Our mission: break the avatar Turing Test

Alan Turing proposed the Turing Test in 1950 as a benchmark for intelligence: a chat conversation where you couldn't tell if a human or AI was responding to you on the other side.

ChatGPT broke this original, text-based Turing Test.

ElevenLabs is close to breaking the voice Turing Test: AI that is indistinguishable from a human on a phone call.

We will break the avatar Turing Test: digital humans that are indistinguishable from a real human on a video call.

But, beyond just photorealistic humans, our ambitions are to bring all characters to life — from a photorealistic human to a cartoon turtle. We are building digital characters that you connect with.

What is a Character World Model? A video world model is a new type of foundational AI model that generates video that is interactive, persistent, real-time, and physically accurate. For example, a playable videogame where every frame is generated in real-time by a diffusion model in response to the player's button clicks rather than pre-programmed within a 3D game engine.

LemonSlice trains video world models that can simulate the world, but are heavily optimized for interactive conversation. This is called a Character World Model: expressive, digital characters that talk, listen, and react in real-time.

LemonSlice released the world's first Character World Model in April 2025 with LemonSlice-1. Our latest model — LemonSlice-2.1 — is a 20B end-to-end diffusion transformer video model that renders every pixel at 20fps on a single GPU.

In contrast to deep-fake or generative approaches to avatars, Character World Models enable any type of character, hand-object interactions, movement within physical spaces, moving backgrounds, and more. It is the Bitter Lesson approach and the only way to ultimately break the avatar Turing test.