A 40-Person Company With 40,000 Paying Customers Wants to Replace the Camera With AI

The episode of No Priors starts as a bit. Sarah Guo welcomes Joshua Xu, then slips into the cadence of the Huberman Lab, introducing herself as a professor of neurobiology and ophthalmology who is here to discuss science-based tools for everyday life. Xu is lost. He asks whether this is the show that interviews technology people like Garry Tan and Alexandr Wang. Guo tells him that version is only for humans. The fake open is the product: voice, likeness, and tone generated well enough to pass as a podcast intro, built by the company Xu co-founded.

Xu is co-founder and CEO of HeyGen, started roughly three and a half years before this taping. He studied robotics at Carnegie Mellon and joined Snapchat in 2014, staying about six and a half years. Early on he worked on machine learning for Snapchat ads, ranking and recommendation. His last two years there were on AI cameras, the software that sits on top of the phone lens.

The turning point, he says, was 2018, when Snapchat shipped the baby filter and a Disney-style filter. It was the first time he watched a computer generate something that did not exist in the real world. Snapchat is a camera company, and almost everyone on it creates through a mobile lens. Xu wanted to replace that lens. HeyGen's mission is to make visual storytelling accessible to all, on the bet that AI could become the new camera.

Elad Gil asks what replacing the camera actually means, given that he uses his own a lot. Xu's first career was spent making people more comfortable creating through a mobile camera. Plenty of people still cannot make good content that way. If you remove the camera, he argues, you remove the barrier to visual storytelling and move the whole production stack forward.

When the company started, the team took video production apart and found two pieces: camera and editing. Camera is A-roll, the human spokesperson, which avatars would later stand in for. Editing is B-roll, voice-over, music, transitions, animation. Customers told them editing is not that expensive; it is a fairly standard service. The camera is not. A CEO who wants to record something may need two weeks of lead time, a crew, and a studio. Two minutes of finished footage can take twenty minutes to shoot, because people have to remember the script.

That camera step, Xu says, is what blocks a lot of businesses from creating new content, which is why HeyGen started by replacing it with avatars. He is not trying to settle every possible future, from Zoom stand-ins to entertainment. The near-term path is the entry point of content: generate the footage, apply AI editing, and assemble the final video. Push the models far enough, especially with GPT-4o and other multimodal real-time systems, and streaming generative video could eventually take over conversations that today have to happen live.

In 2024 the product is still mostly asynchronous. Xu groups usage into three jobs: create, localize, and personalize. On create, someone picks a stock avatar or makes a digital twin, chooses a template, types a script, and generates a video. That loop works best for product explainers, how-to videos, learning and development, and sales-enablement training, the kinds of pieces that used to wait on a shoot date and a post queue.

Localization takes an existing video and turns it into more than 175 languages and dialects, keeping the speaker's voice and face rather than swapping in a mismatched dub. Personalization is video messaging at scale: the same skeleton, different names and details. Xu calls HeyGen a horizontal platform. Marketing, sales, internal webinars, and training teams all show up, and the creative uses keep branching.

One favorite example is a McDonald's campaign that let people send a message to a family member in another language. A generated clip in the conversation gushes about love running out of words. Xu's aside is the point of the product: "AI for everyone, grandma and grandchildren alike." Guo notes that this is a big brand using the tool in a public, consumer-facing way, which is a harder quality bar than an internal training video.

Quality is the first priority across product, business, and research. Xu's framework is an invisible line, call it 90. Anything below that is essentially unusable, because it cannot replace the live production process customers already have. Avatars, he says, are above the line today, which is why they can stand in for a real camera and unlock volume. There is still plenty of headroom: full-body avatars, and the ability to bring more kinds of elements into a generated video.

Full-body generation is the release he is most excited about. Historical avatar work focused on the upper body; gesture and body motion were hard. Academic research has shown the problem is tractable, and HeyGen is trying to take that last mile. Streaming avatars are the other bet. GPT-4o improved real-time text and voice, and an avatar can sit on top as the visual layer for those applications, including virtual sales and support that take customer calls.

Customers do not want full-body motion for its own sake. Xu draws a quality spectrum. On the left, learning and educational content is one-to-many broadcasting; the avatar can be still and professional, and the bar is lower. On the right, high-end marketing, especially ads, has to be dynamic because motion drives engagement and return. Full-body rendering is how avatars move toward that right-hand side and unlock broader marketing and sales. Guo adds a concrete shot: news packages often use a walking-and-talking presenter, a standard camera move that full body would finally let you generate.

The stack is three models: text, voice, and video. HeyGen uses OpenAI's ChatGPT for text, and the same model is the brain of an internal orchestration engine. Voice comes from OpenAI and ElevenLabs. The entire video stack is built in house: avatar creation, rendering, and B-roll generation. Over time Xu expects text, voice, and video to chain into one multimodal model, but that is not how they ship today.

