Day 16: What happens when an AI film crew's engines stop with the chat window? One clip, 14 edits and a voice on the wrong model

Day 15 was about checks that said no. Day 16 was about a crew that kept working with half its machines switched off.

Day 16 at a glance

Claude runs the crew on my film. Two of its machines make new pictures. A runner sends prompts to Midjourney, one at a time. A render queue sends the frames I approve to rented GPUs to be animated. A third machine, the edit room, re-cuts one scene every hour or two. Four critics score each cut before and after: picture, pacing, sound and lip sync.

On October 19 the first two ran until about 6:40 a.m., then stopped for 13 and a half hours. The edit room never stopped. All times here are Eastern.

How did the Midjourney runner go almost 13 hours without drawing?

Day 15 ended with the runner stuck on one prompt. It was Mike on the roof in Scene 18. I had rejected takes of him crying, so the prompt was rewritten. The no-tears gate kept parking it. The runner put it back at the front of the line, and the gate parked it again. Each time, the runner counted the park as a job sent.

From 1:36 p.m. on October 18 to 1:23 a.m. on October 19, it parked that one prompt 30 times. Six times it hit its own limit of six jobs an hour, and waited. It drew nothing. The last real job had finished at 12:52 p.m. on the 18th.

The gate had already been fixed on disk. The runner was still using the old copy it loaded when it started. Claude fixed the runner at 1:33 a.m.:

• a parked prompt can't come back until its text changes,
• the hourly cap counts only jobs that reach Midjourney,
• the runner restarts after any change to a gate.

At 1:37 a.m. the Mike prompt went out for real. Four takes came back. Claude's pre-judge passed three. In the fourth, Mike faced away from the Spire window he should be looking at. I approved take 25 at 11:39 p.m., and its clip the next morning.

Why was there a second Mike in the chair?

Shot 11-31 is Mike alone in the Dream Lab, looking down at his friend in a chair below the frame. Only Mike is in the shot. Claude's no-tears rewrite added the words "his friend in the chair." It sent Midjourney pictures of Mike only. At 2:24 a.m. all four takes had a second Mike sitting in the chair, and Claude's pre-judge failed all four. I never saw them.

Claude fixed the prompt, and then the runner undid the fix. Two tools were saving the same queue file. The runner's next save put the old prompt back. Now rewrites go to an inbox that only the runner reads.

Two fixes later, at 2:41 a.m., take 29 came back: Mike alone, dry eyes, jaw set. I kept it, and approved it that night. A new check now refuses any prompt that names a character in the frame without that character's locked reference.

Why did the engines stop at 6:40 a.m.?

Claude had started the runner and the render queue from inside one of its chat sessions. When that session ended, they ended too. The runner took a planned break at 6:05 a.m. and never came back. The render queue logged its last check at 6:39 a.m.

The edit room noticed at 7:28 a.m. Its health check said both engines were down. It didn't restart them, because its brief forbids it to start Midjourney or rented GPUs. So it wrote a note asking for a restart. It wrote it again in the next eight runs. The hours kept climbing: no new clip for 5.6 hours, then 9.7, 12.7, 15.3, 17.7. Nine polite notes from a crew member who isn't allowed to touch the machine.

The engines came back at 8:08 p.m. Since then the runners start in their own session, so closing a chat can't stop them.

What did the GPUs make?

Not much, and the outage was only part of it. Before 6:40 a.m. the render queue looked for work every three minutes and found none. Every approved frame was already animated. The rest were waiting for a new prompt, or were fight shots meant for Runway.

The one RunPod job ran at 1:44 a.m. It was a new take of Waters spreading his arms in welcome in Scene 11. I'd rejected the last one for its animation on the 18th. One A100 for seven minutes, about $0.15, one clip.

Runway delivered six 15-second takes between 3:07 a.m. and 10:57 p.m., one of them a second try at the same Waters shot. Just after midnight I approved all six Runway takes and rejected the RunPod one.

What did the edit room find in 14 runs?

Fourteen runs started on October 19, one scene each. The numbers are the critics' average score out of 10, before and after. The best finds:

• Scene 11 (1:28 a.m.): 5.9 to 6.3. John's scream played silent, and his ragged breaths shrank to a gasp. A sound filter was fading them out before they started.
• Scene 10 (3:28 a.m.): 4.8 to 5.9. For six days Lilith said an old draft of her line. Mike's "Agreed" came from a file we had replaced. The line check only looked at file names. Now it checks where each voice came from.
• Scene 06 (7:28 a.m.): 4.8 to 6.0. More on this one below.
• Scene 09 (8:28 a.m.): 6.3 to 7.0, the best score of the day. A dissolve into a placeholder card printed "09-11 NEEDS PICTURE" across Lucy's face for half a second. Now it dips to black. The speech checker also heard a phone cord creak as a mumbled "I'm sorry."
• Scene 05 (2:29 p.m.): 5.0 to 6.3. A worker's line, "I can still work!", had gone silent in the 12:29 p.m. cut. Now any silent voice piece fails the check.
• Scene 06 (4:29 p.m.): 5.6 to 6.1. The worker's plea had been recorded with a sobbing tag. Men in this film don't cry. The sob is cut out.
• Scene 02 (5:29 p.m.): 6.1 to 6.5. A 12-second sound file put three extra mug clinks under Lucy's answer. One clink now.
• Scene 10 (8:29 p.m.): 5.9 to 6.4. This is where Slade's voice turned up.

