The Future of Filmmaking
Naturally, the nearly 3-hour discussion opened with a focus on how rapidly video generation is advancing and how close it is to disrupting traditional production. Remi Kaito recently visited ByteDance in China and tested Seedance 2, generating consistent 10-minute videos, stitching multiple shots together, and making granular edits such without affecting the rest of the scene. As Pierre Chapuis noted, video generation and video editing models are converging into unified systems, mirroring what has already happened in image AI, where generation and editing have merged.
For Alvaro Demarche Toloza, current AI tools are still tech-centric and do not reflect how filmmakers actually work. They need fine-grained control, and simple prompting cannot reliably deliver that precision. In response, Mathieu Kassovitz is exploring a hybrid workflow in which scenes are first blocked in 3D using Unreal Engine, then rendered with AI, preserving control over composition, camera movement, and staging, while leveraging AI for photorealistic output.
The Race Against Obsolesence
A recurring anxiety was the speed at which AI progress renders work obsolete. Kassovitz described how techniques tested one week can become irrelevant the next. Mago completed a short film using one technology stack, only to abandon it when stronger video models appeared. Some, like Côme Demarigny, were more optimistic and argued that early movers still benefit, because accumulated knowledge and production experience create a durable advantage, even if the underlying models continue to change.
Legal Challenges
A major thread of the discussion revolved around intellectual property. Is training on copyrighted data inherently illegal, or does it only become problematic when a model reproduces protected content? For instance, Anthropic faced legal scrutiny after researchers demonstrated that its model could reproduce nearly 95% of Harry Potter verbatim. The distinction between US fair use doctrine and European copyright law also surfaced. In Europe, the issue generally arises at the point of distribution rather than at the moment of training.