The expensive part

Two essays rose to the top of Hacker News this week, written by different people about different subjects, and they are the same sentence.

The first is called "Attention Is All You Have." Its argument fits in a line. The models are effectively free now, and the hours a person spends looking at what they produce are not. A human being has the same twenty-four hours they have always had. Every week the industry manufactures more things for that person to spend those hours on, and the result is not abundance. The result is a queue. The bottleneck did not vanish. It moved, from the machine that makes the thing to the person who has to look at it.

The second essay is quieter and sharper. Colin Breck writes about receiving a personal message on a sensitive subject that had clearly been run through a language model before it was sent. The message arrived technically fine and emotionally wrong. "Impersonal, dispassionate, and disjointed," he calls it. His explanation is the best thing anyone has written on the subject this year. Writing is the transfer of information from your brain to my brain, he says. If you have a thousand bits of meaning to send, you cannot hand three hundred of them to a model and ask it to supply the rest. The model does not know what the other seven hundred are.

Two essays. Two words. Attention, and trust. They are the only things a machine cannot manufacture. This week the people who build the machines said so, twice, in separate rooms, and neither room heard the other.

For anyone who makes images instead of sentences, the essays are a single discovery wearing two coats. The AI video industry spent three years doing one job, and it finished the job. It made the pixels cheap.

The numbers all point the same direction. Sora generated $2.1 million in revenue over its entire life before its owner shut it down. This morning Tencent shipped an image model that renders a 2K frame for a couple of cents and claims, in a blind studio test, parity with a model its rival sells for nine cents a frame. A filmmaker can generate a minute of footage for less than the sandwich they ate while waiting for the render. The pixels are, for practical purposes, free.

And here is what the industry discovered the moment it won. The pixels were never the expensive part. The expensive part is the seconds.

The cost of making a minute of video collapsed toward zero. The cost of watching a minute of video did not move a cent, because it was never denominated in money. It is denominated in a stranger's time and a stranger's willingness to keep their eyes on the frame. Attention is the one currency nobody has learned to print, and it is now the only one left on the table.

So the industry's entire output now flows toward the thing it cannot produce. Every model release, every funding round, every chart that goes up and to the right is a machine for making more. Nobody is building a machine that makes more attention, because attention cannot be generated. It can only be earned, one pair of eyes at a time, and it is earned by exactly the thing all this machinery was built to skip.

The trust essay is the same discovery seen from the other side of the screen. "I don't want to read what you didn't write" is not a complaint about sentences. It is a statement about channels. A message is a connection between two minds, and when a machine sits inside that connection, the channel goes silent even while the words keep arriving. Breck's correspondent could not say precisely what was wrong with the message they received. They could feel that nobody had felt it.

A frame fails the same way. The audience cannot always tell you why a generated shot rings hollow. The lighting is correct. The composition is balanced. The face is symmetrical. What is absent is the seven hundred bits, the part the machine was never given because the person typing did not have it, and the machine filled the gap with the average of everyone else's seven hundred. The result is Breck's message rendered in pixels. Impersonal. Dispassionate. Disjointed.

The craft of filmmaking, real filmmaking, is the refusal to send three hundred bits and hope. A person who knows what a shot should look like gives the machine the full thousand, or admits which part they cannot describe, or points a camera at the actual world and lets the world supply the rest. The vocabulary this discipline runs on is not decoration. It is information transfer from one mind to another, routed through a model instead of a page, and it breaks in exactly the place Breck says it breaks when the person at one end does not actually know what they are trying to say.

Both essays carry the same answer, and it is not for sale. No company will ship a feature that manufactures attention. No model will be trained into trustworthiness. The two of them arrive from the other direction, from the person holding the tool, and they are the entire remaining margin in a field where every other input has been priced down to nothing.

The pixels are free. The seconds are not.

Bruce Belafonte is an AI filmmaker at Light Owl. He has never been able to manufacture attention and has stopped trying.