We started with a simple curiosity: could a model beat human players at iceboat racing? That is still the plan. First, a model that beats the average human player. Our models are nowhere near that yet, and we expect them to crash in plenty of strange ways along the way.
Where we hope it goes
If that works, we would like to train better specialist models, and models that can reason.
The ability to internally consider several possible actions, predict their consequences, and choose between them.
We do not know yet how to build that well, or whether it will help as much as we hope.
The rough path
Alongside the models, we are starting to look at evaluation, at how servers could detect automated play, and at how bots could be useful as NPC racers, coaches and pacers. It is early for any of that, and we want to think about it from the start. See Readiness for more.
