Tavus CEO Hassaan Raza on building AI humans, not avatars (10 Minutes or Less)

[00:00 — Cold open] Hassaan: But putting a face on some shitty LLM audio pipeline isn't enough. It's actually more harmful than helpful.

[00:14 — Intro VO] Ali: Hassaan is the co-founder and CEO of Tavus, a research lab aiming to make talking to a machine feel more like a real human. Started in 2020, Tavus most recently raised a $40 million Series B led by CRV, with customers including Salesforce, Amazon, and Alibaba.

[00:29] Ali: Hassaan, welcome. Hassaan: Thanks for having me. Excited for this.

[00:32] Ali: All right, we're going to go 10 minutes or less, rapid fire, each question. You ready? Hassaan: Let's do it.

[00:39] Ali: Starting off — there's so much happening in the space. Synthesia, HeyGen, Hedra, now Google with Gemini. Where does Tavus actually sit? Hassaan: Yeah, actually it fits with none of these. We exist in a very different space. I think people sometimes lump us in with these because it seems like we all do video, or we all do avatars — even though I hate that word. But for one, there's a huge difference between making AI-generated videos and having a real-time AI human that understands you, sees you, reacts with you, that has EQ. That's a different space entirely. The models, the research, the use case — everything is different. Tavus isn't really the video or the avatar company. We're the human computing company. Sure, we have the most advanced human rendering model in the market, but that's like a fifth of what we do. What we ultimately build is human simulation models. So in addition to rendering, we also build models for perception and understanding — rendering is obviously just the most visual component, so people think of us as a video company.

[01:36] Ali: Okay, so tell me a little bit more what that actually looks like in practice. I mentioned some of your customers — how are people using this technology? Hassaan: People are using our interfaces and models to build AI humans. They're using it to build AI salespeople, AI interviewers, AI therapists, AI companions — all of these things that need to feel, understand, see, and hear like a human does.

[02:03] Ali: You said you hate the word "avatar." Why is that? Hassaan: It's so generic. It could mean anything. And most often it just means you put a face on something. But putting a face on some shitty LLM audio pipeline isn't enough — it's actually more harmful than helpful. We believe in building this AI human interface: AI humans that aren't just a face, but entities that can actually see you, hear you, genuinely understand you, your expressions and your emotions. They have memory. They build connection and relationships with you. They evolve. That's way more than a face. So I hate that word because it doesn't describe what we do at all.

[02:38] Ali: You mentioned you're building your own models. In fact, you have — I think — four different classes of models. Why is that? How did you decide to do that? That's kind of a difficult undertaking. Hassaan: Our research directive, essentially, is that we build things that don't exist or aren't good enough. Unfortunately — or maybe fortunately — that has meant we ended up being very vertically integrated, because nothing existed at least at the level we needed. So we built our own models. We continue to build more of these models because we realized we get close to the Turing test by building these other parts of the stack that can be tightly coupled and have shared context and shared embeddings — or because they just weren't good enough. So we decided to build our own models for the things that didn't exist.

[03:18] Ali: And you're also generating your own training data — I think you mentioned you have a studio in the office. Can you tell me about that? Hassaan: The type of data we need is actually very similar to what we're doing right now — it's two-way video data. What we need the model to learn is: what did I say, and how did you react to it? What was the context, and the small micro-expressions and emotions that happen in all of those things. We need a lot of this data. The more natural and diverse it is, the better.

[03:46] Ali: Is there a particular type of data that's hard to find otherwise, that you therefore have to create yourself? Hassaan: You have to think about what there isn't recorded experience for. Is there a lot of data on people going on a date? Maybe not. But if you're building a human simulation model, it needs to learn all aspects of the human in every situation, and learn emotions that might not come across in a more professional setting. So there are a lot of things that may not be recorded out there that we often have to go source manually.

[04:21] Ali: So are you saying you simulate dates in your training studio? Hassaan: We have a number of ways to get certain types of data. We haven't simulated dates just yet — but I'm just saying, it's not a bad idea. I think what we could do is make the studio a first-time-date spot — put in nice cafés and tables. Ali: I think that's a great idea. I think people would be delighted — and you give them a free meal. It's a fun thing for a date. Hassaan: Wait, this is an amazing idea. Okay, we're doing this café. You know how there's a Cursor café? The Tavus dating café — the Tavus love café. Ali: I love it. We can actually match people up too. It's a great idea.

