Creators Cheer Filler-Word Removal and Boo Generative Video

Nathan Labenz opens with an email Laura Burkhauser sent after becoming CEO of Descript, the video editor that launched in 2017 on the idea of cutting a film the way you edit a document. She had been VP of product for several years and took over from founder Andrew Mason. In the note she called generative AI a polarizing topic among creators and declared that Descript "isn't a slop machine and we don't want it to be." Labenz, a longtime paid customer and early tester of the Underlord API, uses the product to cut this very show. He asks the question the email begs: what is slop, and can you keep shipping generative features without surrendering to it?

Burkhauser does not treat slop as a synonym for AI-made junk. To her it is a form of content arbitrage: spotting a temporary hole in a platform's algorithm, pumping in cheap volume, and collecting enough ad revenue or engagement that the whole thing nets out positive even if nobody really watches. The two elements she will not drop are motive and scale. The incentive has to be money in some form, and the production has to be happening at volume. There is a lot of bad art in the world, she adds, and she is generally pro bad art. Bad art is a stage you pass through on the way to good work. She needles Labenz: the first things he put on the internet would probably make him cringe now.

The distinction is easiest to see with avatar meditation videos. Someone who floods YouTube with hundreds of them because the niche has not been policed yet is doing slop. Someone who genuinely wants to be a meditation teacher, uses an avatar, and makes something ugly is just executing a bad idea. The clips can look identical. The nature of the act is not. Labenz admits his comment section is almost uniformly opposed to the AI cover art he now generates for the show. He keeps doing it anyway because the process is fun. Burkhauser says that is exactly her point: playing with a new medium is how you get to good stuff, the same way you would develop a painting style by sitting in front of a canvas and making a lot of bad pictures first.

If so much of what is online still looks terrible, she offers two reasons. The technology is not there yet: you generate five to ten seconds at a time, hope the voice matches across clips, and sometimes roll the dice fifty times to get one usable take. Anyone making good generative work right now is fighting the tools the entire time. The second reason is stigma. People with real taste and skill often will not publish and own that they used these models, so the feed fills up with early players who do not yet know what good looks like. Labenz says his creative teammates at Waymark are clearly better than he is, and he cannot always name why. Burkhauser locates the gap in vocabulary. If you have never studied photographic composition, you know the frame is off without having the words to say so. A few adjectives, an artist's name, or a note on composition from someone who does have those words usually takes the work up a notch.

The same email recruited a generative AI advisory committee. Sign-ups blew past any meeting that could actually convene, so Burkhauser talked to most of the people one by one. She sent the note in the first place because she had just taken the CEO job into a pricing change, always a tender moment with a loyal and vocal customer base. One piece of dual feedback kept recurring: stop spending time on AI features and put that time into the core quality of the app. When you hear that while you are raising prices, she says, you should get curious. Some of it was real: performance, reliability, upload and playback speed, a list the team has been knocking through. Some of what users called core quality was, to her mind, already AI.

Descript has been AI-native since day one. Editing a video like a transcript is an AI idea. Green screen is a visual model. Studio Sound is a model the company trained. Voice cloning, Overdub, and Regenerate — change a wrong word and have the recording say the right one, in your voice, with matching lips — are models too. Once she sorted the complaints, a hierarchy of hostility appeared. Users love AI when it is an effect, a transition, or a button that does something that feels deterministic, even if a model is underneath. Green check mark, keep building that into infinity. Underlord, the natural-language co-editor, sits in the middle: everyone wants it, everyone is excited about killing editing drudgery, and many are mad it is not perfect yet for their use case. Nobody is saying do not build it. They are asking why it is not already better.

The hatred concentrates on generative video. Avatars get some of it; voice cloning and text-to-speech mostly do not. Dig into the anger and two creator-specific reasons show up on top of the broader pause-AI weather. First, people feel they are taking crazy pills: the hype says this stuff is incredible, and then they use it and it sucks. That is a reasonable reaction, she says, because a good experience still requires you to be deeply invested in the model, and the industry does not talk about that enough. Second, the marketing eats itself. How many times, she asks, have you seen a line like Seedance just put a gun to Hollywood's head and pulled the trigger? Sell the technology that way and the people you just announced a gun on are not going to be excited to use your product.

That is not how she talks about it, and not what she believes. She does not think generative video will end traditional film, end recorded media, or put everyone in those trades out of work. It may shift jobs and it may create jobs; it is not clear to her that mass displacement is the main story. Film was a new tool once. This is another one, and the right posture is play and curiosity rather than threat. Labenz, coming back after a sponsor break, says play is one of his own refrains: current frontier systems reward it more than any technology he has used.

Labenz asks how you choose which generative models to ship when they are so hard to benchmark. Burkhauser describes two gates. First, should this even be available inside Descript? Second, should it be the default? Most people will never touch the model picker. People deep in AI swap Nano Banana Pro for photorealistic face swaps and Kling for another job; the average user does not want that sophistication and will accept whatever you put in front of them. Whether a model gets in at all is, she admits, a little bit vibes. The team does not eval every model on earth. They use fal as their inference provider, so if you are not on fal you are probably not in Descript. A custom connector means a new data-license agreement, and they would rather not. Seedance is in the product now because it finally landed on fal.

