Resolves YES if we know that OpenAI has created a game-playing AI that uses Sora by the end of February 2027, NO otherwise.
The game(s) that the AI plays could be anything - Minecraft, XLand, board games, etc. The only requirement is that there is some clearly-defined goal that the AI is shown to pursue across multiple time steps. To be clear, it doesn't matter how many games the AI can play, as long as it's shown to play at least one.
As long as OpenAI uses a model that started out as the first version of Sora (the one they announced in February 2024), it counts, even with fine-tuning or other tweaks. Sora doesn't have to be part of the final product - for instance, it could help to train a separate model, similar to a reward model in RL. If OpenAI uses another model trained from scratch, it doesn't count, even if they call it Sora or use similar training techniques.
See my LessWrong post that explains a very basic way this could be implemented.
I may trade in this market.
1,000
1.00One thing I found when I developed my stock trading model is that it's astonishing how solving one problem generalizes so well to other problems.
While the model isn't very good at other types of data without modifications, it can predict other things. I expect that if someone was going to pay me more to do this instead of spending my time optimizing its stock trading, I could make it output acceptable music and predict other time series data very well, like traffic flow, with about 5% or less of the work required to develop the original model.
If we were able to see the source code for OpenAI's models, I would bet some money, at least at low odds, that there is some chance that Sora and GPT-4 share a somewhat similar architecture, perhaps using the same code blocks for portions of the models.
As such, I think that it would be easy to train Sora to play games such as chess, just in an inefficient way compared to a dedicated chess model. But, like me, they need to focus on making money, and computational power would be needed to retrain the model to do that, so I am betting NO.
Unless it’s paired with something that looks like a traditional game engine modelling the world around it, I don’t think this can happen. Even if you believe Sora has somehow evolved to have an understanding of the physics and lighting in frame in the video, I think it’s a stretch to believe that there’s any understanding or persistence of a world outside this frame. That might work for incredibly basic games like side-scrollers, but anything in a 3D environment where you can look around would be a major challenge.
@singer That's a possibility. See my post https://www.lesswrong.com/posts/bSwdbhMP9oAWzeqsG/openai-s-sora-is-an-agent for a very basic implementation idea.