Aug 14, 2025 · 46m · a16z
Why AI Characters & Virtual Influencers Are the Next Frontier in Video ft Hedra’s Michael Lingelbach
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the a16z podcast, Hedra Founder & CEO Michael Lingelbach joins Justine Moore and Matt Bornstein to discuss how generative video models are transforming synthetic character creation, virtual influencers, and enterprise marketing. They examine the technical shift toward unified dialogue-video models, strategies for scaling digital presence, and the hands-on engineering culture required to build leading AI product experiences.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 2.7 Guest teaching 2.7 Guest disagreement 0.5 The host pushing back 1.1
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 22:32 Rejecting video models as the actor bottleneck
Michael directly rejects the premise that video models are holding back interactive AI actors, arguing forcefully that LLM personality limitations are the real roadblock.
Hardest push from the host ▶ 31:53 Challenging individual personalization scopeMatt pushes back on Google-style individual hyper-personalization, arguing that single individuals are the wrong primitive cohort size compared to shared social experiences.
Biggest teaching moment ▶ 13:44 Explaining physical correlation in video generationMichael corrects traditional video design assumptions, explaining to the hosts that modular lip-syncing looks unnatural because vocalization is physically tied to breathing and body posture.
The host holds their own ▶ 7:36 Deconstructing video as a high-dimensional search spaceMatt demonstrates clear technical expertise by breaking down the math of video search space across spatial pixel dimensions and temporal frame sequences.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Viral Memes as a Go-To-Market Strategy | 2 | 3 | 1 | 1 | Matt and Justine inquire about virality and GTM strategies around meme creation. Michael explains how memes reflect fast creativity enabled by Generative Video models that fill key architectural missing pieces. | |
| Scaling Personal Presence for Content Creators | 3 | 2 | 0 | 0 | Justine highlights creator trends like history videos and digital avatars on Instagram. Michael builds on her points by discussing automated digital presence workflows using tools like n8n. | |
| Conversational Book Characters and Multimodal Learning | 4 | 2 | 1 | 1 | Matt references sci-fi canon like Neal Stephenson's Diamond Age and breaks down the search space of video across pixel dimensions and frame sequences. Michael agrees while contextualizing specialized foundation models. | |
| Character-Centric Primitives and Theater Inspiration | 2 | 3 | 0 | 1 | Michael shares his theater background and explains why Hedra focuses on character as the fundamental primitive of control. Justine details her own real-world voice cloning workflow on the platform. | |
| Unifying Speech, Motion, and Natural Body Language | 3 | 4 | 1 | 2 | Matt probes Michael on why Hedra unifies model components rather than relying strictly on modular workflows. Michael educates the hosts on why overlaying lip-sync fails because speech is intrinsically linked with breathing and body motion. | |
| From Stanford CS PhD to User-Centric Founder | 3 | 3 | 1 | 2 | Matt notes how rare it is for a Stanford CS PhD student to focus heavily on user experience over pure raw model technology. Michael explains why monolithic omnimodal models require proper steerability and unbundled inputs. | |
| Navigating the Time Dimension in AI Video | 3 | 3 | 0 | 1 | Matt and Justine highlight the technical and financial strain of multi-minute video generation. Michael details why human cognition struggles with high-dimensional temporal prompting compared to standard 2D image gen. | |
| The Frontier of Interactive Real-Time AI Actors | 3 | 4 | 1 | 2 | Matt asks if AI actors are around the corner and notes the shifting perception of audio vs video quality. Michael reframes the issue, explaining that LLM personality limitations—not video rendering—are the primary roadblock. | |
| The Evolution of AI Realism & Real-Time Conversations | 2 | 2 | 0 | 0 | The group jokes about early CGI standards and experimenting with weird inputs like talking trash cans and Waymos. Michael points out that video faces foster greater empathy in real-time interactions like customer support. | |
| Holistic Character Modeling vs. Traditional Lip-Sync | 3 | 3 | 0 | 1 | Matt coins the term 'AI-native' applications while discussing enterprise adoption curves. Michael elaborates on why building a single end-to-end model prevents user frustration caused by modular feature walls. | |
| Niche Communities and Personalization at Scale | 4 | 3 | 1 | 2 | Matt analyzes historical media homogenization versus targeted cohorts. Michael slightly reframes Matt's point from hyper-individualized 'personal Netflix' content to cohort-based 'personalization at scale.' | |
| Digital Avatars, Camera Shyness, and Personal Brands | 2 | 2 | 0 | 0 | Justine and Michael discuss digital avatars for camera-shy creators and funny personal brand iterations. Michael gives a candid look into the physical strain and Type 2 fun of founder life. | |
| Effective Vibe Coding and Clear Team Communication | 3 | 3 | 1 | 2 | Matt asks how the engineering team responds to prototype code dumps from a founder. Michael explains that vibe-coding functional prototypes sets clear visual expectations and prevents post hoc team misalignment. | |
| Evolving Beyond Text-to-Video into AI Content Engines | 2 | 3 | 1 | 1 | Justine commends Hedra's quick growth and user-centric focus. Michael previews Hedra's direction, explaining that simple text-to-video prompts will be replaced by AI narrative content engines. | |
| Sci-Fi Influences and Character-Driven Storytelling | 2 | 1 | 0 | 0 | Matt asks Michael for sci-fi recommendations. Michael shares his appreciation for character-driven sci-fi stories like Hyperion and Star Trek Deep Space Nine. |
Statements from this episode (43)
Lingelbach: Enterprise software companies are running Hedra-generated ads on Forbes
“I'm on Forbes the other day, and I see a Hedra generated, like, ad”
Lingelbach: Memes are one of the best GTM strategies for companies
“I think memes are one of the best go-to-market strategies for a lot of companies.”
