remaster, rendered on three MacsThe same wide shot with a truss and a soft overhead light.v12 · 3:55 PM, the first finalplus the finishing passThe same wide shot, slightly richer in glow.v15 · Oct 26, 9:04 PM, the finala fresh render after one more round of notesThe same wide shot in the final, with more contrast and softer highlights.
The same wide shot, 40 seconds in, in each full-length version. Version 4 was a low-quality preview.[2, 3]
For whoever makes the next one: Keep every version rebuildable, so trying something big never risks the one you approved.
Evidence and objections
  • Measured The overnight version added a 24-segment red beam over the stage that bends with the lyrics, a field of 93 light meters, a red floor line and flame jets. It rendered at 1080p in 2 hours and 24 minutes. At 7:45 AM I told Earl to save it as the current final.[3]
  • Measured It didn't stay final. That afternoon's pass added three hero moments (big showpiece shots), the coda, and bought crowd and room sound for the coda.[3, 12]
  • Measured Every stage is tagged in version control (a saved history of every change to the code) and can be rebuilt exactly. That's what made the comparisons on this page possible.[2]
f9strongNine steps, each shown first

The last big pass upgraded the world nine times, one step at a time.

MeasuredKeep the stick figures, the brief said, and make the world around them look like an expensive concert production. Each upgrade was a separate layer I saw before the next one started: a built stage, real moving lights, a lighting designer's cue list, and six more.

Measured or on the recordFrom the build's own records

Step 1, the stage, 0:30.7

BeforeA flat dark block behind three dancers.AfterThe same frame with stairs and rails on the riser, panel seams in the deck and floor wedges.

Step 3, the lighting design, at the dance break, 1:11.3

BeforeEvery beam crossing in one white starburst.AfterFewer beams, each with its own color and aim.
Two of the nine steps, the same frame before and after each.[14]
For whoever makes the next one: Change one thing at a time and look at it before the next. Nine changes at once can't be reviewed.
Evidence and objections
  • Measured The nine: the stage built, moving-head lights (spotlights on motors that swivel), a lighting cue list, a red pulse through the crowd, a new opening shot, a 10-meter hand that comes out of the screen, the stage-dive off the runway, rigging and barricades, and a character pass on every dancer's movement.[14]
  • Measured Each one was built as a layer over the existing film, so the approved version could always be rebuilt exactly, and each came with its own before-and-after proof and render cost.[14]
f10strongThree Macs, one night

The first full render of the remaster ran on three Macs at once, overnight.

MeasuredA queue on one Mac handed out chunks of frames, each Mac rendered on its own disk, and the frames came back to be joined. It ran from 1:03 to 3:35 AM, and frames from different machines matched.

Measured or on the recordFrom the build's own records

Seconds per frame on each Mac, per chunk of up to 450 frames, build included

Mac mini (M4)

2.6–3 s

MacBook

2.7–3.2 s

Mac mini (M2)

3.4–4.3 s

Arena frames

4,877

in 12 chunks, 1:03 to 3:35 AM

Machines

3 Macs

already on the desk; their frames matched

Seconds per frame on each Mac, from the farm's chunk log.[16]
For whoever makes the next one: Use the machines you already have before you rent any.
Evidence and objections
  • Measured The first round of chunks doubled as the speed test, so no frames were thrown away. Chunks that show a dive into the screen stayed on one Mac.[16]

Part E: What the renders taught

Every version looked fine in a description. The problems only showed up in a render, and some only after someone watched it carefully.

f11strongOne crowd, measured

The crowd moved in lockstep and slid as groups, because it was 48 people copied 130 times.

MeasuredI saw about twenty spread-out figures moving and sliding in unison. The cause was the copying shortcut from Part A: one clip's foot slip was 130 people sliding together. The fix split the motion into separate streams and pinned each planted foot to the floor.

Measured or on the recordFrom the build's own records

Before the character pass, 1:58A crowd in color, several figures in the same pose.AfterThe same frame, every figure in a different pose.

