AI Video Is Now Made Faster Than You Can Watch It

Someone has left an AI model running, and it has been making video for hours. Each clip is generated a few seconds before anyone sees it, plays once, and is replaced. Nobody filmed any of it and nobody edited it. People are calling these slop streams, and the first ones appeared this month.

They exist because of one threshold, crossed this summer: a model that makes a second of video in less than a second. Here is what these streams actually are, what they cost, and what they mean for anyone trying to tell real video from generated.

What one of these streams looks like

Open the page and a video is already playing, with the sentence that produced it sitting above. A viewer typed that sentence. On Infinite Slop, built by Pieter Levels on an idea from two earlier streams, Levels says anything written in the chat becomes the next clip, and the model is asked to link it to the one before. About 1,000 people were on it when we looked, writing the channel between them a sentence at a time.

So the machine making video without stopping is only half of it. The other half is that a stranger's idea now reaches an audience in the time it takes to watch five seconds of video.

Nobody is charged to watch, nothing is sold on the page, and no advertising runs against it. Which raises the question of who pays. The answer turns out to be the most useful fact in the story.

Why it is possible now

Cross that threshold and generation stops being something you wait for. It becomes a tap you leave running.

Getting there took a video model retuned for speed instead of quality, reportedly 50 times faster than the version it came from. Once that existed the rest was easy. Within days someone had wrapped it in an endless scrolling feed that looks exactly like TikTok, generating each clip as you swipe to it. He says it took a weekend.

The second part matters more than the first. Speed was a research problem and took a company to solve. Building on top of it now takes an afternoon.

What it actually costs

Hosted video generation is priced by the second of footage. At the going rate for a mid-tier model at 768p, around 8 cents a second at the time of writing, the arithmetic is unkind:

• One minute: about $5
• One hour: about $290
• One day, non-stop: about $6,900
• One month, non-stop: about $207,000

You do not have to take our arithmetic for it. Levels says it would be "very expensive to run this" and that the model provider sponsors it outright. A developer who built an infinite AI TikTok feed over a weekend said his version was "pretty pricey" and that he "would never be able to release this without losing money".

Nobody is drowning the internet at $200,000 a month per channel. These streams exist because a company is paying for the demonstration, and demonstrations stay affordable while there are five of them rather than five million.

So if you were bracing for an infinite-slop apocalypse arriving on somebody's credit card, you have some time.

The part that should worry you instead

The flood, when it comes, will arrive through models you can download rather than through a paid API.

That is already happening. The model behind every clip on this page is MiniMax H3, released on 31 July 2026, and its weights were published three days later. What you have been watching is downloadable now.

Open-weight video models run on hardware you can rent by the hour or own outright, with no per-second meter. Once the hardware is paid for, the only cost left is electricity. Weights that are public stay public: they get copied, fine-tuned and passed around long after the company that trained them has moved on. The curve worth watching is not how fast the newest model generates, but how good the best model is that anyone can run without asking permission.

Something else disappears along with the meter. Hosted services refuse work. Making the clip at the top of this article we tried a prompt that mentioned a well-known racing series and were turned down flat, and recognisable characters appear to be blocked the same way. That is why these streams look the way they do: generic cities, unnamed people, no logos on anything. What reads as a house style is a refusal screen. Filters differ between providers, and looser ones have already produced plenty of AI video full of characters their makers had no right to use.

A model running on your own hardware has no refusal screen. The restraint you see in AI video today belongs to the companies serving it, and it comes off the moment the serving company is cut out.

What this does to detection

Detectors like ours answer a specific question about a specific object: here is a video file, was it generated? Sample some frames, compare them against what generative models leave behind, return a verdict.

A live stream breaks the shape of that question. No file exists, and nothing ends. The honest answer changes minute to minute, because models drift, scenes cut, and a stream that was obviously synthetic at 3 p.m. can be indistinguishable an hour later after the operator swaps a setting.

Checking a stream means sampling it continuously and reporting a moving confidence. That is a different instrument, and most of the tools people rely on today, ours included, are not it yet.

Out of curiosity we put every clip in this article through our own detector. It caught most of them, and got the model wrong every time, since this one is a month old and it has never seen it before. Two it missed entirely.

That is what we are working on.

What to do about it now

Nothing dramatic. Two habits and one caution.

Check the source before the pixels. Who posted it first, and do they exist? A reverse image search on a single frame answers more questions, faster, than any detector, including the one a detector cannot answer: whether a real video is being passed off as something it is not.

Treat a live stream as unverified by default. Unverified, not fake. If it matters, find the same footage somewhere it can be downloaded and examined.

And read any detector, ours included, as one input. Every tool is weakest on models released after it was last updated, and new generators appear faster than detectors learn them. A confident answer about a brand-new model is worth less than the same answer about a two-year-old one.

The melting Will Smith took three years to become a stream that never ends. Whatever replaces the stream will not take three.