Hedra was first to deploy dialogue-centric generative video at scale
“We were, I think the first major audio video foundational model company that was able to deploy, you know, these full body expressive dialogue centric generative videos at scale”
Actor Jon Lajoie created synthetic video series using Hedra
“John LaHua has created who's like a very famous actor behind you know, Taka from the League, created this whole series starting with this, like, you know, Moses podcast to the baby podcast, which he actually started.”
Lingelbach: Generative acting requires personality and control beyond standard video models
“So I think, you know, generative acting is really important, because it's something that's quite distinct from just creating a media model. There's a lot that goes into imbuing, you know, personality and consistency and control into these models that's specifi…”
Moore: Generative video enables indie creators to build studio-budget characters
“I'm starting to see a mix of totally new characters that, like, aliens, like in the Neural Viz series that people never would have been able to create and, like, film before unless you're, like, a Hollywood studio with a big budget.”
Lingelbach: Creators use Hedra with n8n to automate video publishing
“We see a lot of use cases where people combine us with something like N-A-N where they'll use, like, deep research and pull together scripts, they'll bring in a dihedra to generate the audio and video portion, and they'll automate the deployment to their chann…”
Lingelbach: Workflows for autonomous digital personas will proliferate
“I think we're going to start seeing more workflows around the creation of, like, these autonomous digital personas, not just for, like, real people, but also for, like, total fabricated personas as well.”
Lingelbach: Hedra is first to market with low-latency interactive video model
“We just launched this real time interactive video model which is the first to bring this like low latency image, audio video experience really to market.”
Lingelbach: Low-cost interactive video will fundamentally change how people interact with LLMs
“When you're able to bring the cost of this technology down low enough, I think it actually is fundamentally going to change how people interact with these LLMs that are driving a lot of these new experiences.”
Bornstein: Low-latency interactive AI books are near technical reality
“I think we actually can do it now. Like, I think we're actually quite close to having, like, a low latency, fully interactive book, and things like Hidro are kind of the key piece.”
Lingelbach: Audio-only AI agents fail to capture visual learning benefits
“Historically, voice agents, very powerful, but you're missing out on all of the, like, visual comprehension that really benefits people while learning.”
Lingelbach: Claude dominates coding, OpenAI general use, Gemini enterprise
“Claude has become kind of the de facto coding model, opening eyes, that general assistant. Gemini is, like, powering a lot of enterprise use cases now, just due to its, like, cost effectiveness, speed, and general capabilities.”
Lingelbach: AI video will feature bidirectional communication and programmable personalities
“So in my, like, you know, future state, like, you should be able to have a subject in video that you can bidirectionally communicate with, that can follow cues, that has a distinct personality that you can program, and that creates, like, essentially the, you …”
Lingelbach: Interactive AI video requires vertical integration across models and UX
“And I think that necessitates vertical integration, not just in the core model layer, but also in like how you interact with the models.”
Lingelbach: Boundaries Between AI Intelligence and Rendering Models Will Disappear
“I think like with, when we're looking forward as to how the technology evolves, I think a lot of the lines between like what traditionally was like the intelligence portion of a model and what was traditionally the rendering portion of the model are going to g…”
Justine Moore uses a Hedra avatar and voice clone for video intros
“A bunch of intros to, like, my YouTube videos and tutorials now are actually my Hedra avatar using the voice that I've cloned on Hedra.”
Lingelbach: Post-hoc lip-syncing fails because speech rhythm correlates with body gestures
“Putting lips on top of something is not actually driving the animation. Like, your breathing is correlated with how you talk, right? How you move your hands. I mean, I guess I do weird hand gestures. Now I'm getting self-conscious. I would argue it's correlate…”
Lingelbach: Niche AI apps can reach hundreds of millions in revenue
“There applications that maybe aren't wildly popular, but they're so popular amongst a passionate group of segments, these companies can scale to like hundreds of millions of revenue.”
Lingelbach: Iconic AI companies win by solving the UX last mile
“When I think about, like, the most iconic companies in the space, They did that last mile, right? Like, ChatGPT has built out a really good user experience for interacting with LLMs, and they've built things into that core workflow that people aren't coming fo…”
Lingelbach: AI startups should integrate partner models to maximize user experience
“And I think that, for me, it was, like, you need to identify, like, that It's a core pain point and core market segment that you need to serve and you need to do whatever possible to deliver the best experience to them. And if that involves bringing in partner…”
Lingelbach: Unbundling AI inputs allows precise user control across modalities
“The more stuff that you bundle into a monolithic core model, the higher the bar becomes on, like, building that steerability into all the modalities. And I think unbundling the inputs is a good way of still giving people that ability to go in and tweak specifi…”
Lingelbach: Generative AI is moving toward monolithic omnimodal models
“Where I see the field going towards is monolithic omnimodal input, omnimodal output. Where you can provide guidance signals to how closely the model adheres to each of the input modalities.”
