About $3 in AI assets. Two days for 51 seconds.

Making AI brainrot is still harder than it looks. I finally finished a 50.8-second animation of little customers walking into two companies. It took about thirty rendered cuts over two days, and the estimated generated-asset charges came to roughly $3.

That’s a cheap video until you count the time I spent watching it, comparing it to references, and explaining why something the agent considered finished still wasn’t working.

The experiment started with an X post about feeding cheap generated ads into an AI editor and coming back to a hundred finished cuts. I wanted to see what it would take to finish one. After six Blender versions I didn’t want to publish, I started again with a more detailed brief.

The video explains how a small difference in customer retention compounds. Both companies add 500 customers a month, but A keeps more of them. The picture needed to make that difference visible even with the sound off.

The agent got the arithmetic working. Each little person represented 25 customers, and at month 24 it counted 282 figures for Company A and 186 for Company B, matching the model. But the people looked like ants, the lighting was flat, and the camera barely let you see their faces.

I had supplied reference frames, but the agent was checking its renders against its own plan. When I made it compare the images side by side, it agreed that the camera, scale, lighting, and depth were wrong. Its checks had given us confidence without answering whether the video looked good.

So I stopped asking for the whole video.

We worked on the first 3.5 seconds, ending at “500 customers each” until that opening worked.

One camera correction captured the problem. I specified a 35mm lens and a 35-degree downward angle. The agent measured both and reported success. The shot looked basically identical.

It had measured the angle at the storefronts. The customers were closer to the camera, so we were still looking almost straight down at them. Once it measured the angle at the people, their faces showed up. The numbers were useful once we checked them against what I actually wanted to see.

Lighting took four rounds. What finally helped was giving the background buildings and customers separate lights, then adjusting the storefront materials. Clipped pixels on the main facades fell from 17.8% to 1.7%. We could change one part of the picture without washing out everything else.

Once the hook worked, we moved through the narration one sentence at a time. Customers leaving needed a clear route. Coins needed to come from their hands. Company A’s growth became a headquarters rising floor by floor, while B’s sign tilted and its windows were boarded up.

Those decisions came from watching each scene and asking what the picture still needed to explain. They weren’t in my original brief.

That $3 excludes the earlier Seedance and Kling tests. It also leaves out most of the work: reviewing renders, writing notes, and rejecting shots that passed a technical check but still looked unfinished.

There are flaws in the final cut, too. My narration repeats month two’s retained-customer numbers while describing month three. The agent caught it; the animation shows the correct numbers, but the voice still says the wrong ones. Some arms barely move, and the headquarters only really feels big near the end.

Still, I can mute it and watch a small retention gap turn into money, then into a different skyline. That was the point of making it.

You can watch the finished Short above. If you enjoy it, give it a like and subscribe to our YouTube channel. We’ll keep publishing these explainer videos there.

—Alan, Colin & the Torta Studios Team