On the Denoised podcast, we look at how far AI video has moved into live, interactive territory, and why audiences keep falling for synthetic movie magic. We put Lucy 2.5 through a live webcam test, debate how MiniMax H3 handles a stack of reference inputs, and break down how Kimi K3 runs a 2.8 trillion parameter model without lighting all of it up.
Jump to:
01:25 The viral fake tsunami behind-the-scenes video
09:57 Lucy 2.5 real-time video generation demo
15:10 MiniMax H3 and multi-input generation
18:59 How Kimi K3's open-weights architecture works
25:37 Chips, GPU prices, and compute limits
Quick Take
One of the most talked-about pieces of "filmmaking" on Instagram was an AI-generated fake, and the tool we demoed restyles a live webcam feed in real time. We think audiences want to see movies made by hand, while the tools that generate images keep moving further from anything physical. That gap runs through the whole episode, from live activations to shot direction to the cost of running the models.
What We Questioned: Why Audiences Fell for an AI-Generated Behind-the-Scenes Tsunami Shoot
We started with the AI-generated "behind-the-scenes" video of a tsunami crashing into a Miami-style beachfront, which we called one of the most popular videos on Instagram. It shows a crew filming a massive practical water effect, and none of it exists.
The location doesn't hold up. The set blends the Santa Monica Pier with South Beach's Art Deco strip and a skyscraper skyline arranged in a way that doesn't exist.
The production details don't hold up either. Crew members wear shirts reading "effects crew" and stand directly in the shots, which we flagged as a safety violation, since nobody would be in the way on a real set.
Cinematographer Valentina Vee posted a detailed breakdown on X walking through why every element is fake. We think the clip comes from the same creator behind a series of fake disaster-movie shoots.
We think audiences want movie magic they can touch. We trace the appeal to nostalgia for scale models and practical effects, like the physical dinosaurs of Jurassic Park that the public still associates with moviemaking.
We pointed to theme park attractions built on that appetite, including Catastrophe Canyon at Disney-MGM Studios, Waterworld at Universal Studios, and the Earthquake attraction at Universal Studios Florida, whose pre-show recreated stunts from the film before revealing a huge model of an earthquake-destroyed Los Angeles.
The irony runs two layers deep. As we see it, a film like this would never be shot practically, since water at that scale would be a digital sim built in tools like Houdini or EmberGen. We think production is starting to move another step away from that, toward neural networks generating the image outright.
The same tension shows up in synthetic podcasts. We also looked at HeyGen's AI-generated video podcasts, where we found the faces uncanny and the voices off.
We see room for synthetic audio as a study aid, such as a generated deep dive on a technical paper. For industry news and conversation, we'd still pick hosts who are real people.
What We Tested: Lucy 2.5 Restyles a Live Webcam Feed, and Generated Hair Reacts to Your Hand
We ran Lucy 2.5 from Decart on a live webcam feed, giving it prompts and watching the model restyle the video in real time as we moved. It works video-to-video, so the input is whatever the camera sees.
Simpler prompts seemed to work better than stacked ones. Our first prompt asked for an apocalyptic desolate jungle, a battle helmet, and long blonde hair flowing beneath it. The model made the change, but the output looked closer to an iMovie-style effect than a finished shot.
Your hand can push generated hair around. Lucy 2.5 recognized a hand in frame, and pressing on the generated hair squished it as if it were physical. A 1980s rock star prompt with long curly blonde hair, a leather jacket, and Kiss-style face makeup was a lot of fun to play with.
Hair deforms, while other materials stay rigid.
- Chainmail on the arms didn't deform when touched.
- A Davy Jones reference image added the character's tentacle beard but didn't make the face look like him.
- The model treated the tentacles as rigid, even when flicked.
- Facial articulation held up, including a wide-open mouth.
- Hands and faces held up reasonably well during the chainmail test.
Live activations look like a natural fit. We see Lucy 2.5 fitting theme park and IP-character activations, such as turning guests into a character at the end of a ride. A performer could appear as Davy Jones on an LED screen outside Pirates of the Caribbean and riff with the crowd, the way live character comedy shows already work in parks.
Production use is still untested. We ran it on a webcam, and we suspect the model was trained mostly on webcam-style footage. We want to try it from a phone, with multiple people in frame, and as lo-fi previs or reference capture.
What We Debated: How MiniMax H3 Decides What to Do With a Pile of References
The Hailuo 3 model, MiniMax H3, takes a mix of references in a single generation. In the example we discussed, that meant 9 images, 3 videos, and 3 audio clips.
The reaction around it frames H3 as a cheaper alternative to Seedance for multi-input generation, with apparently solid output quality.
