Research in progress
Many small yes-or-no questions.
Asking a big AI model what to do at every moment of a game is slow and costly. We are researching another way: reshape the job into tiny yes/no questions, each answered by a very small model on your own PC.
This page illustrates the idea. It does not describe a finished product, and it may change.
Try it: "gather some wood"
Flip the answers. She reads the list from the top, and the first line that applies names her next step.
Her next step
Walk to the nearest tree.
The big model: asleepwoken, once
Who does what
The big model sets the goal
It turns "gather some wood" into a goal and its smaller parts. After that it is called only when something unexpected happens, such as getting stuck. It never picks a move.
Small models answer yes or no
Each looks at the screen and answers one narrow question. They run on your own PC, with no big model among them.
A plain list picks the step
The answers are read against a priority list. The first line that applies is what she does next, so what she said she would do is what runs.
Why it matters to you: quicker reactions, and less strain on your PC. Back to the stories.