Finished With AI: Albert Bozesan and Robert Sladeczek on Making the AI Sitcom TARMAC

Nuclear waste, one week at a time.

Andrew had not seen TARMAC yet, so his reaction was real. Albert and Robert had asked to focus on their latest piece rather than a history of their work, so this session stays on one show. Andrew asked for the pitch.

Americans have run secret bases in Germany for decades, and Albert and Robert drive past them near Munich. Nobody will tell them what goes on at Ramstein Air Base, so they decided: in the 1970s the base figured out time travel, and Jimmy Carter used the very expensive time machine to get rid of nuclear waste by sending it one week into the future, forever. The team can never be fired, or fifty years of waste lands at once. The waste problem is always a Friday. The rest of the week is up to them.

The idea sat in a drawer for a long time after a late-night writing weekend. In the meantime they got more budget-conscious office sitcoms off the ground as fiction podcasts, including a startup comedy in the vein of Silicon Valley and The Office for one of Germany's bigger networks. The sci-fi idea needed expensive sets or animation. AI tools, and their friend the illustrator Ariane Emmerich, finally made it possible.

Audio plays shaped how they work. Without pictures, a joke has to land on timing alone, and a whole season has to be recorded in a couple of days. That made them impatient with slow production, and AI gave them the same speed for visuals. Chris and Andrew agreed that this kind of industry experience is what separates the best AI work from the rest.

“You just can't fake these things. You can't growth hack these things. There's no course that can teach you.”

Chris Branch▶ 8:43

Blender, Logic and a lightsaber video.

Chris described the pain of learning things the slow way, only to do them in two hours now. Albert pushed back on the idea that those years were wasted. TARMAC brought back the tools they had learned to love, Blender above all, along with the recording techniques from their audio work. They just stopped using them for the slow parts AI can now speed up.

“There's a treasure trove of stuff that people coming from AI should still be learning. Blender, absolutely still a valuable skill set.”

Albert Bozesan▶ 9:27

Robert added Logic, which set off Chris, who taught Logic Pro privately for six years. The week before, at a Dreamina event in London, he had watched Blender scenes used as driving video to control Seedance 2.5. The tools people said AI would make irrelevant are becoming the way to control it.

Their first editing programs explain the partnership. Robert learned on Final Cut Express from YouTube tutorials, keyframing himself onto his own hand because the software couldn't track. He and Albert had been in the same class at middle school and didn't care much for each other, until Albert saw one of Robert's videos.

“I saw his lightsaber video on Facebook and I was just like, I can do that.”

Albert Bozesan▶ 13:06

They kept pushing each other until, as Albert put it, they reached a level of skill where they could no longer ignore that they should probably work together.

Rhymes with classified.

▶ Watch TARMAC - Ep 1: Pilot

The cold open sends a father and daughter racing out on a school night to meet a mothership, with cocoa for the aliens. Then the pilot drops underground at Ramstein, where the Colonel meets Paul, the new recruit: half-brother to one of the team, another secret president's kid who doesn't know which president, with a CV that includes field testing mosquito sprays and picking fonts for label makers.

Paul is shown the Time Spring, built in the mid 1970s, where a shipment of barrels arrives from the States every week and a big lever sends them barrelling into the future. The quantum physicist who pulls the lever explains that every ton sent forward earns a ton of time travel in the other direction. Weektons. It's why she has a Shadow Nobel.

Andrew could tell straight away that they are scriptwriters.

“Just the way the jokes were coming in, and the timing of the jokes, and when to place the jokes.”

Andrew Davis▶ 22:04

Then there was the look. Andrew assumed it must be Seedance 2.5. It was Seedance 2.0, at 720p. The trick, they said, is that practically every single shot is based on a still they put together in Blender.

So what did AI do?

▶ Watch TARMAC - Making an AI-Animated Show in 2026

Paul had found the pair's behind-the-scenes film on YouTube, and they suggested playing it before the deck. It asks the obvious question: if humans designed the characters, acted the voices, wrote the scripts and edited the whole thing, what did AI actually do?

The answer starts in Blender. Albert blocks out each environment with rough objects so the AI knows where the furniture, walls and characters are, which gives very high directorial control. To keep the clicks down on a project with this many parts, he built a custom browser plugin that reads in images from Blender, lets them drag and drop into Magnific, and glues together the prompt.

That plugin, they admitted, technically broke Magnific's terms of service, because at the time Magnific attached no metadata to downloads. They needed it: a distributor will want to know where every frame came from. Their downloader saves the prompt and the model with every file, and was built largely with Claude Code.

The voices were recorded in a single day, planned the way they used to plan audio plays. One actor at a time went into the booth while the others read back to them over headphones from the room next door, so every line is a reaction rather than a recitation. With a TV budget, they would record the whole ensemble together.

Pencil lines, on purpose.

They came to animation through animated ads for entertainment projects, and think it suits AI better than live action right now. Neither of them is a trained illustrator, so the animation expertise comes from Ariane Emmerich, who studied it.