Full-body work exposes a concrete hard problem: connecting voice to gesture. Historically the team trained a TTS model, then fed its output into a video model, and the two sides never really met. Joint multimodal training is what might build that connection underneath. Gil asks how this differs from a generic text-to-video system like Sora. Xu's north star is what a business actually wants: quality, control, and consistency.

That points to two technical paths. One is end-to-end generation, a Sora-style pass that emits the whole clip at once. The other, the path HeyGen took, is to disassemble video into A-roll (the avatar) and B-roll (voice-over, music, transitions), solve each component, and assemble them with the orchestration engine. Logos and fonts should not be generated; they have to be exact. Brand style and color mapping have to be learned from the customer. The component path leaves room to build that system. Sora, in this view, is a partner: a component generator that can feed the same engine.

Research direction is a mix of academic papers, customer problems, and a clear map of what the models cannot do. Every AI model has limits. The product question is how to design around those limits so you avoid the weak spots and amplify the strengths. Video translation is the example. Underneath it is a lip-sync model, plus voice cloning and ChatGPT for the translation. Combined, it preserves the speaker's natural voice and expression in a way traditional dubbing never could, which is a new localization experience rather than a better subtitle file.

Gil puts the fear on the table: deepfakes, stolen likeness, election abuse. HeyGen does not allow political or election content on the platform. Policy forbids unauthorized use of someone's image. Safeguards include advanced user verification, live video consent, a dynamic verbal passcode, and rapid human review of avatars created on the system. Xu says safety is critical to the business, that the company is partnering across the industry on tools against misinformation, and that the guards sit inside the creation flow rather than after it.

Gil also sketches the constructive version. A candidate could send each voter a short clip about the issues that person actually cares about, a hyper-personalized campaign that is technically in reach if the deepfake side is locked down. Xu does not lean into electoral use; the product ban is the current answer. The broader claim he will stand on is that once people can generate engaging, authentic video, they will simply make more of it, because the world is already video-first and production is still too slow and too expensive.

If businesses can generate authentic video, Xu argues, they will use video more to grow. The bottleneck today is that video is expensive and takes weeks or months, which shapes how companies think about communication, marketing, and sales. Cost and time savings are only part of the pitch. The more interesting pattern is customers unlocking work they could not do at all before, which is where he locates the real business outcome.

Real-time avatars are already on HeyGen, and the team is making them faster, toward virtual SDRs and support agents that take customer calls. In about two years he does not think it will be crazy for a lot of today's asynchronous avatar pipelines to become real-time streaming. In about five years, generating an entire video in real time starts to look plausible. That is the setup for the claim he wants to land: generative image is still image, but generative video is not video. It is a new format.

We treat a video as an MP4, an immutable file. Two people on Instagram may see different ads, but if the ads come from the same business they are often looking at the same file. It does not have to be that way. If Xu likes avocado, the Coca-Cola ad in front of him should be a Coca-Cola-and-avocado story; Gil should see something else. That is impossible while making video is expensive, and obvious if ads can be generated from user attributes. Today's player maps to one MP4. A future player could take those attributes and render the best version on the fly.

Guo reaches for education. YouTube is one of the largest learning tools in the world, and it is still the same static video for everyone, even though Bloom's work and common sense say personalized teaching works better. Personalized video has just been too expensive. Xu's customer example is closer to HR than to a classroom: Publicis Groupe generated more than 100,000 thank-you videos for employees worldwide, localized by language and personalized with name and the story of joining the company. Historically that message was one recording from the executive team. Now the communication can be personal at that scale.

The hard part of building video models, unlike many other models, is integrating aesthetics. Video generation is not only a math problem; it is about making something customers love and will watch again. A model with a more optimized cost function does not automatically produce a better picture. That makes evaluation painful and important. HeyGen leans on in-product signals such as A/B tests, because only the customer can judge. The process is not differentiable. They have to wrap a system around it and feed those outcomes back into training.

Gil asks whether that method came from Snapchat. Xu says it is very similar to camera software. You can invent objective scores for lighting and resolution, then learn that higher resolution is not the same as better image quality. The iPhone does not always win on resolution against other phones, and it still produces the pictures most people like. Those early Snap lessons now sit inside HeyGen's model loop.

The company is a little over 40 people and serving more than 40,000 paying customers. Guo flags the ratio: about a thousand customers per employee. The mix is not typical AI early adopters. Xu lists European manufacturers, small businesses, global nonprofits, and Fortune 500 companies, which he takes as a sign they are solving a real production problem. They are hiring across product, design, engineering, AI research, and go-to-market.