Fast world models for generative simulations
We are building generative simulations of the world powered by fast world models.
Instead of using a physics or game engine with hard-coded rules to render each frame and decide how the environment reacts, this is done end-to-end through a neural network.
Unlike text-to-video models, our world models generate frames conditioned on actions, enabling real-time interaction with any environment. Users can explore, manipulate objects, and shape their experience through natural gameplay.
We trained a model to emulate Minecraft that runs smoothly on consumer hardware (1B param model runs at 25fps on an nvidia 4090 gaming gpu while other Minecraft world models run at <3fps). Our advantage in speed comes from aggressive latent compression (128x) that allows us to encode every frame into only 15 tokens while others traditionally use 256.
We are now training on real-world footage and are building a general-purpose universe simulator.
Active Founders
Alberto Hojel, Founder
Studied computer science at UC Berkeley and did research at BAIR under Trevor Darrell. Worked on weather forecasting with ML at Rainmaker. Originally from Mexico City.
Was on track to pursue a PhD but realized I wanted to build something greater while presenting my paper at ECCV.
Passionate about lucid dreaming and immersive content.
Rami Seid, Co-Founder
A high school dropout who got his first internship as an MLE at a robotics lab, I went on to become the CTO of a govtech contracting company. After that, I cofounded my own telecommunications company while doing ML work on the side. Now I'm building a universe simulator at Lucid.
