Fine Tuning Models Ai And Hollywood A Conversation With Oxen S Founder Greg

Fine-tuning models, AI and Hollywood: A conversation with Oxen’s founder Greg Announcing our Series F . Learn more

Community

Fine-tuning models, AI and Hollywood: A conversation with Oxen’s founder Greg

In this interview with Greg, founder of Oxen AI, we discuss his journey from IBM Watson to building fine-tuning infrastructure.

Authors

Madison Kanna

Last updated February 5, 2026

Share

TL;DR In this interview with Greg Schoeninger, founder of Oxen AI, we discuss his journey from IBM Watson to building fine-tuning infrastructure, how Hollywood studios are using AI for pixel-perfect VFX work, the furniture industry's surprising use case that saved $40K per project, and why pairing version control with compute infrastructure is transforming how companies deploy custom models.

I want to talk about what Oxen is, but can you tell me more about your background? I know that you've been training models for years. Yeah. So I've been working in AI since 2012, the very early days of deep learning. Right out of school, I joined a startup that was doing deep learning before any of the TensorFlows or PyTorches of the world. So we were writing our own neural net libraries from scratch in C++ training ConvNets. And then there was the AlexNet moment that really blew our minds, when this convolutional neural network beat all of the handcrafted features on ImageNet. So we just went all in on deep neural networks. That was called Alchemy API, and that company eventually got acquired into IBM, into the Watson Group. Our job was going into IBM Watson that won Jeopardy back in the day and replacing all of the models that were logistic regression models or maximum entropy models with deep neural nets that could do the same but generalize better to the data. IBM would go sell a big contract to be able to do some natural language processing task, and then they'd bring the requirements back to the engineering team. We were part of a fast domain adaptation team that would have to get our models to work in Korean on a financial dataset. We don't even speak Korean, but as long as we get the data, we're able to customize the model and then hit whatever accuracy the customer asked for. So that was in the early days of my career and I did a bunch of reps of just going from a research idea to a production deployment. And almost every single time it just came back to what data we were feeding into the model. That kind of sparked the idea of Oxen. We actually started as a data management layer, version control layer for machine learning datasets. Because it's really just garbage in, garbage out or good data in, good data out. We started to see some traction there, but what's really been picking up is fine tuning some of these foundational models or open source models on the datasets. So pairing the data with the compute and infrastructure and letting a company own their own model end to end is really where we're seeing traction now. Oxen started as managing the data? Yeah, so we started a version control tool to replace Git. We actually rewrote the internals of Git to be able to scale to terabyte-size datasets. This was in the pre-ChatGPT era where these pre-trained models would take a ton of data and we were just saving them in S3 buckets with no sense of version control, no sense of what data trained what model. So that's where we started. Could we build a distributed version control that actually works? If you've ever hit the limits on GitHub, it's like two gigabytes of data that you can put in there. Ours scales to—I think our biggest deployment has 20 terabytes of data in it, and thousands of engineers collaborating on the same repository. You were doing this for a while, but then eventually it changed into what Oxen is today? It turned out that a lot of people were using the version control tool to fine tune their own models as a natural use case for it. Because you have to pair what dataset went with what model, and you might have a bunch of models deployed out in the edge or models deployed in your own infrastructure. And being able to tie that specific version of the dataset to the model weights – both the dataset and the model weights can get pretty hefty. The DeepSeek models are 600 gigabytes alone, the model weights. So how do you store and manage the connections between those? Your customers were basically using this in this way and you started realizing that's what they wanted? Yeah. And they were having to home roll a lot of their infrastructure to build on top of our version control stack. So they would take it, wire it up to their compute infrastructure, and have their own training loops. The time to value for some of those customers was the time that it took their engineering team to integrate the version control tool with their training infrastructure. So we were like, what if we just paired it with a training infrastructure that kind of came out of the box? One click, go from your dataset to a fine-tuned model and then eventually deploy it into your production setting. Can you talk about some customer use cases or someone who might use you, what their use case would be? Yeah, so we work across a couple of different modalities. There's the classic LLM use cases for text in, text out, classification, agents making tool calls or name your LLM use case. What we're seeing is a lot of people who would build the MVP of their application on OpenAI or Anthropic just to see if this is possible. And then they start to run into either accuracy because it's out of distribution, or speed or latency requirements that they need, or even just privacy. I just don't want this data leaving my infrastructure and going to some other cloud. So they either take the data that they've collected in production or do some hand labeling, fine tune their own models and now they have full control over the experience from a latency, cost, or accuracy perspective. So that's on the LLM side. And then on the diffusion model or image generation, video generation side, we're seeing a lot of people who care a lot about the quality of the output. There's the meme of AI slop going around with Sora and all of these image gen models. We're based in LA so we have some clients that are Hollywood grade. They need it pixel perfect on every single frame and they're gonna have their VFX coordinator looking at it and...