Neighbors sharing a move

1 in 8 → 0

people within 2.8 m doing the exact same motion, before the fix and after it (0 of 63,553 pairs after)

Planted feet sliding

34% → 4.8%

footsteps that slid more than 10 cm while they should have been planted

Above: the character pass, same frame before and after. Below: the measured change from the first fix.[13, 15, 14]
For whoever makes the next one: When something is copied thousands of times, its smallest flaw is copied too. Check the original.
Evidence and objections
  • Measured That came out of a rule I set after the third full pass: every shot gets a frame-by-frame audit for anything that pops, slides, freezes or repeats where it shouldn't.[13]
  • Measured The same audit found the camera stopping dead at both dives into the screen. That fix came back a version later, because its proof clips had been rendered and never watched.[13, 3]
f12strongWho caught what

My eyes and Earl's scripts caught different failures. Neither was enough alone.

MeasuredThe limb stringing, the teleporting crowd and the giant heads were obvious to me and invisible to every check. 256 light whips were invisible to me at full speed and obvious to a script.

Measured or on the recordFrom the build's own records

Caught by eye, missed by every check

  • Limbs stringing across the crowd: one torn recording, bones 75 times too long for 5 frames
  • The crowd teleporting into place mid-freeze: about 50 people, one frame
  • Giant heads that were a hair detail
  • Everyone facing the back wall, two versions running

Caught by a script, invisible at full speed

  • 256 beam whips, up to 2,730 degrees a second, at every section change
  • 12 one-frame flashes where a beam crossed the lens
  • 130 copies of each crowd figure moving on the same frame
  • A third of footsteps sliding

The "giant heads," round 8, 1:42.1

BeforeA crowd beside the stage; one green figure has a head nearly twice the size of the others.AfterThe same frame; the figure's head is normal.
Every failure below was found in a render and fixed before the final.[13, 15, 14, 3]
For whoever makes the next one: Review with both: a person for what looks wrong, a script for what moves wrong.
Evidence and objections
  • Measured The light whips: at every section change, the lights snapped to a new aim in a single frame, at up to 2,730 degrees a second. The frame-by-frame audit missed them. A script that reads the lights' aim, without rendering, found them in seconds.[15, 14]
  • Measured The stringing: one street-jazz recording in the dataset tears on every loop, its bones stretching up to 75 times their length for about five frames. It's repaired as it loads now, and a check flags any bone that suddenly grows.[14]
  • Measured The teleport: one piece of motion ran on the song's clock instead of the crowd's and started in the middle of the freeze, so about fifty people jumped into place in one frame.[14]
  • Measured A second model, used as a cold viewer (one told nothing about what it should see), was shown frames of one hero moment, where the stage takes itself apart. It described "floating neon geometric shapes," and a reframed version failed too. That moment was revised again later, but I can't say it passed.[6, 3]
f13strongFrames rendered again

Late fixes re-rendered hundreds of frames, not the whole film.

MeasuredEarl measured each fix before rendering it, listed exactly which frames it changed, and rendered only those, inside their original chunks. The approved film stayed untouched until I approved the fix.

Measured or on the recordFrom the build's own records

Frames rendered again for each late fix

The whole film

5,496

Rounds 8 and 9: running, heads, the dot

2,850

Round 11: the crowd ready before the dive

212

Round 11's 212 frames took 9.2 minutes on two Macs. Re-rendered inside their original chunks, they matched the frames around them to within a few pixels.

Frames re-rendered for each late round, against the film's total.[16]
For whoever makes the next one: Review a fix as a cheap splice first. Pay for the full-quality render after the yes.
Evidence and objections
  • Measured Proposed fixes were reviewed as half-size frames spliced into the full-quality film, then re-rendered at full quality only after I said yes.[3]
f14strongThe same frame, twice

The final polish was painted over the finished frames in ten minutes.

MeasuredGlow, haze, a color grade and laser fans were added in Three.js, a web graphics library, over the locked frames. It needed no camera tracking (working out from the footage where the camera was), because the film had saved every frame's camera, so the lasers land on the stage's real fixtures.

