Michael Lingelbach and the problem of a believable performance

Michael Lingelbach has a useful qualification for building digital actors: he has been an actual one. Before the Stanford research, before the fundraising, before the talking babies acquired microphones, there was theatre. His favourite show to have acted in was Grease. Singing is a talent he still claims. The distance between a school stage and an artificial intelligence company turns out to be shorter than the scenery suggests.

At Hedra, the company he co-founded with Alexander Bergman, a character can begin as an image and an audio track. The difficult part is making the resulting performance feel as though the speaker means it. A mouth can move in time with a sentence and still leave its owner looking like an unusually expensive answering machine. Lingelbach has chosen to build around that awkward gap.

His career offers two ways into it. The performer pays attention to expression and intention. The researcher asks how to represent a world in which actions have consequences. The founder has to make the result useful enough for someone to return tomorrow. All three jobs now share a desk.

A stage, then a simulated kitchen

Lingelbach's academic work concerned a more domestic cast. At Stanford's Vision & Learning Lab, he worked under Fei-Fei Li and Jiajun Wu on visual reinforcement learning and simulation for robotics. These are fields in which giving a machine a sensible understanding of its surroundings is a substantial task before anyone asks it to be charming.

In 2021, he was a co-author of BEHAVIOR, a benchmark built around 100 everyday household activities in simulated environments. Cleaning and food preparation were among the tasks. The project included 500 human demonstrations recorded in virtual reality. Its experiments found that contemporary embodied AI systems struggled with the realism, variety and complexity of those activities. Ordinary life was a demanding examination.

He also co-authored iGibson 2.0, a simulation system for robot learning involving household objects and tasks. By October 2022, he was helping present the BEHAVIOR codebase and research demonstrations at an ECCV tutorial. The work involved building tools other researchers could use, as well as producing results. A simulated kitchen, like a real one, becomes much less useful when nothing works.

There is a connection between those environments and the characters that later occupied his attention. Both demand consistency across time: objects, actions and context must fit together. That connection is a way of understanding the trajectory, rather than a claim that a robotics benchmark somehow contained a secret business plan. The business required a separate decision.

The research preview finds an audience

In 2023, Lingelbach left his Stanford PhD work and began building Hedra with Bergman. A year later, he described the experience as repeatedly humbling. Research expertise had not supplied a complete apprenticeship in starting a company. He credited the friendships formed along the way, including the community around a16z's second speedrun cohort.

Hedra's research preview arrived in late June 2024. When Lingelbach announced its $10 million seed round that summer, he said more than 350,000 users had already generated millions of videos. Index Ventures, a16z speedrun and Abstract backed the round. The abstract question of whether people wanted these characters had acquired a fairly noisy answer.

One early collaboration gave the technology a rather specific assignment: bring back Reid Hoffman as he looked in 2004. Working with Hoffman's team, Ben Relles and creator Shawn Jason, Hedra used an early Character-2 model for the project. Here was a recognisable person, a recognisable period and a performance that had to serve an idea. The demonstration had a story attached.

That is an instructive test for a company devoted to characters. A generation can be technically impressive in isolation. Put it beside someone's expectations of a familiar person, and the audience becomes a much stricter critic. A convincing surface is only the beginning of the audition.

The babies get microphones

By May 2025, Hedra was associated with a wonderfully undignified corner of the internet: AI baby podcasts. Digital infants appeared to conduct the sort of conversations usually reserved for adults with recording equipment. It was a reminder that customers do not necessarily adopt a research system in the solemn spirit in which it was developed.

The company reported more than 2.5 million users and over 10 million videos that month. It announced a $32 million Series A led by a16z's infrastructure fund, bringing its stated total funding to $44 million. Those figures describe that moment in its history. They are useful as a measure of how quickly the experiment had travelled, rather than a running counter for today.

Character-3 and Hedra Studio gave that audience a model and a working environment for making character video. The ambition included marketing teams as well as individual creators. A brand mascot, an explanatory spokesperson and an improbable podcast host could all require the same underlying skill: delivering dialogue with a performance that fits.

Lingelbach's stated ambition went further than a succession of amusing clips. He wanted characters with a history and a personality that persisted between scenes. In his account, a character's intention should affect its behaviour. A filmmaker needs to be able to direct someone, including someone who exists entirely inside a model.

That also explains his interest in interactive video. The appealing possibility is a character with whom a viewer can form a connection and interact. It remains an ambition with demanding technical requirements. The interesting question is how much control a creator can exercise while keeping the performance coherent.

A founder returns to the support queue

The daily work of running Hedra sounds less cinematic. Lingelbach says he answers support tickets to remain close to the people using the product. He sees what confuses them, what they need and what works. A customer unable to complete a task can supply a useful review without ever discussing foundation models.

“It connects me to why I'm building Hedra”

Michael Lingelbach, on answering support tickets

His other habit is walking around San Francisco's South Park with members of the team. These conversations happen when colleagues ask for them. He describes them as a chance to get to know one another and discuss what they are building, rather than scheduled discussions of work performance. Even a company making digital characters needs people to understand one another.

Closeness to the work has also caused trouble. For Hedra's first year, Lingelbach ran the daily engineering standup and offered detailed feedback on features before they shipped. Eventually, he recognised that decisions were waiting on him. The founder's attention, initially a resource, had become a queue.

Hiring an engineering leader named Jon allowed him to hand over the standup. He described a similar change in marketing and sales, with Paul and Justin taking responsibility for execution. There is a small irony in the assignment: a founder working on controllable characters had to become more comfortable allowing actual people room to act.

The marketing team supplies another kind of practical test. Lingelbach says it makes its content with Hedra itself. That means colleagues encounter the same creative problems as customers. Their own campaigns can expose a weak workflow or an unconvincing performance. The product has an audience inside the office before it reaches the one outside.

The next scene gets larger

In February 2026, Lingelbach put his name to an explanation of Hedra Omnia. The direction reached beyond making an appealing speaker. He described a unified base model bringing vision, text and audio together, with an emphasis on learning how environments look, move and change over time. Partners could adapt the foundation through further training on their own domains and data.

There is a familiar research question inside that broader ambition. What does a system understand about the world it depicts? A changing viewpoint, a moving object and a sound might all belong to the same event. Learning their relationship could support more specialised applications. Lingelbach's work had returned, by another route, to worlds and what happens within them.

The company's product surface expanded too. In March, Lingelbach discussed Omnia alongside Avatar, Teams and Agent: longer audio-driven video, shared workspaces and a system for coordinating creative tasks. By October, Hedra's developer offering included an API, Python and TypeScript tools, a command-line interface and connections for software agents. The work increasingly involved helping other people build on the machinery.

Through those changes, the personal thread remains easy to recognise. Lingelbach is interested in what a machine can make, and in what a person can do with the result. His research history gives him a vocabulary for the first question. His habit of reading customers' complaints keeps the second inconveniently concrete.

The theatre connection supplies a final, useful complication. A performer can follow instructions perfectly and still lose the room. Attention is earned in the details: a gesture that belongs, a reaction that arrives on time, a sense that the speaker has something at stake. Lingelbach has built a company around getting machines closer to that standard. Somewhere between the simulated kitchen and the baby podcast, the theatre kid found another stage.

Watch him tell it

Follow the thread

Profile updated 22 October 2026. Growth and funding figures are dated to their announcements.