AI Generates Disturbingly Real Video for Just $20

AI generates hyper-realistic video: from $140 million to $20

A drone jumping, spinning, and falling vertically. A person speaking to the camera with realistic skin, hair, and reflections. None of this was filmed. AI video generation has moved beyond laboratory curiosity: there are clips that, in full screen, are hard to distinguish from real footage. The leap isn't just aesthetic; it's economic. And that's the uncomfortable part.

What does a shot that once required 500 specialists cost now?

The comparison is brutal: 500 special effects experts, one year of work, and $140 million versus one person, one week, and $20. It sounds like sales hype, but the order of magnitude explains why the audiovisual industry eyeing generative models. The breakdown of that bill, with items side by side, shows the true cost displacement. An AI doesn't charge overtime, doesn't get sick, and doesn't negotiate union contracts.

Here comes the first victim with a name and invoice. Actors who don't age and perfect voiceover artists without makeup. The most repeated diagnosis is that this is another group heading toward unemployment, though not everyone agrees: some argue these tools lower production costs and open doors for those who never had the budget to shoot.

The uncanny valley and the errors that betray the machine

Realism has a limit and a technical name. The uncanny valley describes the zone where a robot or generated face looks human enough to cause revulsion, disappearing only when similarity is total. The practical problem today is different: consistency between shots. A single clip looks flawless; chaining ten with the same character, clothes, and lighting does not. Some joke: it works because it's adapted to a specific case; with any other, it fails. The physics of objects, narrative coherence, and synchronized sound remain the wall.

Cinema without BlackRock: the unverified promise

One analytical strand focuses on cultural impact, not technical. If anyone can produce moving images, the argument goes, the cinema and culture oligopoly is broken, and supply multiplies without needing a large investment fund behind it. It's the optimistic version, coexisting with the pessimistic: an ocean of content where the good drowns in noise. The observed reality is more pedestrian: channels proliferating like mushrooms, and the majority of pieces that don't last a full viewing.

Can you make a deepfake in real time?

This question remains unanswered. Software exists to generate a known person saying something they haven't said, and the next doubt is whether this can hold up live, in a video call, with a false identity on the other side throughout the conversation. No one in the exchange offers a solution or date. The suspicion is reasonable: if synthetic video already fools the naked eye in post-production, moving to real time is a matter of computing power, not principle.

What AI still can't do

Let's lower the hype. Generating short, spectacular clips doesn't equal building a 3D video game, which requires sound, music, space, balanced physics, high-level programming, and a narrative sustaining hours of gameplay. The prevailing prediction is that 2025 will be the year of generative video acceptable for professional use, and that a game the size of an old console would signal crossing another frontier. Meanwhile, projections move faster than results: full movies in a year or two, autonomous agents after, and indistinguishable androids at the end of that path. A scenario that, even its defenders admit, is both impressive and terrifying.

The exact point where the analysis stalls: no one knows how long it takes to solve the consistency problem, and without it, the promise of automated cinema remains a collection of dazzling fragments that don't fit together.