Actor Ben Affleck: Disassembling Computers Since Youth, Engaged in Visual Effects Process After Film to Digital Transition
Actor Ben Affleck stated in GQ's One More Question segment that he has been disassembling computers since he was young and began engaging in the visual effects process after the transition from film to digital. He can write "poor Python scripts" and has long used GPUs. He mentioned that this process included machine learning years ago, and before founding Artists Equity, he had a small visual effects company, with some experience coming from working with that group and some from playing video games.
He used convolutional neural networks as an example. The images are written as tensors, containing batches, frame numbers, and RGB values for each pixel in each frame. Convolutional networks perform edge detection and feature extraction within the tensors to determine where the window sill is, then remove the green screen to change the background. He referred to convolutional networks as precursors to transformers, reasoning that the latter can perform more calculations simultaneously.
His visual effects colleagues encouraged him to look at how Google and OpenAI use transformers for video. He said he would call and introduce himself as Ben Affleck, asking if he could come see, and sometimes they agreed. After viewing, he called Matt Damon and said, "We're done," because realistic videos could be generated with just a few keystrokes. For a day or two, he thought he should quickly finish filming the movie he wanted to make.
He then sat with the engineering team to review data and annotations, and the conclusion reversed. The model lacked domain control in filmmaking, the shots were inconsistent, and it could not generate a film simply by typing. He refused to give value to external companies that did not understand copyright and portrait rights, using his own capital to start a second company, spending months building a dataset with numerous cameras, LiDAR, fixed-point equipment, and physical and virtual studios, and applied for patents. The company hired inference experts, data architects, and visual effects personnel.
This company, InterPositive, was established in 2022. Netflix announced its acquisition on March 5, 2026, with the entire team integrated, and Affleck serving as a senior advisor. Netflix disclosed in its 10-Q that the enterprise merger completed in March was valued at approximately $587 million, all in cash. The filing did not name names but corresponded with the acquisition announced at the same time. Affleck stated that the dataset was not trained by capturing H265 clips from peer works. The computational scale for video generation is also larger than that of language models.
This is about content providers buying back training data and tools, not an actor showing off skills. The seller is InterPositive, which refused to sell film data to model labs, and the buyer is Netflix. Money flows from the cash acquisition of streaming media to closed studio datasets, patents, and engineering teams. The beneficiaries are Netflix and Affleck's team, who gain controllable tools while avoiding public scraping controversies. The pressure is on video model companies that still rely on internet video clips for training but cannot access native film data.
Source: Public Information
ABAB AI Insight
Affleck's technical path started in visual effects, not in model labs. After the transition from film to digital, green screen compositing was already using convolutional networks for edge and feature extraction, with GPUs handling the parallel computation of each frame's pixel tensors. He had previously run a small effects company before Artists Equity. In 2022, he founded InterPositive, which does not use compressed clips from peer films for training sets but instead re-samples in real and virtual studios using cameras, LiDAR, and fixed-point equipment. In March 2026, Netflix acquired the entire team for approximately $587 million in cash, and he joined as a senior advisor.
The capital movement occurred after he viewed Google and OpenAI's video transformers. He initially assessed that realistic images could be produced with just a few keystrokes and told Matt Damon that the industry was finished. Upon reviewing annotations and controllability, he discovered that the model did not understand shot continuity, missing shots, background changes, and lighting. External companies proposed collaborations and offered to take over AI, but he refused, reasoning that value would remain in labs that did not handle portrait and copyright issues. Thus, he used his capital to build a second company, treating data production rather than data purchase as an asset.
Structurally, this is similar to the self-built tools in the effects industry post-2012, rather than a technology investment associated with an actor's name. Industrial Light & Magic and Weta kept motion capture and virtual production on the studio side, while the AlexNet generation of convolutional networks elevated GPUs from gaming graphics to image recognition. Netflix rarely makes acquisitions; it previously bought the virtual avatar company Ready Player Me in December 2025 and this time brought the entire InterPositive team into its product technology line. The industry phase has shifted from "models can generate a video" to "who controls the authorized, reshootable native shot data."
Essentially, this is a transfer of pricing power. The pricing power of public videos lies with the scrapers, while the pricing power of native film data lies with those who can build studios, hold portraits, and have distribution channels. Video inference is more expensive than language, with labs trading scale for consistency, while studios use LiDAR and storyboard rules for control. What Netflix paid for was not Affleck's Python, but a set of data and production teams that do not rely on peer films and can integrate into its own production process.
ABAB News · Cognitive Laws
Being able to generate does not equal being able to control.
Data can be purchased or reshot.
If domain rules are not integrated into the model, value remains on set.