Indico Data elevates insurance AI with out-of-the-box capabilities for global carriers. Read more.
  Everest Group IDP
             PEAK MatrixĀ® 2022  
Indico Named as Major Contender and Star Performer in Everest Group's PEAK MatrixĀ® for Intelligent Document Processing (IDP)
Access the Report

Unstructured Unlocked season 2 episode 10 with Andrew Carr of Cartwheel

Watch Indico Data CEO Tom Wilde step in as co-host alongside Michelle Gouveia, VP at Sandbox Insurtech Ventures, in season 2 episode 10 of Unstructured Unlocked with Andrew Carr, of Cartwheel.

Listen to the full podcast here: Unstructured Unlocked season 2 episode 10 with Andrew Carr of Cartwheel

 

Tom Wilde: Welcome to another episode of Unstructured Unlocked. I’m your host, Tom Wilde.

Today, we have a really fascinating guestā€”Andrew Carr, the founder and chief scientist of Cartwheel. Cartwheel is a generative motion company, and we’ll let Andrew tell us more about that. This will be an interesting episode where weā€™ll dive into the art of the possible, maybe taking a different view of unstructured data. Weā€™ve spent a lot of time on this podcast discussing how to extract intelligence and information from unstructured data, but what if we take the opposite approach?

Andrew, welcome to the show.

Andrew Carr: Thank you. I’m excited to be here.

Tom Wilde: Tell us a little bit about your background, especially as it relates to the explosion of new technologies weā€™re all trying to interpretā€”generative AI, agent architectures, and so on. And what is the core problem that you’re solving at Cartwheel?

Andrew Carr: Certainly. Iā€™m still somewhat new to the fieldā€”I started working in machine learning in 2016. This was just before transformers became mainstream.

My background is actually in applied mathematics, where I focused on high-dimensional geometries and probabilities over geometries. I found it fascinating, but it was a niche field with only a handful of people deeply engaged in it.

I eventually realized that my expertise had direct applications in machine learning. My early work in the field involved medical imaging, where I focused on reducing the cost of MRIs. Then, I worked on audio technologies before moving into predictive physics at Lyft.

At Lyft, I was on the prediction team, where we worked on forecasting how cars move relative to other cars. Specifically, I worked on predicting the trajectory of the “follow car”ā€”the vehicle directly behind another. At the time, we were experiencing too many harsh braking events, which, while safe, made for an uncomfortable ride.

I really enjoyed that work. However, I quickly realized that self-driving technology would take much longer to reach maturity than I initially thought. I even began to suspect that general-purpose autonomous driving might be what we call an ā€œAI-completeā€ problemā€”essentially, we may need to solve intelligence itself before we can fully solve self-driving cars.

That realization led me to shift my focus to intelligence more broadly, so I joined Google Brain to work on AI problems. After that, I became a fellow at OpenAI, where I worked on program synthesisā€”teaching models to write code.

At OpenAI, I was working on a model that could write Python really well. As I looked for useful applications, I discovered Blenderā€”a 3D animation and modeling software thatā€™s free and open-source. Blender has a built-in Python interpreter, meaning you can write Python scripts to automate tasks inside the software.

Since I didnā€™t know how to use Blender myself, I decided to hook up my AI model to it. The model could generate Python scripts that would then execute inside Blender, allowing me to create animations and effects that I wouldnā€™t have been able to produce manually.

Tom Wilde: So Blender is primarily a three-dimensional engine?

Andrew Carr: Yes, exactly. There are several 3D enginesā€”game engines like Unity and Unreal, or 3D modeling engines like Maya and Blender. These tools let you create 3D objects, place cameras and lights, and render realistic images or animations.

The lightbulb moment for me was realizing that, unlike text, images, and video, animations generated by AI werenā€™t easily editable. Artists donā€™t get the same level of control over AI-generated animations as they do with traditional tools. I saw an opportunity to change that, and thatā€™s what led me to start Cartwheel.

