AI video is moving beyond clip slop

This is Lowpass by Janko Roettgers, a newsletter on the ever-evolving intersection of tech and entertainment, syndicated just for The Verge subscribers once a week.

Hollywood is cooked — or so a growing number of people on social media would like you to believe. Their purported proof: AI-generated clips of Daniel Craig riding a Vespa through an Italian city, Godzilla fighting King Kong, or The Avengers zooming through Manhattan.

In reality, cheap slop like this won’t replace Hollywood blockbusters any time soon. However, a new generation of AI video solutions could upend how studios work. That’s because, until recently, AI companies basically tried to sell Hollywood on the same idea as those Twitter guys, with a slightly more palpable spin. The pitch, in a nutshell: AI video will allow everyone to make movies faster, cheaper, better — one prompt at a time.

“The premise was: Substitute your camera for our video model,” says Luma AI CEO Amit Jain, whose company used to make that very same pitch to studios. But when it began partnering with the entertainment industry, it received a crash course in the way Hollywood actually works.

“It’s not sufficient to just produce a clip,” Jain says now. “Because then what?” Clips generated by video models are typically 10 to 16 seconds. “That’s not a shot. That’s not a sequence. That’s not a scene,” Jain says. “Churning out short videos is not enough.”

Now, AI companies like Luma believe they have found a better way to sell Hollywood on AI. The gist? Don’t just use AI for video — use it for everything.

Luma has been working on AI agents that can help with the entire production process. Jain compares this transition to the way software development with AI has evolved, with companies like Anthropic moving from simple vibe coding to agentic workflows.

“It’s not sufficient to just generate a little bit of code,” Jain says. “We need these systems to do long-horizon, end-to-end work. That’s what solves problems for people.” AI agents can do the same for Hollywood, he believes.

Luma isn’t alone with that approach. Just this week, Google unveiled a new version of its AI media authoring platform Flow that also emphasizes agentic end-to-end work over simple clip generation. “There’s this huge evolution that’s happening in generative tools,” says Google Labs VP Elias Roman. “Moving forward, they’re going to become much more like agents.”

In the new version of Flow, an agent guides the user through multiple steps, from starting with a concept to fleshing out plotlines to developing characters to setting the desired look and feel. And when it’s ultimately time to generate video, the agent uses the things it learned along the way to achieve a specific outcome without having to be prompted about every single detail.

One issue this is aiming to solve is consistency. Generative AI has long struggled with keeping characters looking the same from clip to clip. In the new version of Flow, users can add a character they developed for a project to a prompt simply by tagging it, just like you’d add a colleague to a Slack conversation.

A new generation of video models is also better at understanding physics, the look of a certain era, and cinematic languages. Google’s Flow is powered by the company’s new Gemini Omni world model, while Luma has developed Uni-1 as a unified model that doesn’t need extremely complex prompts anymore to make sense of an envisioned world.

Luma recently teamed up with Amazon to produce The Old Stories: Moses, a companion special for MGM’s House of David show. While shooting Moses, actors would perform in front of LED walls showing backgrounds generated with Luma’s video models, while their costumes were rendered with AI as well.

If a shot turned out to look not quite right, all it took was one new prompt to generate a new asset. “This level of production would take about six weeks to eight weeks per hour of television,” Jain says. “Now, it’s taking them a week.”

Some studios are increasingly embracing that change. Netflix acquired Ben Affleck’s AI company InterPositive in March, and launched its own AI animation studio the same month. Two major Hollywood studios already use Luma’s AI agents, Jain claims. He declined to name names, but the company has been publicly celebrating some smaller wins: Luma recently announced the launch of a joint venture with indie studio Wonder Project, which made Moses.

These developments undoubtedly will lead to job losses, even though the scale of the impact is still unknown. If studios can make a TV show in a month instead of 10, it won’t be sending out checks for those other nine months. The counterpoint AI boosters like to raise is that this will lead to more productions. This could be a silver lining for Los Angeles in particular, which has seen production days plummet in recent years.

Now, we’ll just have to see if Hollywood uses this tech for something people actually want to watch.