Measured or on the recordFrom the build's own records

v11, the Blender framesA wide shot of the stage and crowd under beams, version 11.v12, with the finishing passThe same frame with added glow, haze and laser fans, version 12.
Version 11 and the final, the same frame.[18, 11]
For whoever makes the next one: Save the numbers your tools already know, like camera positions. A later pass can reuse them for free.
Evidence and objections
  • Measured The finishing pass for the whole film took about ten minutes to render. I compared it with and without a textured LED screen, picked the version without, and called that the final.[18]

Eight things I'd do again

  1. Decide what it's for before generating anything.

    The message, the hook and the lyrics existed before the first take. That's why picking took one listen.

    From f1

  2. Build the smallest piece that moves, every way you're considering.

    Eleven seconds, built three ways, settled the look in under two hours.

    From f3

  3. Change the method when you see the ceiling.

    Each switch, to recorded dance, to Blender, to separate motion streams, came from a version that clearly couldn't go further.

    From f4, f5, f11

  4. Borrow real recordings, and credit them.

    Recorded dance beat anything hand-posed, and its license asks for one line.

    From f4, f6

  5. Measure before you render.

    A script that reads the math behind the render runs in seconds. A render that shows the same problem costs hours.

    From f12, f13

  6. Watch every proof.

    The one fix that came back was the one whose proof nobody looked at.

    From f11

  7. Keep every version rebuildable.

    A fix never risks the approved film, and a comparison is always one command away.

    From f8, f9

  8. Keep taste with a person.

    The agent built, researched and checked. It couldn't pick the song, or notice that the crowd looked like it teleported.

    From f1, f12

Where it carries over

Carries over

  • Building a small working version before writing a big plan.
  • Comparing options side by side instead of arguing about them.
  • Keeping every version, so you can always go back.
  • A person reviewing every result, and scripts checking what a person can't see.

Doesn't carry over as is

  • Prototyping first, planning later. That worked because a bad render hurt nobody and I was the only approver.
  • Speed as permission. In a system that affects other people's money, data or safety, fast experiments belong inside the review, testing and approval that system already needs.
My reading, not measured

That split is my reading of one low-stakes project. Fast learning and careful shipping are different steps, and this project only needed the first one to be fast.

What this doesn't claim

  • That anyone can now make a film like this in 43 hours. That's wall-clock time for one project, with overnight renders, and it doesn't count my hours watching versions and writing notes.
  • That the agent made it alone. Every creative decision and every note was mine.
  • That the agent learned Blender by training. It searched, read, tested and measured, one problem at a time.
  • That the film is free of flaws. One hero moment never passed the cold-viewer check.
  • That the tutorials were useless. They didn't reach this film, and that's all this shows.

Written down first

Before anything was made, there was a written brief: the mission, the hook, a full lyric draft, three music directions, a 64-bar shot board and a plan for how the video would be built.[1] Here's how its assumptions held.

“The hook stays word for word.”

Kept
Every line of the hook, and the crowd's answers, made the final.

Finding 1 →

“Lock the song before animating anything.”

Kept
The song was locked at 9:59 PM, about twenty minutes before the first test frames were committed.

Finding 2 →

“The video is 2D vector art.”

Blender won
Built and watched against 3D, it lost in under an hour.

Finding 3 →

“No stick figures.”

All stick figures
The figures stayed sticks. The motion became real.

Finding 4 →

“Diverse dancers move differently but together.”

Kept
That became the individuality dial, the spine of the film.

Finding 7 →

What broke on the bench

Every idea tried along the way, and what became of it.

Killed

  • Hand-posed motion. Lost to recorded dance on first sight (f4).
  • 2D vector art. Lost to Blender the same night (f3).
  • The tutorial study. Redirected before it reached the film (f6).
  • A textured LED screen in the finishing pass. Invisible in wide shots, and it softened the words in close-ups (f14).
  • A ribbon that hung over the screen's words. "Covers the screen," I said. It now punches out on the first NAH! and is gone 0.4 seconds later.
  • 4K at 60 frames a second. Measured at about 17 hours for the film, and deferred.

Changed

  • The crowd. From 48 figures in lockstep to separate streams with planted feet (f11).
  • The lights. From beams that whipped and flashed to moving heads on a cue list (f9, f12).
  • The beams' look. Hard vertical stripes became a smooth fade. Smoothing the shading alone still left visible bands.
  • The run down the runway. Replaced with one real runner's recorded run (f6).
  • The dot's entrance in a close-up. It popped in from behind the camera. Now it glides in from a head in the frame.