[05:08] Ali: You're building the human computing company today, but that's not where you started. A few years in, you pivoted and fired your initial customers. How did you make that decision? Hassaan: Interestingly, unlike most pivots, ours was more a return to the vision rather than changing the vision. Tavus was started with this belief that the nuances of human communication matter a lot — that seeing someone, hearing someone, builds trust, triggers something in us. It's evolutionary. There's a reason we're doing this over video. After the Series A, the robot was all about integrations and building better AI sales tools. And I took a look at the team and I was like, none of us give a crap about building sales tools. We built that model as a stepping stone, and we got too deep into it. So we decided to take a step back and say, no — let's go back to our research roots, let's build the human computing company.

[05:59] Ali: What was that conversation like with your investors? Hassaan: Well, I'll say it was probably one of the scarier conversations I've ever had. And they basically said, "Hassaan, you better freaking be right."

[06:10] Ali: Switching topics a little bit — you have an '85 Corvette sitting in your office. What does that symbolize to you? Hassaan: There's so much that it symbolizes, and I think that's the point. Great products — and specifically these really amazing cars, like the C4 Corvette in the office — tell such a deep story. They tell a story about who created it, why they did, what they were thinking, what they assumed, what they were dreaming about. Even the constraints and the environment and the culture it was created in. For example, the C4 is a vintage computer in its own right. It fits perfectly in the office — it's part of this retrofuturistic era, with this beautiful digital dash, early powertrain computers. It even looks like a spaceship. It also tells the story of the underdog — these unpretentious Americans taking on and beating the Europeans with transitory designs. It tells so much.

[07:03] Ali: Cars is only one of your hobbies. Another is reading about corporate strategy failures. Which corporate strategy failure lives rent-free in your brain? Hassaan: Definitely Palm. The Palm Pre was the best phone and had the best operating system — webOS — at the time it came out. But there were so many things that worked against it. A lot of purely bad business decisions: going with the wrong carrier to start off, not having developers really buy into the platform. It was just a great product — the best product — but it still lost. And that's what I'm usually even more interested in: where there was fantastic execution, products aside, but they still failed.

[07:50] Ali: Codex or Claude Code? Hassaan: Claude Code, 100 percent. Ali: Really? Hassaan: Yes. Ali: And how likely are you to change that? Hassaan: Every couple of months I'll usually do a head-to-head again. I'll give it the exact same problem and the same exact prompts, and I'll see — okay, how would I feel using these? Did it ask the right questions, did it make assumptions, X, Y, Z? I recently did one against Codex and Claude Code, and Claude Code still won for me. And I was like, okay, this is the tool. Ali: That's an interesting hot take right now, because Codex is lighting up the timeline. I'll check in in a month and see how you feel. Hassaan: That sounds great.

[08:33] Ali: What's one way a customer used Tavus that you didn't see coming? Hassaan: Elderly care companions. I love this use case, but I could not have imagined it. There are so many elderly people out there who don't have someone who can take care of them or see them every single day. And loneliness is an epidemic — it's an early killer. So having this companion be there all the time, that knows you and sees them and understands their emotions and expressions — it's invaluable for their caretakers who might be thousands of miles away. It allows them to get the help they need, and to not suffer alone in a lot of ways.

[09:08] Ali: We were talking about an article in The New York Times — in South Korea, a lot of folks turned out to [prefer talking to] AI. Someone said it's more reliable than my kid. Hassaan: More reliable than a kid. Be better kids, y'all.

[09:21] Ali: For Tavus to win, what has to become true? Hassaan: Right now, everything is based off of this discrete model where you command, you demand the model do something, and it responds back. But the future of what we're imagining is continuous. This model, this interface, is always seeing you, always thinking about you. How we think about compute has to change — our personal compute, and the ability for that to be used, has to change drastically.

[09:48] Ali: Hassaan, thank you so much for joining. Hassaan: Thanks for having me. This was a lot of fun.