How to Build a Consumer AI App That 10 Million People Can't Put Down

We're processing more than a trillion tokens per day. A lot of people think for consumer AI, PMF, product market fit, is very important. But I have a different opinion. You're no longer a tag. AI knows you. We can rank 100 models in one hour, two hours. What have you learned about yourself?

Hi, I'm Joshua Liss. I lead media entertainment here at Nebius, and I am thrilled to be your host for the Nebius for Startups podcast series. Today, I am joined by Lifan Wang. Lifan is the co-founder at Keon Labs, formerly known as FlowGPT. Keon Labs is building the infra layer for data and creativity. Why don't you tell us what that means?

Yeah, sure. Previously, we're known as Flow GPT. We built an app called Emoji, which has more than 10 million users. But what we're actually building, we're actually working on every day, is the infrastructure that actually scales the product. We're actually testing more than 100 different models every week. using the real-time signal from the users, and we're processing more than a trillion token per day.

That's insane, a trillion tokens a day. So, 10 million users, 100 different models that you're doing post-training on. Per

week.

Per week. Yeah. And a trillion tokens a day. Yeah. That's amazing scale. So, what does it mean to be the infra layer at that scale for data and creativity use cases?

Yeah, so I think at the beginning, when we were building this product, we feel like we just need to provide a really good experience for users. That's our original belief. And then we figured out, why don't we try different models? And that actually turns out very important. The retention, I think trying different models, varies more than 10%. We figured out that we need to build better models. The conflict between consumer app and AI model is that AI model is improving, keep improving, but the cost is just so high and we cannot afford the pretargeting model for the consumer use cases.

And especially with your business, because you were telling me how much time your users are spending with you. Tell me what that number is and what that means for how you scale a business like yours.

So our users are spending more than 150 minutes per day per user. How many minutes? 115 minutes?

More than 150.

Per

day, per

user.

That's incredible stickiness.

That's one of the top. time-consumed

apps on the app stores. Congratulations.

Thank

you. Okay. So scaling a business like that is quite challenging. Maybe you can give me a bit of an overview of the different types of users of your platform. What are they coming and doing with Kion Labs on a typical day?

I see. So, the use case is more on the entertainment side. Okay. Which we call role-play.

Okay. Entertainment role-play. Okay.

Yeah. So, for traditional content consumption, it's very restricted, I would say. Okay. For example, when people are reading a novel, people have to wait for the author to update. People cannot control the novel, where it's going. And also, it's not super immersive. You are actually just seeing others doing things. but um for us for for the years on emoji people are actually role playing in the story people can you know like say what they want to do yeah in the in the story

yeah

and they