Fred interviews Greg Schrodinger on Oxen, AI video workflows, and the future of creator tooling Fred sits down with Greg Schrodinger, founder of Oxen, to trace his path from early deep learning and IBM Watson to modern generative AI video production. The conversation covers how Oxen combines model access, data storage, training, and studio services to help creators and teams build better AI workflows. They also dig into real production examples, including a Subaru campaign, a Frank’s RedHot Super Bowl ad, and collaborations with Neil Blomkamp, showing how infrastructure, iteration, and creative control all come together in practice. Key topics Greg’s AI origin story He started in AI in 2012, back when deep learning was just taking off. His early work at Alchemy API focused on small language models for translation, sentiment, and entity extraction. The company was acquired by IBM Watson, where the team helped modernize systems with neural networks. How Oxen began Oxen grew out of a practical need for storing training data from motion capture and fitness applications. Early products centered on two building blocks: data storage and GPUs for training. That foundation later expanded into a broader platform for generative AI applications. What Oxen does today Greg describes three major pieces: models, a data storage layer, and a model router. The storage layer emphasizes provenance and versioning, which matters in production workflows. Customers use the API to build their own tools, pipelines, timelines, and canvas-style interfaces on top of Oxen. Training custom models on real production data Greg walks through a small fine-tuning demo using a “50 Cent” concept dataset. The example shows how open source models like LTX and MiniMax can be trained with captions and video frames. He explains how LoRA-style training lets creators fine-tune likeness, style, and locations. Why data reuse matters in video generation Generated clips can become training data for the next iteration. He argues this makes cheaper, high-quality open source models strategically important. He also notes the risk of building on closed systems that may stop supporting a model mid-production. Oxen’s target users The company mainly serves builders who are creating their own tools on top of Oxen’s infrastructure. Their customers include studios and teams working on everything from micro-dramas to feature films. Greg emphasizes that the UI is intentionally light so customers can build their own experiences. The studio side of Oxen Oxen also operates as a studio, making shots for clients while feeding lessons back into the platform. Greg says the studio and platform are roughly fifty fifty. This loop helps the team improve product, model access, and workflow design based on real work. Neil Blomkamp partnership Greg explains how Neil Blomkamp became an advisor after early collaboration through music video and production work. Neil’s projects at Synaptic and Barley helped show how AI could be used in commercials and feature film contexts. The partnership highlights how serious filmmakers are adopting these tools for production. Working with unrestricted video models Greg talks about the need for fewer creative filters in certain film and commercial workflows. Some platforms block violence, gore, or celebrity use cases, which can get in the way of legitimate production needs. He says Oxen can work with model providers to remove filters when proper authorization is provided. The Subaru campaign and the realities of AI production The team built AI-heavy shots for Subaru, including puppy and car scenes. They learned that maintaining dog size, perspective, and consistency across frames is harder than it looks. The final workflow relied on storyboards, keyframes, compositing, and a lot of iteration before video rollout. Vibe coding and automating creative workflows Greg shares a side project that automates launch videos for new models. He used a Claude skill to research models, generate shot lists, evaluate outputs, and stitch clips into a finished montage. He also vibe coded a front end in less than twenty four hours to scrub timelines, preview clips, and inspect generated assets. How AI is shifting product development Greg says design and implementation are now close in speed, and sometimes implementation is faster. The emerging workflow is to prototype quickly with a rough model, then bring in human polish later. He sees this same pattern in film, where a low fidelity version can be enough to validate the story before making it shine. Community, in-person events, and the human factor Fred and Greg close by reflecting on the value of community, live events, and working together in person. They connect that to the continued appeal of movie theaters and physical media. The broader theme is that even in an AI-first world, in-person creative communities still matter.