Voice agents with Realtime Video — Sidney Primas, LemonSlice

My name is Sidney. I'm the CTO and founder of LemonSlice. And LemonSlice is on a mission to break the avatar Turing test. What we mean by this is making an avatar that is indistinguishable from a human on a video call.

And all of this is, of course, making the avatar photorealistic. But there's actually a long tail of technical problems that we care a lot about, that I'll be talking about today, that we're planning to solve. And those problems are things like getting the emotionsright, getting the object interactionsright, getting the micro-expressionsright, and even figuring out an internal state for these avatars so that they feel like a real human being.

So what I'm going to do in this oh, listen. What I'm going to do in this talk is talk about our technical bet to solving the avatar Turing test, what we've built today, and the hard technical problems we've solved today, and our bet on how we think this problem will be solved in the long term.

But before we dive into that, I want to give you guys a sense for what our product is and how it works. So actually, yesterday was a pretty big and exciting day for us. We had a big launch yesterday.

We had a partnership with Microsoft that brought an ex-US president to life, Teddy Roosevelt. And they built his presidential library. And inside of the presidential library, they did a replica of the Oval Office. And you can walk in, and you can have a full interaction with Teddy Roosevelt in real time.

So definitely not Turing test solved yet, but a big milestone towards trying to solve that problem. And the coolest part of all of this was we had a special guest. But hate him or love him, he showed up.

So we had President Trump, who came to the opening of this ceremony and interacted with our avatar. So we actually got some footage, real footage, not fake, AI-generated footage of Trump interacting. So let me show you guys this real quick.

Do you consider the Panama Canal your greatest achievement?

Well,right then, yes.

Good answer.

The canal stands as one of my proudest battles, no question.

But greatness is a strange thing. It isn't always the biggest or boldest job that matters most. The Panama Canal showed what America could achieve if we held steady and acted fast when the world dragged its feet. That said, I measure my greatest work by the lives improved.

Parks set aside, food and drugs made safe, the square deal given to all, not just to a few. Still, when I stood in the mud watching those steam shovels, knowing ships would pass through, changing the world's map forever, I felt I'd left a mark that would last.

Perfect. So you guys get a sense of the experience. The funny thing is, he was scheduled to be there for a quick minute, one interaction. And he actually stayed for 10 minutes. He left the room and then came back again to talk to Teddy Roosevelt even more.

So you could see there, the avatar is there. It's a full-body avatar. It has hands. It does actions. It does movement. And all of this is obviously done in real time. I'll talk a little bit more about the why we think this is an important problem later in the talk.

But a big part of it is actually related to this interaction.

Humans struggle to focus and pay attention. It's often hard for a lot of people to spend a bunch of time reading. And even voice, it's hard to really lock in and understand what's happening. So we think we're biologically wired to best understand things when there's a visual component as well.

So our bet here is, in the long term, we think most interactions between AI and humans will have a visual layer. And we're building that visual layer.

So now I want to talk about our approach to building these avatars. It's a very different approach than what most other avatar companies use. Essentially, what we do is we take these world models, and we focus them on humans.

And the reason we take this bet is, even though it's harder to get the initial model working, it's harder to train the model. It's harder to deploy the model. Once you have a model, you get all of these nice emergent properties that we all know about, where you just very easily can solve things like full body movement, like object interactions, like movements in the scene, all the way down to the micro-expressions as well and emotions.

And all of those kind of don't come for free, but come more for free than when you use the other approaches that other people use. So let me show you guys real quick the product in action so you guys can get a sense.

So

let me expand this a little bit.

How is it going? I am here to listen and help with anything you need. We can chat about something in particular or play a game if you would like.

So the lip sync might be off here a little bit. But that's just because of the AV setup here. But as you can see, we have actions. They can wave. They really can do any actions. We have movement.

We have hands. And the way this works is it just can use any single image to create the avatar. And so, yes, you can have photorealistic. You're also going to have Pixar. You're also going to have cartoons. It just literally takes a single image.