Inside fal they try to keep representation from the labs that sit on the public leaderboards month after month, because you are always a week away from one of them having the best thing again. The default is a more formal process: external evals, plus their own tests on common customer use cases, plus an A/B test against the current default. If the live test matches the internal evals, they ship. Image generation currently defaults to Nano Banana Pro. Video defaults to Google's Veo, and they are considering replacing it with Seedance. Internal evals are a panel of trusted people scoring outputs, not a fully automated pipeline. Labenz notes that aesthetic quality has always been a hard thing for models to judge. Burkhauser is not sure you should try to automate it.

She had dinner with the CEO of Midjourney, who told her they still have the most aesthetic image generation because he keeps his thumb on the scale. Google, in that telling, lets a democratic panel or some automation pick, and the winner is always a generic pretty blonde woman. Burkhauser thinks it is reasonable, on a judgment-heavy task, to have a vetted human with taste make the call forever. Studio Sound is her in-house example. It was originally built by a cellist with a good ear; the first eval was that person listening and saying this one is better. After he left, the team had to write down what his ear was doing: thirty-seven different things that make one form of background-noise removal better than another. They recently evaled Studio Sound against every new provider and kept their own.

She does not believe generative video will be winner-take-all. Oscar-grade effects and cheap-enough product-page video for Amazon are not going to be won by the same model. Some jobs are bulk and do not care about consistency across time or voice; some jobs are worth thousands of dollars per generation. The useful question is who your core customer is and what the hero use case is. Seedance is the live dilemma. It is strong when you want an opinionated edit that makes artistic choices you did not ask for. Descript users mostly want B-roll, and B-roll is not supposed to steal the scene. Seedance can be too expensive and too much of a scene-stealer for filler. That is why the default is still hanging. She also thinks orchestrator agents will eventually have to pick the right model from project context, hitting a cost, quality, and use-case bullseye.

Underlord currently "sees" video by captioning it: frame-by-frame text descriptions, with tricks to fake eyes and ears. Burkhauser says it does okay and that working multimodally is the agent-quality team's number-one priority, with a major upgrade due in a month or two. Labenz has seen Gemini hallucinate hard cuts that were just him moving his head. The build-versus-buy line is a strategic bullseye. Descript wants to be the world's best at editing recorded media. Regenerate is the example: change an answer after the fact, or update a two-year-old explainer because California added a law, without re-recording the whole video. Jump-cut smoothing is about to ship, so a choppy edit looks like the speaker simply moved. Pure generation they do not want to own. Companies spending hundreds of millions there are still going to lose to Google, she says, and she does not want to set money on fire that way.

Part of her feels possessive about human facial expression and unsettled by clones that purport to be her and are not. Another part would love to have an authentic conversation and then, in post, fix the makeup, the lights, the outfit, and the stupid line. The killer use case she wants is augmented human recorded media. They invest where they have data nobody else collected, where they can build without breaking the bank, and where a frontier lab is unlikely to care. Removing retakes is much less interesting to Google than voice consistency between three-second clips. User thumbs-up and thumbs-down tell the team where the quality bar is being missed, whether on a deterministic action like remove-retakes or on an Underlord request. She tells Labenz to give the agent constraints: get rid of filler words unless you cannot make a clean cut, in which case keep them.

Underlord is supposed to be truly open-world, not a menu of twenty-eight tools that dies on idea twenty-nine. Evals grade a random sample of real user queries with a swarm of LLM judges, averaged across old and new versions. Grade one is you did not break anything, target near 100 percent. Grade two is you did what I asked, target 90 percent. Grade three is you did it well, and she wants that at 80 percent by year end, because that is when people start to trust a co-editor as doing the job at about their own level. Rough cut is already a bright spot. Visual work is not, which is why multimodal is the priority. Filler words showed up as a hotspot in the thumbs, and the team spent a couple of sprints on it. The model bet is a generalized harness with low-level tools, on the assumption that general intelligence keeps improving. When the next Claude drops, they want it evaled within fifteen minutes and in the product. They are not spending research money trying to keep up with the labs.

A design principle follows: Underlord should not be able to do anything in the editor that a human cannot, and vice versa. It is a collaborator, the way Descript has been a team tool since day one. Today it lives at the project level. It needs to live at the drive, and then outside Descript entirely, via MCP, as a collaborator you bring into Claude Cowork. Burkhauser already works that way: she asks Cowork to look across a week of Notion and Slack, propose six clip ideas from things she actually said, dump scratch scripts into new Descript projects, and then turn finished recordings into LinkedIn quotes with a skill she wrote. A user she is obsessed with triggers an edit-podcast skill off a Zoom recording and only opens Descript to check. Labenz has a v1 of his own that trims everything before "welcome to the Cognitive Revolution" and after the thank-you. The API, he notes, does not even offer fully deterministic fine-grained edits; you prompt Underlord. Transcription can be a deterministic tool. Orchestrating media, visual edits, and layouts stay in Descript. Think of it as telling Claude Code to hire Underlord as your video team. If Google or OpenAI decided this job was exactly what they wanted to be great at, she could probably do nothing. A reliable video editor is high up the tree. You climb a lot of branches before you volunteer to maintain one forever.

