Quick summary:
- Independent filmmaker Jenny Krakovsky wrote, directed, scored, and produced "Boss," a feature-length AI film set in 1960s Liverpool, in a 30-day public sprint
- Style consistency now takes a three-line prompt and a contact sheet, while human performance and lens optics remain the technology's biggest gaps
- Vlogging the build drew hundreds of VCs, producers, and streaming services to her DMs before the film was finished
During a London heat wave, Jenny Krakovsky stood in her bedroom and filmed herself bouncing around like a maniac. She'd written a garage-rock song about a dog for a club scene in her film, and the AI-generated singer kept delivering it flat. She wanted Iggy Pop energy. So, she performed the song herself, sweaty and self-conscious, and fed the clip to the model as a reference.
That's what making a feature film alone looks like right now. Krakovsky spent 10 years directing commercials and documentaries before setting herself a challenge: make "Boss," a gangster comedy about 1960s Liverpool housewives who take over their husbands' crime business, in 30 days, using only AI tools. The script had been sitting on her computer for years, waiting on financing she knew wasn't coming. She decided to shoot it with AI instead and document the whole build in public.
On this episode of "The Disruption Is Now," Krakovsky walks host Greg Matusky through what the tools can do, where they still fall short, and why the smartest move she made had nothing to do with prompting.
Watch now:
Key takeaways
Launching early beat waiting for the technology to be ready
Krakovsky tested AI video models for three or four years and kept walking away frustrated. The tools weren't there. She suspected plenty of other filmmakers were doing the same thing, and once she launched her challenge, she found out she was right.
"I think it's probably the case that a lot of filmmakers were sort of like waiting in the wings," she says. "I knew that I would be using the technology at a place where it's not quite there. It's at like 11 o'clock, right? It's like an hour to midnight. It's almost there."
Going early carries real risk. "As a filmmaker, you're only as good in real life as the talent that you work with," Krakovsky explains. "If you work with terrible actors or terrible cinematographers, you're going to look terrible." The same logic applies to models that render wooden faces and warped hands.
But she ran the numbers the other way, too. Wait until the tools are polished and the experiment stops being interesting. Nobody cares about a sprint anyone could run. Her window was narrow, and she took it.
Three lines of prompt is all that’s needed for style consistency
When Krakovsky started with AI video, prompts sprawled into essays for every shot. That era is over. Her entire visual world runs on one long sentence covering her camera, her lens, the 1960s Liverpool setting, and a documentary aesthetic, paired with a contact sheet of reference images she points each generation toward.
"I give it a few lines and because it knows it's 1960s Liverpool, it does all the rest by itself," she says. Across a full feature, she's had to remind the model of the period maybe once.
Sometimes the model goes further than the prompt. The script called for washing lines strung between houses during the film's climactic car chase, and every one of the roughly 20 generations she ran for that scene included the laundry. Then came a final aerial shot she used as a bookend. She never mentioned washing. The model draped the backyards with it anyways because that's what terraced Liverpool neighborhoods looked like.
Human performance is the last thing AI can't deliver
Ask Krakovsky where the technology fails, and she doesn’t hesitate. "One of the reasons we all love to watch movies especially is human performance. We want to see the best actors giving their best performances," she says. "AI is just not a place where it can replace that yet."
She's pragmatic about the tradeoff. As an early-career filmmaker, she was never going to land big talent giving career-best performances anyway.
The other gap is glass. Krakovsky loves anamorphic lenses, which produce oval-shaped bokeh and barrel distortion at the edge of the frame. AI can bend the frame convincingly. It cannot fake the bokeh because it can't yet emulate how light moves through a lens. She expects that to change fast, noting that the way light bounces off surfaces and materials has improved 200% just in the time she's been using the tools.
Until then, she’ll just have to make the fixes herself.
The audience knows it's AI and that makes the comedy work
"Boss" plays its gangster tropes straight, then undercuts them. The car chases are fast and serious until the getaway car plows through a washing line and ends up draped in underwear. Housewives run the crime syndicate. Everything cool gets a ridiculous detail.
Krakovsky argues the AI medium makes that register land harder. "You’re kind of having a wink and a nod to the audience," she says. "They know it’s AI, you know it’s AI, and yet you can still enjoy that scene and the comedy value that it presents."
Matusky compared the effect to what Rick Rubin loves about professional wrestling. The audience is as inside the performance as the performers, and nobody breaks character. Watching "Boss," he felt in on it. The stiffness that creeps into AI-generated humans read as a note on where the technology stands, and the campiness invited him deeper into the film rather than pushing him out.
The vlog did more for the film than any prompt
Krakovsky decided from the start that the vlog was part of the project. Every stage of the build went up on YouTube, Instagram, and X, and that decision did more for the film’s prospects than anything she typed into a prompt box.
"If I’d have just built this film in my bedroom alone, then nobody would have cared at all," she says. "The only reason anybody cared is because I said I would vlog it as I go."
The audience that showed up surprised her. Hundreds of people landed in her DMs, including venture capitalists, producers, and streaming services. The content was the marketing strategy, and it worked before the film even existed.
The vlog also revealed how peers were tracking the tools but waiting for someone else to go first. "Basically, there are thousands of mes sitting there kind of watching from the sidelines," she says. "It's quite a big risk for your reputation. It's quite a big risk for your ego. It's a humbling endeavor."
Key moments
- How a commercials director first spotted AI's potential in Midjourney (1:57)
- Filmmakers waiting in the wings and the case for launching at 11 o'clock (5:48)
- The three-line style prompt and contact sheet that keep 1960s Liverpool consistent (8:12)
- The two things AI video can't do yet (9:00)
- Why satire works when everyone knows it's AI (12:24)
- What Rick Rubin's love of pro wrestling says about watching AI films (14:24)
- Scoring the film with AI music (20:31)
- How to sync a singer to a finished track (22:42)
- Director days, producer days, and a DM inbox full of VCs (25:46)
- Release plans and the buyers circling an unfinished film (29:13)
Q&A with Jenny Krakovsky, filmmaker and director of "Boss"
Q: When did you start working AI into your professional workflow?
A: Initially, especially as a serious filmmaker who does it professionally, I could see that the video models were a long, long way off, but I could tell from some of the image models like Midjourney that there was a lot of potential here, even just for things like doing treatments or mock-up videos, images for imagery search.
Once the video models became viable, I'd play with them, but I would get frustrated because I could see they weren't quite there yet. Then I'd say in the last year, the technology got to a point with tools like Runway and Higgsfield and other platforms that I thought this is almost there.
Q: How did you pick which script to make with AI?
A: If you’re going to make a first-time feature film, you're probably going to have to really restrict yourself in terms of locations and talent and things like that. This particular film is a period piece, and as a first-time feature filmmaker, you'd never be allowed, unless you had very, very wealthy family members, to make that kind of film as your first feature.
So, I thought, well, this is an interesting experiment. I can't shoot this script where I am. Why don't I try and do it with AI?
Q: How polished is AI footage straight out of the model?
A: It depends on the scene, really.
For the car chase, I ran about 20 generations and picked the best ones. But there are always things you need to adjust regardless. For instance, a character might end up standing in the wrong spot in a phone box, or a detail will drift between shots. I have plenty to fix at the moment. That's the reality of where the tools are right now.