And

let me show you this. Oh, no.

Perfect. And I'll show you kind of a quick interaction here as well. And then we can talk about how we built this. Here, within the same video call and on the same kind of inference setup, we can easily change clothes.

We can change the scene she's in. And

I can talk to it, too. I'm not going to do it here.

It sounds like you might be trying to say a story. Would you like to tell me a story or hear one?

Yeah. So you can see here, too, that there's physics on the earrings. They move. The water moves. So the video model has a good understanding of physics and can use that to make the avatar feel more realistic.

Perfect.

Now, the way this is used, just for kind of everybody's information, is we're mostly the API layer. So we provide an API. People bring their own LLM. People bring their own, usually, voices. And then we're the visual layer on top of it.

So people use this for language learning. People use it for AI sales calls. Basically anything today where you have a voice agent, you now also kind of have a video agent, if it's on your screen or on your computer.

Great. So now let's talk about how we built this today. This is our approach. And I can't talk about all the technical details. But I'm going to try to give you enough information to make it interesting. So what we do is we get started with training our own video DIT model.

The big thing here for us that really matters is the audio. The audio turns out to be very important for getting emotionsright and the facial expressionsright. People care about that a lot. So we really focus on getting great audio data that we can train on, and then actually also getting the audio encoders veryright.

Most audio encoders today are trained on basically audiobooks, which is very monotone, very simple, don't have a lot of emotions. So if you want to have a very expressive model, you can't use those audio encoders. And you actually have to spend a lot of time getting the audio embeddingsright so that the video model is super expressive.

So that's kind of like step one for us. And this is basically just a video model. People can call this a world model. But it's basically a video model that understands the physics of the world. And now the big challenge is, how do you take that video model?

And how do you make it real time and interactive? And so the way that works is the interactivity part is interesting. So usually, video models are bidirectional. So they can look into the past. But they actually also can look into the future and what's about to come.

And they basically generate videos all at the same time by looking at all of the latents that are being generated. For us, you can't do that. You only can look into the past. And so here on this Make Interactive, you can see that we basically train a model with an attention mask so that the model can only look into the past.

So when you do inference, it never can see the future because the future doesn't exist because you haven't given it those inputs yet. So for example, it doesn't see audio in the future. It only sees what has been said in the past.

And then once you have that, we make it real time. And there's a bunch of things that goes in here. But the biggest thing to speed up video generation is, usually, you spend a bunch of steps denoising these video models.

So let's say 30 steps. You spend 30 steps removing the noise to generate the beautiful videos. And what we need to do is go from 30 steps, bring it down to one step. So basically, just in a single step, go from pure noise to a pure video model to make it real time.

And so those are the two big things that make it interactive in real time. Now, the hard part here, what makes this difficult, is this problem here on the left, which is called error accumulation, which is a problem anybody in real time video generation and world model is familiar with.

I'll show you this video. And then we can talk about it briefly.

And I'm a professor and ophthalmology at Stanford School of Medicine.

So you can see, I mean, that's a very bad example of error accumulation. But you can see, as the video continues to generate, more errors are introduced. And the reason for that is because, since you only can look backwards, you're looking actually backwards at videos that you've previously generated.

But each video you generate, or each video block you generate, has some error in it. So now you're looking in the past. You're looking at the error. You're adding more error to it. And then just the error compounds over time.

And this is especially hard because, I mean, ideally, these video models are endless. The Teddy avatar is generating continuously, nonstop, frame by frame, for eight hours straight with no reset throughout the entire process. We have another one that's going to be generating for 16 hours straight.

So it's a very hard problem to not have any error accumulation over long periods of time. And

we came up with a new way to solve this problem that is different to the best of our knowledge than what everybody else does today. And so I can't talk about that yet. But because of that, we can basically generate these very long videos.

I mean, you guys can try it out that essentially have no error accumulation, no noticeable error accumulation in them.

