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.