Lingelbach: AI video tools face a trade-off between control and model priors
“The more control we surrender over to these very large data-driven models, the more priors we can bake in to make that easier to get to that natural distribution of content. But at the same time, the more control the user has to kind of surrender. And I think …”
Bornstein: Video product design is qualitatively different than 2D media
“Three dimensions is way, way bigger than two dimensions. Right. And actually like a qualitative difference, like the product you need is probably different for three dimensions and for two.”
Lingelbach: Teaching 3D interfaces is harder than teaching reference prompting
“Because these three D systems, it's hard to teach people how to use a three D interface, probably even harder than teaching them how to prompt with references.”
Moore: Monolithic AI video generation wastes compute and time on minor edits
“If you say, like, I want a two-minute video talking about this into a model that does it all at once, like the script generation, the video, the audio, like, that's like two whole minutes of points of tiny things you want to change, and then you've wasted, lik…”
Lingelbach: LLMs, not video models, limit real-time AI actors
“Oh, a hundred percent. But I think that the bigger limitation is not the video model right now. I think the bigger limitation is actually that I think large language models still have a lot of work in terms of making people feel very authentic.”
Lingelbach: LLMs Are Scaling Maturely, But AI Video Remains Very Early
“I think LLMs are now kind of in this, like, scaling paradigm, right, where, like, people are putting more compute in, they're doing more, like, RLHF, they're doing more reasoning. But, like, you know, video is actually still really early.”
Lingelbach: Real-time video faces deter customers from yelling at AI support agents
“We see this now when we're talking to customer support companies that are thinking about integrating real-time video. It's a lot harder to, like, yell at An assistant that has a face that's staring at you that can, like, wince.”
Moore: Most revolutionary consumer AI stems from previously impossible capabilities
“I think some of the, like, most life-changing stuff comes out of things that you could never do with the technology we have before, that people are just randomly, like, Throwing stuff into a model and seeing what happens”
Lingelbach: Startups must translate prosumer AI trends into enterprise products
“It's our job as startups to kind of take these like signals from the prosumer users and then productionize that and make that something that, you know, enterprise users can come and, you know, rely on to make the next generation of their content needs.”
Bornstein: Enterprises Resist Replacing Core Databases with AI Alternatives
“It's relatively hard to get a big company to swap out their database for some AI powered database, you know, or like, you know, core workflows, billing workflows, whatever, but like these brand new experiences, that's kind of the exciting part of the market an…”
Lingelbach: Enterprise accounts are Hedra's fastest-growing customer segment
“Enterprise is more fastest It is our fastest growing segment now.”
Bornstein: Individual personalization is the wrong primitive for media consumption
“Like, I really do think, like, it turns out in an individual person, Is the wrong, like, primitive, or the wrong, like, cohort size for personalization, right? It's just like, it's, you know, you see this across all sorts of media consumption now.”
Moore: Friend's animated AI penguin account has millions of followers
“I have a friend who runs a penguin, like, an animated penguin AI account of herself as, like, a fluffy penguin with, like, her personality that, it sounds crazy, but she has, like, millions of followers.”
Lingelbach: Founder life requires giving up work-life balance
“It's definitely one of those things where you have to give up the notion of having, like, work-life separation, or, like, a personal life or anything.”
Lingelbach: Real-Time UI Prototyping Is Changing How Creative Software Is Built
“And I think this is like changing how we build software, especially. I think creative tools are really exciting because there's just As you're talking about so much dimensionality and so many opportunities to rethink how we interact with media and being able t…”
Bornstein: Only 12 to 24 Silicon Valley Teams Truly Master AI Product Velocity
“I think there's a relatively small set of companies in Silicon Valley right now that Get AI, building AI-native products, and just, like, work their butts off and ship products at high velocity. That's kind of the magic recipe right now, I think, and there's, …”
Lingelbach: Traditional text-to-video prompting will not remain the primary video AI paradigm
“I think that traditional prompt text to video is like very much not what the paradigm will look like.”
Lingelbach: AI video creation will shift to autonomous content engines
“And I think in the same way, like we've seen this transition from like, you know, auto completion and code generation to now coding agents. I think we're going to see something very similar with content engines, and that's a lot of what we've been working on.”
Moore: Very few AI companies care about what users actually want
“Like, you truly really care not just about training the best model for this use case, but also what your users actually want and what they're doing in a way that I think very few AI companies today do.”
Lingelbach uses Claude with GitHub and Linear as an AI product manager
“One of these really exciting use cases of AI is like automating a lot of, you know, these processes that you have to go through as a founder. Like, I recently started using, like, Claude with, like, linear, and Notion, and GitHub integration as, like, a micro …”