Great. And so the last two hard problems we worked on to enable this is hardware optimization.

There's a bunch of stuff that goes into this. The big thing here is making it low cost enough so you actually can use this. Usually, when you generate an AI video, it's five seconds. And you just share that.

We have to generate minutes and hours of AI video and still have it be cost effective. And so the cool thing here is, we've been able to make the model small enough and efficient enough so that the costs are about the same as a voice model, which is crazy to me.

Because think about a voice model, how much

data is streamed there compared to a video model. It's much more pixel heavy in a video model than with voice. And the costs are about similar. So the cool thing there is, this enables us to do consumer use cases.

And a lot of our customers are consumer companies that basically use the video for consumer and entertainment applications. And then the final thing to mention here is the model harness. I feel like the model harness is something that is often overlooked but is actually super important and super hard.

Getting the model harnessright is a huge technical challenge for us. And I feel like a lot of our value, actually, in productizing this is in the model harness. And the way to think about it is, you just have a bunch of separate threads.

All of it is managing real data streaming through our system. And you have basically a bunch of stuff you do on a GPU and a bunch of stuff you do on a CPU. And you have to orchestrate this perfectly in a way that the video always remains real time.

There's never any stutter that happens inside of the video. And this is especially hard when you have things like interrupts. You have queues. You're buffering data. You have to clean the queues. And so getting this orchestrationright at production, at scale, has been a ton of work.

And honestly, I feel like over time, a lot more of the value of the things we build will be in figuring out the model harness. I think it's especially true for any real time applications.

Cool. So this is our technical approach of commercializing these world models for avatars. Let me talk a little bit about what we're actually working on actively now. So the big thing we want to enable next is having more of an emotional engine that's driving these avatars, having them be more aware of what's actually happening inside of the conversation, and then be able to react to what's happening.

So react means theright emotions at theright time for theright duration, and same for the actions. And today, when you interact with these avatars, you can try. It still feels a little bit I mean, it definitely still feels

awkward. And I think a big part of that is, they don't emotionally react to you. They're not listening to you. They're not emoting in theright way when they're talking. And so this is what we're trying to solve. And so let me show you guys some of kind of the new videos we're able to create with basically better emotional control.

So this is a real time video.

Oh, my god. Go. Go. We won the game. I can't believe it. We won. Oh, my god. Oh, my god. Oh, my god.

Oh, shoot. So you can see, the emotions just matter a lot. They make a huge difference in connecting with this avatar. You feel much more connected with that avatar. So that's one thing. And then this is the next generation model, not in real time yet.

But there's just a lot more.

I think the most important thing is just to take a moment and breathe. Life gets so busy, you know? Do you think you'll stay there for a while? Yeah. I really think I need to stayright here for a long time.

So there'll be a lot more basically natural interaction.

I don't know what to do about the ability.

I don't know. Sorry. There'll be a lot more natural interactions with themselves and then also with their environment. And the big thing here is, our model can already do this today. It has the capabilities to do this. It's just not controllable enough to make it real time with the conversation and not deterministic enough to make it useful with the conversation.

So a big part of our work here is making these actions actually controllable at theright time. And to do that, we're building this emotion engine that is just literally predicting based on the audio input from the avatar and the text input for what the avatar is going to say, what the action they should be doing at this moment in time.

So target launch for this next model is

in basically one to two months is what we're targeting.

Awesome.

So where do you want it tonight?

Allright. Great. So that's where we are today. I just want to talk briefly about let's see what the time is. Almost done.

I want to talk briefly about where we see this heading. So I strongly believe that, in the end, there will be a single model

that is the EQ layer for AI. The way you can think about this model is, the model will literally take in directly the user video and the user audio, so directly the user talking to you like you are in a FaceTime, and put out the avatar video and audio, and do all of that in an end-to-end model, in a single model end-to-end.