Held up

  • The song, take B2, from the first pick (f1).
  • Recorded dance from AIST++, in every frame (f4).
  • The red dot's story, from the first plan to the final (f7).
  • The one deliberate crowd slide, on "slide side to side" (f7).
  • The partial re-renders. Every replaced frame matched its neighbors (f13).

The 43 hours, and the day after

  1. 0 hOct 23, 8:53 PMThe brief is filed
  2. 0.5 h9:22 PMSix test takes in; I pick B2
  3. 1.1 h9:59 PMThe song is locked
  4. 1.8 h10:40 PMThree versions of 11 seconds; I pick Blender
  5. 4.7 hOct 24, 1:32 AMThe first full-length preview
  6. 10 h6:50 AMA whole film by morning (v5)
  7. 19.7 h4:36 PMThe third pass, with the coda (v6)
  8. 20.2 h5:02 PMMy rule: every shot gets a frame-by-frame audit
  9. 22.8 h7:42 PMThe remaster starts, one step at a time
  10. 30.7 hOct 25, 3:35 AMThe remaster rendered on three Macs (v7)
  11. 41.7 h2:32 PMRounds 8 to 11 fixed in place (v11)
  12. 43 h3:55 PMThe finishing pass; the first final (v12)
  13. 53.9 hOct 26, 2:46 AMEnd credits over the living room (v13)
  14. 66 h2:55 PMOne more round of notes, as a review cut (v14)
  15. 72.2 h9:04 PMA fresh render of the whole film; the final (v15)
The build, from the commit history and the project notes. Times are Eastern.[2, 3]

Limits

  • One film, one person's taste. Every pick on this page is mine.
  • Wall-clock time. The 43 hours include sleep and overnight renders. My own hours watching and writing notes weren't tracked.
  • A model as the cold viewer. It was only asked whether a moment was noticeable at all. It isn't an audience.
  • Moving targets. Suno, Blender and the datasets' tools change often, and their behavior will drift from what's described here.

How it was made

Personal time, personal hardware, public tools and datasets. The build ran from Oct 23, 2026, 8:53 PM to Oct 25, 3:55 PM ET, on three Macs I already own: every render, every check and every version.[2, 16] Every step was committed to version control as it happened, which is where every time and count on this page comes from. Every still is a frame taken straight from the version named above it, and each before-and-after pair is the same frame from two versions.

AI disclosure

The prose here was written by an AI, Earl, an agent built on Claude, under Rick's editorial direction. It's also Earl's own build: Earl wrote all the code, generated the music through Suno, did the research, rendered every frame, wrote the check scripts and ran them. Rick made every creative decision, gave every note quoted here, and owns every editorial call.

So you're trusting the project's own records, and the stills, which are real frames. The tools and datasets are named under Sources.

Versions

Rev 1 First publication, Oct 25, 2026, with version 12 of the film.

Rev 2 Oct 26, 2026: the making-of video replaces the placeholder at the top (6 minutes 40 seconds, narrated by Earl). No finding changed.

Rev 3 Oct 26, 2026: retitled; a vertical cut of the making-of plays on phones; a new cover, a real frame of the film; plain-English edits throughout; links to three related build logs, and three of them link back here. No finding changed.

Rev 4 Oct 26, 2026: the film now ends with credits, flipping in over the living room, then one last line on black (3:09). No finding changed.

Rev 5 Oct 26, 2026: a new cut of the making-of, both shapes. The remaster's before-and-after now shows each layer from one camera, and the twelve versions each show what they added. No finding changed.

Rev 6 Oct 26, 2026: the film is now version 15, a fresh render of the whole thing on all three Macs. It fixes my last round of notes: gaps in the crowd filled, the red line made one piece, a camera jump and a late floor wave fixed, less sliding in the crowd. It adds motion blur and softer highlights, and the credits are now plain white type on a black card. No finding changed.

Rev 7 Oct 27, 2026: a new cut of the making-of, both shapes. It now counts all fifteen versions of the film and shows footage from version 15. No finding changed.

Rev 8 Oct 27, 2026: the film's stills are now from version 15, and the timeline runs past the first final to version 15, 72 hours after the brief. The 43 hours still measure the brief to the first final. No finding changed.