Of the other six runs, five gained between 0.1 and 1.2 points. One scored lower, so it threw its own work away and went back.

How did my pick end up as a slideshow?

On October 17 I added a five-second take to Scene 06. John stands at the curb outside the Paradise, with Mike in the window behind him. Then we see John's face as he decides not to cross. But the cut kept playing an older copy. It was made on October 12 from 15 drawings of that take, stretched 4.8 times. For two days half my take never played. The rest crept along like a slideshow.

The 7:28 a.m. run found it. Now the whole take plays at real speed, and a new check flags any stretched copy of a clip I picked.

How did a voice from the wrong model last 15 days?

Slade stands at the bar with a radio on his lapel and reports, low: "They're going into the Dream Lab. Stay on them." My rule is that every voice in the film comes from the newest ElevenLabs model, eleven_v3.

Slade's line was made on October 5 with an older model. It had no record file. So the line check couldn't tell. The edit notes even called it the approved v3 take. The lip sync critic in the 8:29 p.m. run caught it, and the voice history confirmed it.

Claude recorded three new takes on eleven_v3, low and cold, into the radio. Take 2 fits his mouth with no stretching. It starts one frame after his lips part. The old file is archived. The run also listed every other voice that has no record file, so each one can be checked against the history.

What did I judge on October 19?

I wasn't at the review page much. Two moments:

• 8:08 p.m., on my phone: five clips rejected, each with the same two-word note, "new frame." One was Scene 10's walk past the vices. The 8:29 p.m. edit run took it out of the cut and sent a new prompt for that shot to Midjourney.
• 11:39 p.m.: 54 frame verdicts saved within four seconds. I approved 41 and rejected 13. Seventeen of the approvals were Scene 22. The other takes of each approved shot, 81 frames, were hidden so they won't come back.

One approval had a note. In Scene 60, Victor and John face off inside a mirror chamber. I asked for John to read bigger and stronger than Victor. The animation queue holds that frame until it's fixed.

What did Midjourney do?

The runner sent 32 jobs: 21 between 1:37 and 6:04 a.m., and 11 between 10:05 and 11:58 p.m. Claude's pre-judge ran 120 checks, and 82 passed.

• Scene 60-11, the mirror chamber. At 4:40 a.m. all four takes failed. John had black slicked hair, Victor had no silver temples, and the men had real bodies where the script wants reflections. The new prompt went out at 5:54 a.m. All four passed. Take 26 is the one I approved with the note.
• Scene 10-08, the walk past the vices. The magician at the billiard table has a "bald crown ringed by gray curls." A little chrome robot beside him fans a deck of black kings. At 10:10 p.m. Midjourney put a real crown on his head, on all four takes. Claude's second rewrite, at 10:48 p.m., says plainly: nothing on his head. In take 44 the robot holds up a hand of kings. The magician sits at the edge of the frame, looking down. No crown. I approved it at 11:39 p.m.
• Scene 55-02. Three of four takes drew a second Otis, white beard and cap, standing where John should be.
• Scene 62-09. The script wants a fist crushing a chess king. Three takes rested the fist politely beside it.

What got posted?

Three X posts from Day 9 went out, at 10:09 a.m., 12:07 p.m. and 10:03 p.m. They showed John lacing a boot with no foot in it, the founder saying the narrator's line, and the spy hoisted twice. At 9:09 a.m. the weekly review found 12 queued posts that still talked down to my characters. My rule against that was two days old. The post check now blocks them.

The numbers for October 19

• My verdicts: 59 frames (41 approved, 18 rejected), in two moments, 8:08 p.m. and 11:39 p.m.
• RunPod: 1 pod, 1 clip, about $0.15.
• Runway: 6 takes of 15 seconds.
• Midjourney runner: 32 jobs. Pre-judge: 120 checks, 82 passed.
• Engines down: about 6:40 a.m. to 8:08 p.m.
• Edit room: 14 runs. 13 improved their scene, 1 reverted.
• Voice lines fixed that had been wrong for days: 5 (Lilith's wording, Mike's "Agreed", John's pitch, the sobbing plea, Slade). Two more went silent in a single render and were caught the same day.

So what did Day 16 teach me?

A machine that only runs while someone is talking to it will stop when the talking stops. That was true of the engines, and it was the fix.

The edit room did its best work on the quietest day. No new footage came in, so it spent 14 runs listening to the cut it had. It found a silent scream, a stretched pick and a voice from the wrong model. Follow the run on the Making Death Match series page.

FAQ

Why did an AI film crew stop making images for 13 hours?
Claude had started the Midjourney runner and the render queue from inside a chat session. When the session ended, at about 6:40 a.m. on October 19, they stopped too. They came back at 8:08 p.m. Now they start in their own session.

Why didn't the AI editor restart them?
Its brief forbids starting Midjourney or rented GPUs. It flagged the stopped engines in nine runs in a row and left the restart to a production session.

How did a wrong voice model stay in the film for 15 days?
The file had no record of which model made it. The line check only compared file names. An edit note called it the approved take. A lip sync critic caught it on October 19, and the line was re-recorded on eleven_v3.

How much did the GPUs cost on October 19?
About $0.15 on RunPod for one clip. Runway delivered six 15-second takes on top of that.

While you wait for the next update: play Halfsies

Halfsies is our iPhone game, live now. One cut. Two halves. Each day you get three objects and one attempt at each: cut it exactly in half by eye and the scale tells you how close you got. No restart needed. Download Halfsies on the App Store