Internally, inside of the model, it will do the audio understanding, generate what it's going to say, model its own internal state, like own internal emotional state, and then based on all of that, produce an output

video and audio. And all of this will be trained in one model. This will happen. We're already seeing early papers that are doing proofs of concepts around this. What we're not saying is that this EQ model will be very intelligent.

It will have very high EQ. And it'll be very good at interacting with people. But there will be a separate model that will drive the IQ that will basically give this EQ model input to do all the magical things that AI can do today, like do the tool calling, do the deep thinking, do all the intelligent stuff.

But there will be an end-to-end layer up front. And so that's our bet. We're working towards building that. And we can talk more about some of the benefits of this approach. But I feel strongly that, within two or three years, you'll be seeing these kinds of end-to-end EQ models coming on the market.

That's it.

Maybe take some questions if there's any questions. Yeah, go ahead.

So obviously, EQ could probably define thousands of different configurations of emotion per user. So obviously, the inputs are pretty clear, like video and audio. And you can process that responsibly by the model. How do you think about the output of the EQ model and, I guess, the granularity at which you will be able to control that in the future?

In the future? Well, I think, in the future, in the long term, it's going to be more of an internal state where the model will have an internal state in latent space that defines what the emotional state is or the goals are of the model.

And they will be more implicit. So they won't be human interpretable necessarily.

Today, basically, the way we think about actions and emotions is basically words. If you can describe it in words, you can generate those actions and emotions.

So I guess this state of the EQ that you expect the model to hold would be, I guess, what is the ideal state of the avatar in response to whatever you're decoding from the EQ?

Exactly. Exactly. And track that over time,right? Have an internal state that tracks that over some history of time. But yes, exactly.

You guys are trying to pass the Turing test, the avatar Turing test.

Yes.

I just wondered if you passed it with your first example. Did Trump know that wasn't Roosevelt?

There's a big debate about that. Twitter is debating. Gavin Newsom said that he definitely doesn't know. And he got tricked by the ghost of Teddy Roosevelt.

Is this a good question? So the avatar is looking at humans. What is the way that companies can take this to be a real human model?

You're asking about what the way method is or the test is? Yeah. No, that's a really good point. We haven't done this yet. But we're in the process of figuring out our own version of the Turing test for these avatars, which will just include real people.

And we're planning to run that this year. And we won't pass it this year. But it'll just be cool for us to track it over time to see when we can actually solve the Turing test and ideally allow other people to also go through the test as well.

But yeah, we plan to create a test and then publish around the test.

Oh, yeah.

So are you also planning to

create a digital twin? Extending this to make a digital twin like give more background information and then create an avatar that can give and tag emotions?

Not for us. That's not what we plan to do. Maybe people can use our technology to do that. But that's not our goal. But you could imagine people using our technology, giving it theright context, having theright harness around it, and then having their nice IQ brain around it to do exactly what you're saying.

But we don't want to solve that problem.

How do you imagine the cost per hour of producing this scaling over the next two years? Is it a big challenge producing it in the real world, or is it very expensive?

Yeah.

We would still like it to be cheaper. The good news is it is OK, almost done. Last question. The good news is it is surprisingly going into this, we didn't know how cheap this would be. I've been very surprised at how inexpensive it is.

Again, the cost of this is at the same level as an audio model in terms of what we charge for it. And we're all hoping for this cost to go down. Part of it is for consumer. You just need very low costs.

And then part of it is, as the cost goes down, we can do higher resolution, which will help us a lot. But I guess the way we're thinking about it is, there's algorithm improvements and hardware improvements. And what has been true for some time now is, things have been improving on all vectors faster than expected.

And we're just hoping that continues, basically. I think there'll also be very cool architectural updates to move to more of a token approach instead of a diffusion approach that will make this type of video generation way cheaper. That's it for me.

Thank you, guys, for showing up. We are hiring. If you guys are interested in solving these types of technical problems, come talk to me afterwards. Or if you're just interested in this at all, come talk to me afterwards.

Appreciate it, guys.