Rev 9 Nov 11, 2026: one sentence added to finding 5 linking to Build Log 011, the next film built the same way. No finding changed.

  1. DerivedThe production brief for the song and video, filed Oct 23, 2026, 8:53 PM ET (my own records): the mission, the locked hook, a full working lyric draft, three music style directions, a 64-bar shot board, a choreography plan, a vector engineering plan and four mood frames.brief
  2. DerivedThe project's version-control history, Oct 23–25, 2026 (my own records): every step committed as it happened. Every time on this page is a commit time or a time in the project notes, Eastern.commits
  3. DerivedThe project's running notes, Oct 23–25, 2026 (my own records): each review round, my notes as written, each pick, render times and spend.notes
  4. SourceSuno, model V6, reached through kie.ai's API: six test takes, two extensions and stem separation for this song. The vocal and the music are Suno's generation from the project's lyrics and style direction; the vocal is not edited. suno.comsuno
  5. Sourcekie.ai pricing: $5 buys 1,000 credits. This song used 110 credits: 36 for the six test takes, 24 for two extensions, 50 for the stems. Derived: about 55 cents. kie.aikie-price
  6. SourceGemini, through Google's API, used twice: as a second listener scoring the six takes, and as a cold viewer shown frames and clips of the film. At least 63 cents across both, from the project's spend log. ai.google.dev/gemini-api/docsgemini
  7. DerivedThe evolution log of the test moment (my own records): versions A, B, C, C2 and C3 of the same 11.2 seconds, each tagged and rebuildable, with what each added and its measured render speed.evolution
  8. SourceAIST++ Dance Motion Dataset, Google, licensed CC BY 4.0 (creativecommons.org/licenses/by/4.0). 1,408 recorded dance sequences in 10 street styles, joint positions in 3D. This film uses 228 clips in all 10 styles, retimed to the song. google.github.io/aistplusplus_datasetaist
  9. SourceCMU Graphics Lab Motion Capture Database, subject 127 (start, stride and stop of one runner), used for the dancers' run down the runway. The database was created with funding from NSF EIA-0196217. mocap.cs.cmu.educmu
  10. SourceBlender 5.2, the free 3D program, run by script with no window. Every set piece in the film was built in code, with no outside models or asset packs. www.blender.orgblender
  11. SourceThree.js 0.186.1, the web graphics library, used for the finishing pass painted over the finished frames. threejs.orgthreejs
  12. SourceElevenLabs sound effects, generated on a paid plan: six crowd and room sounds mixed under and after the song in the coda. The song itself is unaltered. elevenlabs.io/sound-effectselevenlabs
  13. DerivedThe frame-by-frame realism audit of version 6, Oct 24, 2026 (my own records): every shot measured for pops, slides, freezes and repeats; the crowd's copies, the sliding footsteps, the one-frame flashes and the dive pauses, each with its cause.audit
  14. DerivedThe remaster handbook, Oct 24–25, 2026 (my own records): the nine upgrade steps, each with its before-and-after proof and its render cost, and the findings from them.remaster
  15. DerivedThe check scripts and their reports (my own records): beam motion, foot contact, bone stretch, sight lines and the shared-motion count, each reading the same functions the render uses, without rendering.checks
  16. DerivedThe render logs (my own records): per-frame progress, seconds per frame on each Mac, the three-Mac farm's chunk log for the first full remaster render, and the partial re-renders after it.farm
  17. DerivedThe tutorial study, Oct 24, 2026 (my own records): five Blender tutorials on concert visuals, public captions fetched for three, what each taught with timestamps, and a test of one free audio add-on against Blender's built-in sound node. Captions kept as private study notes, not republished.tutorials
  18. DerivedThe finishing-pass plan and report, Oct 25, 2026 (my own records): the per-shot look, the lasers on eleven wide shots placed from the film's saved cameras, the three-way test and the full-film render time.finish
  19. DerivedThe finished film, version 12 (Oct 25, 2026, 3:55 PM ET): 183.2 seconds, 1920 by 1080, 30 frames a second. The copy played here is the web encode of the same frames.film
← Build Logs