Dumme, a startup putting AI to practical use in video editing, is already generating demand before opening up to the public. The Y Combinator-backed company has hundreds of video creators testing its product, which leverages AI to create short-form videos from YouTube content, and a waitlist of over 20,000 pre-launch, it says. Using a combination of both proprietary and existing AI models, Dumme’s promise is that it can not only save on editing time but also — and here’s its big claim — do a better job than the contracted (human) workforce who is often tasked with more menial video editing jobs, like cutting down long-form content for publication on short-form platforms like YouTube Shorts, TikTok or Instagram Reels.
Founded in January 2022 and a participant in startup accelerator Y Combinator’s Winter 2022 program, Dumme co-founder and CEO Merwane Drai said he was originally focused on building a search engine for video. But around six months ago, the team realized that a better product might be to repurpose the same AI models they were developing to edit video clips instead.
Joined by co-founders Will Dahlstrom (CPO) and Jordan Brannan (CTO), all with AI backgrounds, Drai realized Dumme may have landed on the right product-market fit after their app went viral, crashing their servers.
“We didn’t really expect that it would get a lot of traction or anything, so we just put something out there,” Drai explains. “Then what happened is that overnight, we woke up to overloaded servers — like, nothing actually worked. So we took everything down and actually put together some sort of waitlist,” he continues. “The next morning, we probably woke up to 5,000 people in there, which was interesting.”
The team later discovered that a TikTok creator had posted a short video about the product, which sent a flood of traffic to their site.
“It actually never calmed down from that,” Drai notes.
The product, pronounced “dummy,” appealed to creators because it aimed to simplify and speed up the work involved with video editing.
Image Credits: Dumme
Using Dumme is as simple as the name implies. To get started, the user pastes in a YouTube video’s link, then clicks on “generate” and the AI will output a number of short videos showcasing highlights from that ingested content. The company says it’s using YouTube as the source, instead of supporting raw video footage, in order to outsource content moderation — that is, if it’s allowed on YouTube, it’s allowed in Dumme.
The processing time and the number of resulting clips will depend on the length of the original video.
But as an example, an hour-long video podcast might take around 20 minutes to process and you’ll start receiving clips after about five minutes, Drai says. When complete, creators can download the video clips, which are under 60 seconds by default, and upload them to any platform that supports short-form content, like YouTube Shorts, but also other platforms, like Reels or TikTok.
Image Credits: Dumme
How this all works on the backend, of course, is much more complex. The company says that, initially, Dumme will learn as much as it can from the source video via the metadata. It then transcribes the video and tries to understand the semantics of what’s being said while also looking at the frames to try to decode the emotions of the person speaking. These findings are correlated and passed to a language model that tries to determine what parts of the video are worthy of clipping. That’s then handed off to another model that tracks active speakers and handles cropping.
Dumme says it’s working with existing AI models like GPT-4, a fine-tuned version of Whisper, and others it built in-house — like the model that tracks the active speakers in a video frame. One of its models is also trained on a bunch of YouTube Shorts to learn what makes for a good opening hook to pull viewers in. And, though not yet live, the team is also experimenting with an open source model, LaViLa from Facebook Research, to better understand the context of the video.
The AI work is being done on GPU Cloud provider CoreWeave, not AWS, as it’s more affordable, the company tells us.
Because Dumme relies on AI that processes spoken words, the tech is not appropriate for things like long gameplay videos or others where people aren’t talking. Drai says the startup is initially targeting YouTube creators, podcasters, and agencies — the latter, they believe, would be the best bet for monetizing the product.
Image Credits: Dumme
Agencies, explains Drai, today often outsource this type of work with hit-or-miss results.
“They just pay contractors in cheap jurisdictions to edit their own content. And the problem is that it’s still actually pretty expensive and it takes a lot of time — it takes weeks, not minutes,” he says.
Asked how he feels about creating a technology that would actually put people out of work, Drai was not worried.
“The way I think about it is that, eventually…I think this is like telling me that math teachers are going to [be put] out of work because there’s something called a calculator…,” he explains. “People are going to adapt. And then there’s going to be someone teaching you about the calculator, right? So I think it’s just a matter of adapting to this,” Drai says.
Currently, the pricing being considered involves tiers where a business would