AI production · 6 min read

When production is nearly free, taste becomes the bottleneck, and the senior skill is knowing when to stop generating.

Grip, the tyre series in my portfolio, has a cast spread across [N] countries and three generations. Before AI, that meant casting calls, location scouts and a budget few tyre companies would sign off for films that barely mention tyres.

So yes, AI made it faster. It also made me slower, and the slow part is the one worth paying for.

The speed is real

A pub somewhere in Europe, an old farmer in a worn canvas jacket, warm light from the lamps over the bar: available on request, in minutes, as many times as you like.

That changes production budgets, and it changes the question a creative director is there to answer.

When a shot took a day to set up, the hard question was can we get it? Now you can get almost anything a hundred times over, and the hard question is which one? That question doesn’t speed up with the machine. It slows down, because there are more answers to look at.

Where the time went

The farmer is a good example. He’s meant to be tired and a bit bitter, halfway through an argument about a new law he doesn’t care about. The smile is supposed to arrive late, and only once, when he gets to his tractors: five in forty years, sixty tyres.

In the first trial takes he smiles almost the whole way through. He glances over his shoulder at the bar, grins into the camera, and turns into a nice old man in a nice pub. Every take was technically fine, with good light, believable skin and no melted hands, and every take was wrong. If he’s warm from the first second, the smile at the tyres has nothing to land on.

That’s where the days go. Roughly:

The third line is the job. Generating is cheap and rejecting is expensive.

Regression to the mean

Generative models are trained on what already exists, so their first answer is the most probable one. Ask for an old man in a pub and you get the most probable old man in a pub: friendly, a little twinkly, ready for a beer advert. The camera has the same habit. (Left alone, AI video would turn every scene into a drone shot over Iceland.)

Unchecked, everything drifts toward that average. You can already see the same polished, vaguely cinematic, forgettable look everywhere.

The only thing between your work and that average is someone willing to look at a perfectly decent output and say “no, again”, and then to know when to stop, because “again” can go on forever too.

The new senior skill

We used to hire senior people partly for craft, because they could make the thing. That’s less scarce now.

Judgment under abundance is what’s scarce: knowing what you’re looking for before you generate, recognising it when it shows up as take 41 of 60, and stopping there instead of making 60 more in case something better turns up.

Architects have their own version of this: when you can draw a hundred options in an afternoon, the skill is knowing which one deserves to be built.

So the tools made me faster, and I spent the time I saved making a farmer less friendly.

Which turns out to be the expensive part.