InterPositive and the Filmmaker-Approved AI

Netflix’s acquisition of Ben Affleck’s InterPositive came with a remarkably convenient story about artificial intelligence. Filmmakers had built it, filmmakers would control it, and the technology would protect the creative judgment that other AI companies supposedly failed to understand.

A respected actor and director was bringing AI into Hollywood on Hollywood’s terms.

I can see why Netflix wanted that story. I am less convinced that the technical explanation supports everything the story invites us to believe.

I build AI tools for video. When Affleck talks about cinematic understanding, proprietary data, and adapting models, I want to know what changed in the system. His explanations contain enough engineering to sound specific, but the important distinctions keep slipping. The product concerns post-production changes to filmed material, so those are the operations against which its claims should be judged.

The base model is already there

In the interview clip published by Bloomberg, Affleck explicitly describes taking existing open video models and continuing their training. He says those models had learned from online video, and that his team taught them to meet cinematic standards.

So the starting point is not an intelligence created entirely from material his team recorded. It is an existing model, with existing capabilities acquired through earlier training.

That matters because the proprietary-data argument is doing ethical work as well as technical work. It gives the impression that this system avoids the uncomfortable dependence on other people’s creative output associated with general video models.

Additional training does not establish that.

Recording your own footage establishes something about the material you added. It does not establish the provenance of the base model’s data, and it does not remove capabilities learned from that data. If the concern is that a system learned from actors and filmmakers without their consent, the relevant question includes what it learned before InterPositive began adapting it.

Open weights make a model available to work on. They are not a certificate that everyone represented in its training data consented.

That is the gap I wanted addressed. Explaining how you resume training does not address it.

What became cinematic?

There is nothing inherently unimpressive about adapting an existing model. A deliberate dataset and good training can produce a useful specialization. But “cinematic understanding” bundles together several capabilities whose relationship needs demonstrating.

Does the model preserve an actor’s performance while changing the background? Does it maintain lighting across a shot? Does it recognize why a cut works? Does it understand an instruction expressed in filmmaking terminology? Does it produce a more controllable image?

Those are different tasks. Improving one does not establish the others.

I also find the suggestion that filmmaking expertise arrives as a final layer too convenient. The model already learned visual relationships from the material used to train it. InterPositive may improve some of those relationships, add controls, or make particular operations more reliable. Which ones?

The description should make it possible to answer that question without borrowing the authority of the word cinema.

Knowing what directors call something is not the same as doing it well. Neither is being trained by a director.

Useful software is not a new kind of intelligence

A production workflow can be a substantial achievement. Connecting models to editing tools, maintaining references to source material, making changes reversible, and giving artists controls they can actually use are difficult problems.

I have no objection to someone building that on top of existing research. Much of useful software is built that way.

What bothers me is the inflation of the achievement. If the product is a well-designed workflow around adapted models, explain the workflow. If the adaptation produces better results, show the improvement. We do not need to pretend that everyone outside Hollywood lacked the crucial understanding of images until a filmmaker supplied it.

Nor does a specialist product automatically outperform a general model connected to good tools. Specialization is a method, not a result.

The comparison has to include what the alternatives can actually do. A multimodal system is not merely a text assistant that knows some camera terminology. A video-generation system does not become incapable of visual correction because it was trained for broader tasks. A smaller adapted model might outperform it on a particular operation. That is something to demonstrate, not something the product category establishes.

Protecting whom?

The language about protecting artists deserves scrutiny of its own.

Keeping creative control with a director may be valuable. It does not tell us what happens to the people whose work the director can now complete with fewer paid hours. More control for one artist and less work for another can happen in the same production.

That does not make the tool bad. It makes “protecting artists” an incomplete account of its consequences.

Which artists? Protected from what? Who receives the savings? What happens to consent, compensation, and bargaining power when a task becomes cheaper to automate?

A filmmaker can sincerely care about the industry while selling technology that changes the distribution of work within it. Sincerity does not tell us who gets the savings or whose work becomes less secure.

Affleck’s GQ interview makes the institutional preference quite explicit: he wanted the technology developed within filmmaking and regarded Netflix’s position inside that industry as important. That is a business choice as well as an ethical position. It places the technology with an established buyer and distributor of creative labor.

A studio’s familiarity with that labor does not make its interests identical to those of everyone it employs.

What Netflix bought

My view is that Affleck’s credibility is a substantial part of this product’s value, and that the public pitch uses it to carry claims the technical explanation has not established.

He can get the meeting. He can recognize production problems. He can make AI adoption feel less like an intervention from outside the profession. Netflix gains a tool, a team, and an unusually effective person to explain why the tool belongs there.

That is a serious commercial advantage. It should not be confused with a demonstrated technical moat.

Calling the acquisition a public-relations move alone would go further than the evidence allows. Calling the technology filmmaker-approved, however, does not go nearly far enough to establish that it is better, ethically cleaner, or protective of the people whose work it changes.

The public-relations value is readily visible. The comparative technical value remains much harder to inspect.

That imbalance is why I find the pitch irritating. The explanations become expansive around human creativity and oddly imprecise around the model.

I would like to see the same difficult shot attempted with InterPositive and with a capable alternative workflow. Preserve the performance. Change the background. Keep the lighting and detail consistent. Show the failed attempts, the controls, and what the artist still had to repair.

That would tell me more than another assurance that the technology was built by people who love cinema.

Affleck has explained how his team continued training a model and built a product around it. He has not, in those explanations, shown that the adaptation establishes the ethical separation or special cinematic competence being sold alongside it.