[naed90]

Hey, developer of Oasis here! You are very correct. Here are a few points: 1. We trained the model on a context window of even 30 sec. What's the problem? It barely pays any attention to frames beyond the past few ones. This certainly makes sense as it's a question of the loss function of the model during training. We are running now many different training runs to experiment with a better loss func (and datamix) to solve this issue. You'll see newer versions soon! 2. In the long term, we believe the "ultimate" solution is 2 models: 1 model that maintains game state + 1 model that turns that state into pixel. Think of it as having the first model be something resembling more of an LLM that gets the current state + user action and produces the new state, and then the second model being a diffusion model that takes from this state and maps to pixels. This would win the best of both worlds.

[throwaway314155]

This stuff is all fascinating to me from a computer vision perspective. I'm curious - if you have a second model tasked with learning just the game state - does that mean you would be using info from the game itself (say, via a mod or with the developer console) as training data? Or is the idea that the model somehow learns the state (and only the state) on its own as it does here?

[naed90]

That's a great question -- lots of experiments will be going into the future versions o Oasis. There are quite a few different possibilities here and we'll have to experiment with them a lot.

The nice thing is that we can run tons of experiments at once. For Oasis v1, we ran over 1000 experiments (end-to-end training a 500M model) on the model arch, datamix, etc., before we created the final checkpoint that's deployed on the site. At Decart (we just came out of stealth yesterday: https://www.theinformation.com/articles/why-sequoias-shaun-m...) we have 2 teams: Decart Infrastructure and Decart Experiences. The first team provides insanely fast infra for training/inferencing (writes from scratch everything from CUDA to redoing the python garbage collector) -- we are able to get a 500M model to converge during training in ~20h instead of 1-2 weeks. Then, Decart Experiences uses this infra to create these new types of end-to-end "Generated Experiences"