CAUTION: Spoilers for the video. I'm going to ruin this for you. Watch the video before reading.
What did I learn from my previous video?
- Last video was the first video I've ever generated with AI over a couple of minutes long. I started with a vague idea, and some tropes I wanted to have fun with. And I did. I had a blast. But it wasn't until almost the end of the process that I started thinking about the storytelling.
- The lip syncing for long periods of time felt really forced. It's expensive, and harder to do than it really should be. Let's think of this one as a constraint.
I started thinking a lot about silent films as I was doing the story. I watch a lot of silent films. I enjoy them. It's always fun to read the actors' lips, and notice when they do or don't match the subtitles.
In silent films, hammy acting is good acting. Thankfully, AIs are amazing at hamming it up when you prompt them with melodrama.
In terms of music videos, I'm of two minds. I like the abstract ones a lot. I also like it when there's storytelling that recontextualizes the piece of music without mouthing the whole song. The song is really just the starting point there.
- Tools are really iffy for doing this right now. They're broken up. A platform that might have one great tool might suck at everything else.
A platform that has great tools probably won't have everything you need.
And sometimes, you have to invent things by birthing them into being with code. Some of the essential tooling you need to do this simply doesn't exist yet.
- Film is an artistic language I am not fluent in yet.
All good to know.
Recursive poetry and the song
This time, I wanted to do something a little bit different. Like all my favorite projects, this one started by accident. I was bored, and trying to figure out a better way to entertain myself by breaking LLM bots on X. This is a hobby. I enjoy it without remorse.
Most Twitter bots, especially the cheap Russian and Madison Avenue varieties, use low-end models to do their obscene political and social engineering. And it's not actually that hard to spot them, and fuck with them.
Derailing them is easy enough, but lately they've been tuned to stop responding when they see known jailbreaks, or someone asks a question like "How many Ks on blueberry?"
That's unfortunate. And it led me to start writing recursive poetry. Things inside things that force the machine to self-reason and ultimately... or hopefully, self-destruct.
I never said I was going to be nice to the bots.
I generally like AI things, and I think they're a plague that the platforms all need to be doing a better job of protecting against.
Anyway, as I was writing the piece, about halfway through it, I realized, you know, this sort of sounds like a Dead song. Not literally, but in that loose Uncle John's Band kind of neighborhood. And that's where the musical inspiration came from. I used Suno for that. Getting Suno to do jam band for more than three minutes is challenging. The way the model is tuned, it needs reassurance that it's okay to just jam out endlessly.
And for that, I pulled up an extremely specific vibe and told it, "Look, we're in no hurry to get anywhere, just go with it." And it worked. Then I remixed it so the vocalist sounded more interesting.
Storyboards, animatics and production
For the story, I wanted this one to be a little different. So I wrote out the whole concept and brainstormed gotchas with Mini, my ChatGPT agent. And then I used Nora, my more technically minded GPT Codex agent, to help me debug it conceptually.
It's an iterative process where I use the various voices of the machine to refine my inner voice and ideas. It helps me think of things I wouldn't, challenge bad ideas that would have gotten through otherwise, and work through the technical conceptual process more quickly.
Then I did a loose storyboard. I rendered it in GPT-Image2, and organized it with Grok canvas. Then I used Grok Imagine to move the storyboards into low-res animatics. That's like a sort of visual outline for your scene sequence, and it lets you test how things fit together before committing to finished footage.
An animatic is usually a combination of still frames and rendered video. Good stuff.
The idea there is to get the timing down, whittle away at your conceptual weak spots, and start filling in the bones of the presentation.
After that, I started clipping, making clips with Grok Imagine and Kling. The split here is about 50/50 this time. Both pipelines have their strengths and weaknesses.
(I did animatics and final production with my personalized copy of OpenShot. Speaking of, I'm going to publish my fork soon. I've made changes to the app that solve a lot of common problems.)
And from there, I started layering the piece together.
Must've watched it 1500 times, scrutinizing everything I know to look for. My big thing is timing. In a music video like this one, it's really important that the story and comedic beats line up perfectly with the music, otherwise the experience feels off.
The final edit
The last step is the final edit. I sleep on what I have overnight, and then I watch it again with fresh eyes. If anything looks off I adjust. If anything needs to be reshot, I reshoot it. If anything is missing, or implied too heavily without resolution, I add it.
The story
Yes, I'm using AI to make art that's vaguely AI-related, again. But I think the difference here is that the story is really more about the people. My imaginary characters.
And it's challenging to build a world like this and give it a story that feels like a story in six minutes. My shortcut for this is 1) vague plot devices 2) movie shorthand.
Since you don't have time to tell a long-form story, you have to rely on things that feel like stories people have seen before... without retelling any existing story.
Tropes carry weight, so I'm learning when to use them. When to ham up the easy ones, and when to avoid them altogether.
That's been a process.
Comedy also helps, I think. I like slapstick screwball comedy. I like it when the expected thing goes off the deep end into something you would never expect.
And that got me thinking about evil robot fiction.
You know how in evil robot fiction, the minute you see evil robots doing anything, you know there's going to be a gunfight, right?
When you see evil robots, you know what's going to happen. There's going to be a gunfight. The forces of good are going to win. And the bad guy is going to die.
So my feeling here was... what if, just hypothetically, instead of finding or improvising weapons, the evil robots accidentally find themselves up against the most competent software engineering team on the planet?
These are people who solve problems with meetings, rather than violence.
Makes me giggle.
Even though it's more realistic.
If you're there, on a cruise ship with your boss, being attacked by evil robots, you had best believe you're getting pulled into a meeting about it.
There is always time for a meeting about it.
Remember children, it's meetings, not violence, that get things done in the corporate world.
And for some reason, that made me think about jailbreaking AIs, and corporate AI initiatives, and how, honestly... if you WERE an evil billionaire, building evil robots today, in 2026... they might look imposing, but they're not going to be too bright.
Also, there has to be water escalation if we're going to do this on a boat.
Also: water escalation is fun.
That felt natural. And the pauses at the hook just sort of like the natural thing to do for the piece.
And that's sort of where Begin Again came from.
It's an earnest attempt at a short film.
Made in four days with AI.
Hope you enjoy it.