Michael Lingelbach

A theatre kid who built computers

Michael Lingelbach grew up in a small town in Florida^[26]. He did local theatre productions from elementary school and sang a cappella in high school, and he says he built his first computer with his dad at about six^[8]. The stage is still his favourite joke about his career: on Heather Cooper's Intelliverse podcast he said his acting career did not work out, so he had to go and do a startup^[8].

Unfortunately my acting career didn't work out so I had to go do a startup.

Michael Lingelbach, on Intelliverse^[8]

A Hedra hiring profile lists earlier research posts: a postbaccalaureate role at the Salk Institute for Biological Studies, a scholarship at HHMI's Janelia Research Campus and an internship at NVIDIA^[34]. His own site does not list them, so treat them as unconfirmed. His Google Scholar profile does include a 2022 Nature Communications paper on neuronal ischemic injury^[4].

Simulated homes for robots

At Stanford he was a PhD student in the Stanford Vision and Learning Lab, working with Jiajun Wu and Fei-Fei Li on visual reinforcement learning, simulation for robotics and embodied AI^[1][3]. Reports differ on the exact field of the PhD, so this page sticks to the lab. His papers from those years build virtual worlds for robots to learn in: ThreeDWorld in 2020, iGibson 2.0 in 2021, BEHAVIOR at CoRL 2022 and BEHAVIOR-1K, a benchmark of 1,000 everyday household activities, at CoRL 2023^[4][42].

BEHAVIOR-1K is his most cited paper, with 721 of his 2,308 Google Scholar citations^[4].

The Neovim years

Alongside the research he became one of the people behind Neovim's built-in language-server (LSP) client. His first pull request to nvim-lspconfig, in November 2019, added support for Microsoft's Python Language Server; he went on to land 209 merged pull requests there and 131 in Neovim itself, and was listed as a default code reviewer until April 2022^[5][6]. His GitHub account, mjlbach, has 97 public repositories and 320 followers^[33].

Leaving the lab to direct characters

In 2023, in the final year of his PhD, he took leave to start a company^[7]. Work on Hedra began around August that year with Alex Bergman, a Stanford friend whose lab worked closely with his^[7]. VentureBeat later described both founders as Stanford PhDs^[9].

The idea came from wanting control. He says his drawing skills peaked in second-grade art class, and that a director like Wes Anderson would be dissatisfied if he just had to roll the dice on a video model^[7]. The name is geometry rather than mythology: Hedra is the suffix of polyhedra and tetrahedra, which he described as the foundation of a geometric solid^[8].

I wanted to direct the characters.

Michael Lingelbach, on Emergent Behavior^[7]

Character-1 and the talking babies

Hedra launched Character-1 on 18 June 2024 as a free research preview: a model that turns a single image and an audio track into an expressive talking, singing or rapping character^[11]. Two days later he was on the Emergent Behavior podcast explaining why audio drives the whole performance, breathing and gesture included, rather than only the lips^[7]. On 1 August 2024 Hedra came out of stealth with a $10 million seed round from Index Ventures, a16z speedrun and Abstract Ventures^[9][1]. That week he told Intelliverse that Hedra had passed roughly 1.7 million videos created, about four weeks after launch^[8].

Memes did much of the marketing. On the a16z podcast he called memes one of the best go-to-market strategies for a lot of companies, and credited the viral talking-baby podcasts, and creators such as Neural Viz building whole character universes, with Hedra's rise^[25]. His launch playbook was to give leading AI artists early access and let them post, rather than pay influencers, and he has singled out Heather Cooper's and Uncanny Harry's early Character-1 work^[27][8].

Studio, Character-3 and the Series A

Hedra added voice cloning in October 2024^[35]. On 6 March 2025 it launched Hedra Studio with Character-3, billed as the first omnimodal model in production, reasoning jointly over image, text and audio, and served on Together AI^[14][15][16]. The Character-3 API followed on 2 May^[17]. On 15 May 2025 Hedra announced a $32 million Series A led by a16z's infrastructure fund, with partner Matt Bornstein joining the board, bringing total funding to about $44 million by Hedra's count^[18][19].

On the PMF Show that August he said Hedra had reached an eight-figure revenue run rate after the Studio launch, was signing an enterprise contract every couple of days with no outbound sales, and had a team of about 23^[27]. Those are his own figures on a podcast, not audited ones. He has also kept publishing, as a co-author of Phased Consistency Models at NeurIPS 2024 and of the MagicInfinite family of talking-video and video-distillation papers in 2025^[40][41][4].

Real time, then world models

In July 2025 Hedra launched Live Avatars with LiveKit, real-time streaming avatars for voice agents, reported at under 100 milliseconds of latency and $0.05 a minute^[23][24]. Character-3 HD followed in November 2025^[29]. On 5 February 2026 he wrote the launch post for Hedra Omnia, a unified base model across vision, text and audio that partners can post-train on their own domains^[30].

Through 2026 Hedra has opened its models to developers through an API, SDK, CLI and MCP, and its homepage now describes the company as visual inference models and infrastructure^[31]. No press coverage of that shift had appeared by October 2026, so that wording is Hedra's own.

Hollywood, prompts and characters

He argues that AI characters create new kinds of entertainment rather than replace actors. He told VP Land at AI on the Lot in June 2025 that he does not think AI is going to replace humans in Hollywood, and that prompting from language alone is underspecified^[21][22]. On the a16z podcast with Justine Moore and Matt Bornstein he described the character, not the clip, as the unit of control^[25][26].

On AI LIVE

He has been a special guest on AI LIVE twice, both times as a LinkedIn Live before the episode archive on this site begins: Ep 41 on 14 August 2024, two weeks after the seed round, and Ep 67 on 27 March 2025, three weeks after the Character-3 launch^[12][13]. Neither recording is in the archive, so this page does not quote either show.