The Algorithm Doesn't Want a Good Movie. It Wants a Watchable One.

In 2019 I was cutting a documentary about a man who restored church organs. Four years of his life, on my timeline, twelve minutes at a time. The film was never finished. The money ran out because the platform that had optioned it changed its commissioning strategy mid-production, and the new strategy was not "quiet films about old men and old instruments."

I went dark for six months after that. I want to be honest about the six months, because when people ask what happened to the industry, they usually want a technical answer about generative video, and the technical answer is not the interesting one.

Here is the interesting one. The tools did not kill my film. The metric did.

What a recommender actually optimizes for

A streaming algorithm does not evaluate whether a film is good. It cannot. Goodness is not a signal it receives. What it receives is a sequence of behaviors: plays, pauses, exits, completions, rewatches, and the position of the scrubber at the moment the person leaves.

From those it derives one thing: how long did this hold attention relative to its cost to acquire. That is the whole value function. Everything else, the awards, the reviews, the cultural weight, is noise in that equation.

This is not new. Broadcast television ran on the same calculus with worse instruments. But broadcast had a floor, a schedule, a slot, and a human being in a room deciding that a film was worth the slot because it was good. That human being was the entire difference. They had taste and they had authority and the two together could overrule a number.

The recommender has no such room. There is no slot to argue for. There is only the feed and the probability that a given user stays.

Why generative video arrives here, not somewhere else

Which brings us to the thing everyone wants me to write about. Generative video is not coming for cinema because it can make images. Cinematography was never the bottleneck. A competent camera department can make beautiful images on a Tuesday.

Generative video is coming because it collapses the cost of coverage. The unglamorous, expensive, hours-long process of shooting enough material to have choices in the edit. That cost is what makes a film slow and deliberate and expensive and therefore worth arguing about.

When coverage is free, so is everything downstream. And when everything is free, the only differentiator left is the metric. A system that can generate a season of television in a weekend will be pointed at the metric, not at the art, because the metric is the only thing the system can hear.

The craft argument, stated properly

I am not going to tell you that AI cannot make something moving. It can. It already has. There are shots generated in the last year that would have taken a full crew a week and a lot of money, and some of them are beautiful.

What I will tell you is that a beautiful shot is a unit. Cinema is not a collection of units. It is an argument made through sequence, and the argument is made by a person who has decided what to withhold. Withholding is the entire craft. The cut that skips the kiss. The reaction shot held two seconds too long. The scene that ends before the line lands.

A model trained on completion will always give you the line. It has no instinct for the withheld thing, because the withheld thing is an absence, and absences do not appear in training data.

A note on the tools themselves

The generative video tools are not the enemy and I dislike the framing that they are. I have used them. They are astonishing. What frightens me is not the tool, it is the budget meeting where someone asks why we need a crew when the tool produces the coverage for free.

That meeting is happening in every studio right now, and it is being won by the tool, not because the tool is better but because the tool is cheaper and nobody in the room has the authority to say the film is worth the difference.

What survives

The films that will survive this are the ones that can justify their existence without a retention graph. There are more of those than the industry believes. They are just not in the business of being watched in fragments, and the business has reorganized itself around fragments.

I do not have a solution. I have a practice. Shoot less, mean more, and make the argument in the edit rather than in the coverage. If you cannot afford to be expensive, be deliberate. Deliberation is the one cost the models have not collapsed.

The frame is always lying.