For the entire history of generative media, one rule held: making video takes time. Generate 15 seconds and wait 2 to 5 minutes. That rule just died. AI video generation is now faster than real time, and it changes what any of us can make with a text box.
What just happened
On August 27, Wharton professor Ethan Mollick pushed “generate” on a web interface and got back a reasonably high-quality video in less time than it took to watch it. He called it a line being crossed. By August 29, people had hooked the model to Twitch streams. One livestream hit 5.75 million views before platforms shut it down.
Here’s the timeline nobody’s putting together for beginners:
- MiniMax launched H3, an open multimodal model that generates video with native stereo sound (up to 15 seconds at 2K).
- Fal.ai took H3, trained a version called H3 Max, then optimized its own inference engine for a 35x speedup over the official endpoint.
- The result: generation faster than playback. Infinite video, generated on the fly.
- Twitch and YouTube kicked Fal off for streaming it, so Fal launched its own live service, fal.live, an “AI television” where the chat directs every scene.
- Entrepreneur Pieter Levels built infinite-slop, a live stream where anything you type in chat becomes the next video, linked to the previous one.
Wild, right? Twitch banning AI video generators for being too fast has to be a first. This is the first time AI video generation has outrun human attention, and the whole industry is still scrambling to react.
What H3 Max actually is (in plain English)
H3 Max is a video model, but it’s not just a video model. MiniMax built H3 as a general-purpose “multimodal” generator, meaning it understands text, images, video, and audio together. You can say “use the camera move from Video 1, have the character in Image 2 sing, and match the vocals to Audio 3,” and it just handles it. That kind of instruction-following was science fiction a year ago.
The “Max” part comes from Fal’s optimization work. Same base model, post-trained for cost and quality, then squeezed through an inference engine tuned for speed.
The 35x speedup nobody’s talking about
Speed is the whole story here. The official H3 endpoint already priced video at under a third of mainstream models per second. Fal’s version runs 35x faster than that endpoint. Levels.io described the jump as 50x faster than the old way of generating video.
Think about what that means practically. Before: you plan content around a 5-minute wait per clip. Now: you type a prompt, and the video appears while you’re still reading your own prompt.
Here’s the before-and-after in numbers:
| The old way (2024-2025) | H3 Max today | |
|---|---|---|
| Time for a 15-second clip | 2 to 5 minutes | faster than real time |
| Resolution | mostly 720p to 1080p | up to 2K |
| Sound | separate step, often silent | native stereo audio |
| Cost per second | mainstream rates | under a third of mainstream |
| What you do while waiting | make tea | nothing, it’s already done |
That row about sound matters more than it looks. Audio used to be the forgotten stepchild of AI video. H3 generates dialogue, effects, and music together with the picture, so a clip comes out finished instead of half-done.
What you can do with real-time AI video right now
Here’s what real-time AI video generation unlocks today. The obvious uses are the ones you’d expect, but some of them are new:
- Product demos. Describe your product, get a video in seconds. For a small business, that’s ad material on demand.
- Explainer content. Talking-head intro, product shot, background loop, all generated in the time it takes to brew coffee.
- Live “AI TV” channels. Chat-directed streams that never end. Sites like fal.live and infinite-slop are crude today, but they’re the first version of something bigger.
- Video editing shortcuts. Generate transitions, b-roll, or placeholders instead of hunting through stock libraries. If you want the full workflow for replacing expensive editing software, our guide on AI video editing walks through it.
Here’s the part I find genuinely interesting: the cost. Cheaper per-second pricing plus no waiting means the economics of video flip for solo creators. You no longer need a production budget to test a video idea. Generate, watch, iterate.
How to try it yourself (for free)
You don’t need to be technical. Three entry points:
- fal.live is the most fun. Open it, type a prompt in the chat, watch it render on screen in seconds. It’s a live stream, so it’s always on. Expect chaos; that’s the point.
- Fal.ai platform gives you the H3 Max model directly with a simple web interface. Good for generating individual clips with prompt enhancement.
- hailuoai.video hosts the original MiniMax H3 hands-on demo, including the audio and 2K features.
Try the cheapest experiment first: generate a 10-second ad for your own product or service. Show it to someone. That exercise alone tells you more about this tech than any article.
Here’s a sample prompt to get you started: “A cinematic drone shot flying over a neon city at night, rain bouncing off the lights, then a slow zoom into a coffee shop window.” Type it, hit generate, and watch how fast your description turns into footage.
The catch: everyone’s calling it slop
Let’s be honest about quality. Latent Space, the newsletter that covered this story closely, called the live streams “pure slop”: no plot, weird imagery, clearly machine-made. I watched a few minutes of fal.live and yes, it’s a fever dream. There’s no pretending otherwise.
But here’s the thing: this is the worst it will ever be. Every generation of video AI has followed the same curve, from unwatchable to usable to very good, and each step has come faster than the last. Sora looked rough at first. Now look at it. If you want a free alternative that already works well today, our Seedance 2.5 review covers a solid one.
The slop phase still matters because it’s teaching us what to watch out for. If you’re publishing anything, you should know how to check whether a video is AI-generated, because your audience will be seeing more of it whether they want to or not.
What this means for creators and small businesses
Three takeaways, in order of importance:
- Video production just got a lot cheaper. The barrier to testing video content is now a text prompt and a few seconds. That favors small players.
- Differentiation shifts to taste. Anyone can generate video now. What you choose to make, and how you edit it into something worth watching, becomes the skill. This is why understanding the full pipeline matters: generation is one step, and it’s the easiest one.
- The flood is coming. Faster generation means more AI video everywhere, and every major AI video generation tool is now racing to ship its own real-time mode. That also means the tools to identify and filter it matter more. We have a guide on spotting AI video with free Gemini tools if you want to stay sharp.
For creators, I’d treat this like the early days of stock photo sites: the tools get commoditized, the people who pair them with good judgment win.
One more thing worth saying. Real-time generation doesn’t replace storytelling, and it doesn’t replace editing sense. It removes the waiting. If a 10-second idea takes 10 seconds to make, you’ll make a lot more ideas, and most will be bad. That’s fine. The few good ones are the point.
Takeaway
Try fal.live today, even for five minutes, and generate one video about anything you work on. That’s the fastest way to understand why AI video generation just crossed a line. If you’re creating content, start planning for a world where video is as cheap as text, because it’s arriving faster than anyone can watch.