Today, at Cartwheel, users type in text, and our model generates 3D motion animations. Crucially, these animations are fully editable. You can adjust the characterā€™s movements, speed, or appearanceā€”all in a way that is useful to professional artists. Our goal is to accelerate creativity rather than replace human artists.

Tom Wilde: That really resonates. There are typically two ways to start a businessā€”you either have deep expertise in a specific business use case, or you start from the technology side and work toward a viable commercial application. It sounds like you approached this from the technology side.

Tell me about your journey from realizing the potential of your technology to identifying the right market problemā€”the so-called ā€œkiller appā€ for Cartwheel.

Andrew Carr: Absolutely. Once I left OpenAI, I was working solo in my basement, trying to refine the technology. Then, I got an email from my former colleagues at OpenAI.

They said, “We just interviewed someone who pitched us an idea identical to what you’re working on. You should probably talk to him.” That person turned out to be my now co-founder, Jonathan, who has deep expertise in animation.

Jonathan has worked as an animator, was part of the motion design team at Google, led rebranding efforts, and even started his own agency. He understands exactly what animators need, which made him the perfect partner.

Together, we identified gaming as a major market. Game developers care deeply about the craft of animation and storytelling, but they also need speed. Cartwheel allows them to move faster while maintaining high-quality results.

That said, weā€™re still in the early days. We only recently launched the tool and are just starting to see how users adopt it.

Tom Wilde: So the core value proposition is enabling anyone to animate, not just trained professionals?

Andrew Carr: Exactly. Right now, animation is expensive and time-consuming. If I wanted to create a short animated film, it could take years and cost hundreds of thousands of dollars. With Cartwheel, anyone can type in a description and generate an animation in seconds.

Itā€™s also useful for professionals. Animators can take on more ambitious projects or increase their output by automating repetitive tasks. Our goal is to make everyone an animator, whether theyā€™re an individual storyteller or a seasoned professional.

Tom Wilde: I often describe generative AI as a new kind of programming languageā€”one that allows people to program using natural language instead of code. Would you agree with that?

Andrew Carr: Absolutely. Iā€™d say it raises the floor of what people can do. Everyone now has access to capabilities that were once reserved for highly trained professionals.

At the same time, it raises the ceiling for existing professionals. Even experienced developers can use AI tools to write better code faster. At Cartwheel, weā€™ve seen junior engineers save weeks of work by using AI-assisted programming tools.

Tom Wilde: A big topic in AI right now is whether large language models will become commoditized. Do you see Cartwheelā€™s technology as defensible?

Andrew Carr: Yes. Weā€™re fortunate in that motion is its own data type, separate from text, images, and video. General-purpose AI models like GPT-4 simply canā€™t generate motion because they werenā€™t trained on that kind of data.

At Cartwheel, weā€™ve built everything in-houseā€”our own training stack, models, data, and hardware. That full-stack approach gives us a competitive edge.

More broadly, I believe specialized models will always outperform general-purpose models when trained on domain-specific data. Thatā€™s a key insight for companies in insurance, finance, and other industriesā€”if you have proprietary data, training your own models will always give you an edge.

Tom Wilde: This has been a fascinating discussion. Itā€™s exciting to see how Cartwheel is pushing the boundaries of generative motion.

Andrew Carr: Thanks, Tom. I appreciate the conversation.

Tom Wilde: Thanks for joining us, and thanks to our listeners for tuning in to Unstructured Unlocked. Weā€™ll see you next time.

Check out the full Unstructured Unlocked podcast on your favorite platform, including:

Subscribe to our LinkedIn newsletter.

Resources

Blog

Gain insights from experts in automation, data, machine learning, and digital transformation.

Unstructured Unlocked

Enterprise leaders discuss how to unlock value from unstructured data.

YouTube Channel

Check out our YouTube channel to see clips from our podcast and more.