Because they are strong in 3D, the sets are 3D and the characters 2D. Then they worked backwards. From experience, they knew which 2D styles Nano Banana Pro handled well and Seedance 2 animated well, and had Ariane design a handmade style to match. The outlines are deliberately rough and pencilly, because video models at the time couldn't keep clean outlines clean.

Chris called it the kind of decision only years of generating can teach, going back to the struggles of Midjourney 3 and 4, when, as one of the guests put it, you could spend 50 bucks on a clip. Then he put their production deck on screen.

“This is the first time we've been sharing this actual deck.”

Albert Bozesan▶ 37:19

Before this, it had only been shown internally, inside Albert's AI animation class at Curious Refuge. Chris also asked anyone watching with funding to look at the show: what they had achieved in their booth sessions was a showcase of what a real studio budget could do.

Five people who'd better see it.

The message they most wanted to send was for solo creators. A lot of people come to AI amazed that they can now do everything alone. TARMAC was built to keep the parts of filmmaking that spark joy: writing, riffing on the page, working out whether a line suits the actor, table reads, eating together.

Chris Auber, who voices Paul, is an American stand-up comedian working in Munich, where the English-speaking talent pool is excellent but small. He and Martina Peroni changed lines on the day and pushed back, which no AI voice will do. Chris pointed out that Gossip Goblin made the same choice in Sweden, recording English-speaking actors locally rather than using virtual voices.

There was a practical reason too. The pilot took a month, not an afternoon, and they had asked five people to help. Owing those people a finished result kept them going, in a way that only disappointing themselves never would.

With Ariane's designs, they asked a question most AI productions skip. Nano Banana Pro posed her characters into the stills that Seedance then animated, so Albert asked her directly whether drawing every pose was work she wanted to keep.

“Is the task of drawing all those individual poses, the stuff we think we want to do with AI, is that a task you want us to take away from you?”

Albert Bozesan▶ 43:41

Robert, who knows her from school, relayed her answer: keep the parts of animation that are fun to bring to life, and skip the long back-and-forth stretches of TV animation in favour of building the world. Special moments, like Paul's jaw dropping, would go back to her, because push a design too far and Nano Banana just goes anime. They want everyone on the team paid fairly, and future episodes will bring her back for much more detailed character sheets.

“There's the interesting parts of animation that still are fun to breathe life in.”

Robert Sladeczek▶ 44:08

Chris asked about the backlash from illustrators, 3D artists and VFX. One of the pair had just come back from SIGGRAPH in Los Angeles, where most talks were about AI and people were excited. Their view: the people online claiming to speak for every creative aren't at the conferences, and nobody significant in AI film is coming out of nowhere. They put themselves in the middle of the spectrum, and definitely more excited than not.

Closer to the hundred than to fifty.

Chris asked the question the deck raises: with human-made Blender files behind every shot, could TARMAC get full copyright?

“This is one of the most copyrightable AI things that I've ever seen.”

Chris Branch▶ 50:20

Their jobs put them in front of a lot of lawyers, with a lot of different opinions. It's a spectrum, they said, not 100%, and TARMAC is closer to the hundred than to fifty. Albert, who until very recently worked at the Munich studio Storybook Studios, had the same debate there with people who consulted on the EU AI Act. Even they concluded that someone just has to try it.

“Basically the recipe is you just try to put as much human touch in it as you possibly can, and that's what we did.”

Albert Bozesan▶ 51:25

Then Albert made the least glamorous point of the night, and the one he most wanted every viewer to hear.

“To anyone out there who wants to make their stuff copyrightable, step one is documenting every single asset.”

Albert Bozesan▶ 53:23

Every background and every asset for TARMAC sits in a clean file system, ready for a buyer. Keeping it tidy while being, in their words, comedy nutcases was hard. It helped that they had studied management and technology, a degree one of them thought was three wasted years until the computer science and GitHub basics turned out to come naturally.

A set you can shoot forever.

Robert brought up the Blender mockup Albert built for the Colonel's office. Its cameras sit where a multi-camera sitcom like The Big Bang Theory would put them, with clearly defined angles that repeat, and a separate tool they vibe coded for Blender pre-renders the rough compositions in layers.

Once a set exists, it can be shot from as many new angles as they like without generating anything. Just typing officer in an office, they said, deteriorates over time. Instead they add the detail and the Easter eggs in Photoshop, and use a few Nano Banana iterations to fix problems.

Mary Unfiltered asked whether they had planned every set for the season up front. They built only what the pilot needed. A pilot is the expensive episode, and they expect about 80% of episode two to happen in sets and with characters they already have. Family Guy-style cutaways will be the exception, each with its own set, and part two had one they didn't want to spoil.

A library of Paul.

Chris said he had not seen anyone show this level of control, continuity and documentation. Why share it at all? Because a lot of people got secretive about AI workflows early, and they believe the market is still being built. They can become competitors fighting over cinema screens in a couple of years. For now, everyone needs to reach the same base level together.