The orchestration happens before the generation. As we understand it, a model handling that many references first has to infer what you want, then route each asset to the right place in the shot. We see that intent step as the core of these models, or as we put it on the show, "the orchestration of the generation before the generation."
Direction beats long prompts. We've had success with grayboxing, rough 3D motion references, and images marked up with camera paths, and prompts are our least favorite way to direct. Detailed prompts with timecodes help, but our ideal would be an annotated storyboard passed in as a reference image.
Camera language is the test we want to run. That means drawing an arrow for a camera move, a circle where the camera stops, and a prompt telling the model to hit that circle at the 2-second mark. Whether H3 can read that kind of markup is still open.
Open weights are coming. MiniMax plans to release H3's weights, though they aren't available yet. That would put a model we'd place near Seedance for inputs and quality on a local workstation, likely a powerful one.
We don't know whether the release will be a distilled version that runs on a single GPU or the full model needing server hardware. Either way, local weights open the door to refining it and building your own versions.
We count Hailuo among the labs worth tracking alongside Alibaba, Tencent, and the team behind Kling.
What We Explored: How Kimi K3 Runs 2.8 Trillion Parameters by Activating a Sliver of Them
As an open-weights model, Kimi K3 lets anyone download the weights, inspect the architecture, and run it themselves. On the show, one of us called it supposedly better than Fable 5, a claim we can't confirm.
We think that openness matters given what closed models cost to use, including the weekly usage caps we keep running into.
Mixture of experts keeps a huge model fast. As we described it on the show:
- Kimi K3 has 2.8 trillion parameters.
- By our recollection, that network is split into roughly 900 experts. As a simplified illustration, one might handle a finance problem and another image understanding.
- For a given prompt, the model activates around 16 of those experts, about 2% of the whole network.
- The approach follows the mixture-of-experts design we associate with DeepSeek.
Compressed context lets it hold more. As we understand it, Kimi K3 applies latent compression to its context window, which matters for tasks like coding across an entire codebase. We compared it to compressing raw video into H.264, which keeps most of the information in a fraction of the space.
We described it as the same family of technique that lets image models trained on billions of images run on a consumer GPU with around 20GB of VRAM.
We think architecture is carrying gains that scale no longer delivers. One of us guessed this design is how Kimi K3 could approach Opus-level performance at lower token cost. We noted that models have already scraped much of the public internet, and, as we paraphrased him, Yann LeCun has argued scale alone won't be enough.
For a deeper technical walkthrough, we recommended Caleb Writes Code's explainer, which covers Kimi K3 in 13 minutes.
What We Debated: Why Chips and Power Are Becoming the Limit on AI Models
We noted that more Chinese model developers appear to be optimizing for Chinese-manufactured chips rather than NVIDIA hardware. One of us recalled MiniMax being among them, though we weren't certain.
As we see it, those developers are trying to get out from under the limitation NVIDIA has become.
Consumer hardware is feeling it too. One of us had read that NVIDIA's consumer RTX cards are set to rise another 30% in price, a figure we haven't verified.
One of us pointed to Apple and other companies raising prices as well.
Electricity, water, and silicon set the pace. We see the data center buildout as a question of infrastructure and compute, and we see those three resources as the bottleneck.
Bottom Line: AI Video Is Getting More Directable While Compute Sets the Limits
Across the episode, the thread was how much direct control creators get over these models and what it costs to run them.
- Fake tsunami shoot. We think audiences still picture practical effects, even as we see real productions using digital sims and starting to move toward generation.
- Lucy 2.5. Real-time restyling of a live feed, with generated hair that reacts to touch, looks well suited to activations. In our test, chainmail and tentacles stayed rigid, and production footage, phone capture, and multi-person scenes remain untested.
- MiniMax H3. Multi-input generation depends on how well the model parses intent, and promised open weights could bring it to local machines.
- Kimi K3. Mixture of experts and compressed context show how we think architecture is carrying gains as scale flattens.
- Chips and compute. Silicon, power, and water set the pace, and Chinese labs appear to be optimizing around NVIDIA's constraints.
The tools gaining ground take direction from images, motion, and physical gestures, and open-weights releases could move more of that control onto hardware creators own.
Links from This Episode
Tools & Platforms
Lucy 2.5 - Decart's real-time video-to-video model, which we tested live
MiniMax H3 - multi-input generation model with open weights planned
Kimi K3 - open-weights mixture-of-experts model
Seedance - multi-input video model we compared H3 against
Kling AI - made by one of the Chinese labs we follow alongside Hailuo
Companies
News & Analysis
WATCH THE FULL EPISODE ON YOUTUBE →
