BTW, there is a gradio space here:

https://huggingface.co/spaces/Overworld/waypoint-1-small

And our streamed version:

https://overworld.stream

The context seemed to last a few seconds. I went from a mock up screenshot of a fantasy video game, complete with first person weapon. Then as I moved forward the weapon became part of the scenery and the whole world blurred and blended until it became some sort of sci-fi abstract space. Spinning the camera completely changed look and style.

I ended up with a UI that closely resembled the Cyberpunk 2077 one complete with VO modal popup. I guess it must have featured a lot in the training data.

Really not sure what to make of this, seems to have no constraints on concept despite the prompt (I specifically used the word fantasy), no spatial memory, no collision, or understanding of landscape features in order to maintain a sense of place.

Accurate to my experience hacking on this model today, but I don't think anyone's blowing smoke about it.

Thinking back to where GPT-3 was 5 years ago, I can't help but be a little bit excited. And unlike GPT-3 this is Apache.

If you think this is cool you might also be interested in https://github.com/MineDojo/NitroGen which is kind of the opposite (and complimentary).

10,000 hours training data seems quite low for a world model?

60fps training data goes a long way ;)

An RTX 5090 for 20-30fps for the small model: That is not as unreasonable as I had feared :D

Hi,

Louis here. CEO of overworld. Happy to answer questions :)

great work! Will the medium model be also open/apache-licensed?

Medium is going to bc cc by sa nc 4.0. We may reevaluate in the future and make it more lenient. Small is meant to be the model for builders and hackers.

this is like an open weights version of DeepMind's Genie