Then Robert shared something they learned along the way that isn't obvious: Nano Banana can only focus on so many things at once.

“Generating the characters within the scene gives you a less qualitative output than generating them separately on, say, a grey background and putting them back in.”

Robert Sladeczek▶ 1:01:50

The difference shows even at 720p, and it saves money.

“You save a ton of credits that way. Honestly, it's probably embarrassing to admit how little we spent on this episode.”

Albert Bozesan▶ 1:02:09

Both are in the Magnific Creative Partner programme, and the pilot used no more Seedance credits than that provides, because so many images could be reused. They now have a library of Paul in every pose and expression to drop into any new location, and can relight him in Blender without AI. Like a sitcom: reuse the sets, write good scripts and stay low budget. Albert teaches the approach in his AI animation course at Curious Refuge, and goes live with students about once a month.

When everything's marked, nothing is.

A viewer asked about the EU's newly announced rule that AI-generated deepfakes must be labelled. The guests are for transparency, and say good expert consultation means it barely touches entertainment: for a film, the label goes in the credits, not on screen. Deepfakes that look real and get posted online should be marked. A fictional product that everyone knows is fiction loses nothing.

Andrew's problem is where the line is. A slight AI edit to a photo of himself would need a label. The same photo changed completely in Photoshop wouldn't. And does the audience even care?

“I don't really care if it's AI generated or not. A good story is a good story.”

Andrew Davis▶ 1:05:48

The guests pointed to the retouching notices some countries require on ads: a good impulse, aimed at protecting teenagers from beauty standards, but now on every ad, so nobody notices. When everything's marked, nothing is marked. They like that the EU separates AI free, AI modified and AI generated, though nothing says exactly where generated begins. They would argue TARMAC is AI modified, because so much of it is real, but can see why someone would point to the final animated pixel and call it generated.

Then the comments. They have had to delete a lot of comments calling TARMAC slop. Critical comments get a reply and a real discussion. Someone who opens with slop is not going to change their mind.

The background joke really happened.

Part two lands Paul at the base after a flight he never planned to take, with a tour of Europe by coffee, and Agent Norbert Rowbird telling him this is his time to shine, dimly, underground. Paul's first idea for the time budget uses 0.05% of it, and the episode ends on an airman explaining what would happen if they ever crashed one.

Then the guests' connection dropped. While one of them rejoined on an iPhone, Andrew gave his verdict on the pilot.

“It ended in a, I want to know more. And I think that's so key.”

Andrew Davis▶ 1:13:34

Back on, Cori Campbell asked whether any characters are based on real people. Not characters, they said, but jokes. As comedy writers they keep a running list of things that happen in meetings, on the street and on the subway, and squeeze them into a scene the moment they fit.

In the background joke, the Colonel asks Paul about his background and Paul turns round to look behind him. It happened at a conference, when a non-native English speaker took the question to mean the wallpaper on his computer, which was XCOM. They never found out where he was from or what he studied. About half the people who see it call it one of the funniest things in the pilot. The other half don't get it. The moonlighting line came from a fancy dinner, where a father asked his son, about 14, whether he knew what moonlighting meant.

They have outlines for the whole season and are definitely making more, with episode two planned for the autumn, when it gets less hot. With each Seedance update, they said, it only gets easier.

Figure out distribution.

Laurence Quinn asked about refusals. The easiest fix is to remove whatever triggers the filter and put it in differently. Seedance wouldn't do the ass joke, so the line went through as sass, and the voice actor's original line was cut back in during the edit. The coffee joke didn't trigger anything.

When Seedance says no, they go somewhere else, and Kling is very reliable. For TARMAC the deciding factor is lip sync on 2D characters, and Seedance's lip sync is what unlocked the show. Albert also recommended Comfy Cloud, which he uses regularly for his Curious Refuge course and which TARMAC didn't need.

“It's a great way to use very high quality models, I don't want to say completely uncensored, but without having to be afraid of a lot of the filters that are on most of the commercial sites.”

Albert Bozesan▶ 1:25:38

Mary asked about localisation: the work Pixar does to make posters without text, or to swap the broccoli in Inside Out for a vegetable Japanese kids actually dislike. If international audiences wanted TARMAC, they would regenerate the shots. They would not auto-dub. Germany has a whole industry of actors who voice American stars, the Amazon film that was dubbed with AI was terrible in many languages, and the jokes don't survive translation. For the German version they would oversee the translation and rewrite the jokes themselves.

Their one ask of the AI space is distribution. Showrunner and Escape are working on it, but the talent in these communities needs to reach people because the work is good, not because it is or isn't AI. Chris said that is why AI LIVE exists. TARMAC's pilot is done, season one is outlined, and they are looking for a distribution partner.

One last detail, for the audio nerds. Chris brought up Kavan the Kid's line that sound is 90% of a film. TARMAC went further: it was made sound first, with the visuals put on top.