Google DeepMind and Primordial Soup put Veo into a working film production

What the partnership actually commits to

Google DeepMind is partnering with Primordial Soup, a storytelling venture founded by director Darren Aronofsky, to produce three short films using DeepMind's generative AI models, tools and capabilities — including Veo, its video generation model.

The structure is specific. Each of the three films is helmed by an emerging filmmaker, mentored by Aronofsky and supported by DeepMind's research team. The first, 'ANCESTRA,' is directed by Eliza McNitt and premieres at the Tribeca Festival on June 13, 2025. That's a dated, public deliverable, not an open-ended research collaboration.

Hybrid production, not fully generated film

The announcement is careful about what the technology does. 'ANCESTRA' uses what Google calls a hybrid production model, integrating live-action performance with advanced tools and emerging AI research. The model isn't replacing the shoot; it's being placed alongside filmed footage.

That framing matters. It positions Veo as one element inside an existing craft workflow rather than a one-click replacement for it, and it acknowledges that the interesting work happens at the seam between live-action and generated material — where a director decides what each is good for.

Filmmakers as a research channel

The partnership's vision is to put the world's best video generation model into the hands of top filmmakers, pushing the boundaries of the technology for more creative and emotional storytelling.Montana Labs

Read plainly, this is a feedback loop. Google gets its models used by demanding practitioners on real projects, with its own research team in the room. The phrase 'emerging AI research' suggests the tooling isn't finished — the films are as much a testbed as a showcase, and the friction filmmakers hit becomes product input.

The implication: model quality now gets judged on set

By routing Veo through Aronofsky's mentorship and a Tribeca premiere, Google is choosing to have its video model evaluated by filmmakers and festival audiences rather than by benchmark clips. The bar shifts from technical fidelity to whether a director can bend the tool to a scene.

For teams building on generative video, the signal in this announcement is the pipeline, not the pixels: the useful pattern here is embedding a model in an existing craft workflow with expert users producing dated, public work, then feeding what breaks back into research.