Labenz pushes on a stranger horizon: OpenAI has publicly aimed at an autonomous AI researcher by March 2028. If coding agents can sustain the effort themselves, specialist models and custom sense organs for video stop looking like a moat. Burkhauser gets excited about automating more labor if society does the work to make that a good future, and she is skeptical of the timeline. When someone makes a claim like that, get it on paper. What can the system do without human oversight, and what can it not? The picture usually gets more reasonable than the topic sentence. She thinks we overstate near-term labor and social change from AI and understate the longer-term shifts. She cannot tell you what the labor market looks like in five years or where Descript sits in it. She is confident the company has rituals that let it shift strategy quickly.

Will podcast editor still be a job in two or three years? She does not know. Will people be employed to tell stories? Yes. What kind of stories, in what media, for whom, is all subject to change. If what you are good at is interviewing people and getting what is interesting out into the world, that remains a job, especially if you embrace new media and new channels. Labenz's colder question is whether the winners and losers are even the same people — whether the ones displaced can pivot into the new winning positions. Burkhauser says economies always reallocate, and if this one happens on a compressed clock, societies need systems that catch people through extreme disruption. She does not subscribe to permanent, irrevocable losers unless that is the society we choose to build.

The practical product problem is closer. Individual button clicks and Underlord prompts can now burn a few dollars of credits in one go, a weird new world for software that used to let you click for free. As a consumer, Burkhauser hates the feeling, even when the deal is good: a dollar for someone to cut your clips beats an afternoon of doing it yourself. AI still costs Descript real money, so free and unlimited is not an option. Pricing is built around cadence. Hobbyists should be able to make one really good thing a month, creators one a week, business teams several a week. Credits sit in a shared pool rather than per feature, because a podcaster may never need AI speech and may need clips several times a week. The health gate is that less than 5 percent of monthly active users should have to buy a top-off. If half the park is hitting the limit, it is not a fun amusement park. She thinks this is a temporary shape. The consensus destination is outcome pricing, maybe on exports: you are not charged until you get the value. Models are still too expensive for that world, so everyone is living in the uncomfortable one.

Labenz's last zoom-out is whether a future of infinite slop can be avoided. Models are still expensive, which is one brake; cheaper generation without a new social contract could take that brake off. Burkhauser says the people who ask her this are usually in tech or economics, and with respect to both, they miss that content is a little bit business and a little bit art. Artistic expression is not as clearly driven by free-market nihilism as other parts of the world. The camera did not kill painting; it permanently changed the medium, because people lost interest in photorealistic painting once a machine could do that job. Art reacts to the technology and the culture of the day in ways that surprise us. The truly creative work will move first. Businesses will then want a hunk of it.

She reaches for Miranda Priestly in *The Devil Wears Prada*: a designer puts cerulean in a spring collection, and years later you buy it in a bargain bin at Marshalls. Content works the same way. People with an aesthetic eye will do unexpected things with this technology and in defiance of it. Marketers will steal the look. It is easy, especially in a skeptical moment, to look into the future and see nightmares. People pitch her a world with no human creators, just a phone that watches your eyeballs and generates the next seconds so you never look away. David Foster Wallace, she notes, already wrote that in *Infinite Jest* in the nineties. Maybe it happens. She is more excited to see what artists do with the tools and against them, and she thinks those things will win in the marketplace even in a world of slop.

The conversation leaves a short working list. Decide whether a clip is slop by asking if the motive is money and whether the production is scaled; do not give your own ugly experiments a moral pass, and do not club every failed first try into the same bin as algorithmic flooding. Stay in the medium long enough to grow opinions about prompts and styles. If you are stuck knowing something is off without knowing why, look at work you admire until you have the adjectives. When you instruct an open-world editor, add the constraint Labenz picked up on the air: remove filler words unless the cut would be dirty, in which case keep them.

Use the thumbs. Burkhauser is explicit that those votes decide which hotspots get the next few sprints; filler words became a focus that way. If you have a paper trail, copy her Cowork loop: mine a week of Notion and Slack for ideas you already said, dump scripts into Descript, record, then run a skill that matches how you like clips to look. The point of sharing amateur edit-podcast skills, Labenz says, is less the quality of any one workflow than the discovery that you can do this at all. And when a lab publishes a date like March 2028 for an autonomous researcher, slow down and write the bet on paper until the unsupervised abilities are named. The topic sentence is usually louder than the claim.