press0

Next will be writing the inference stack in C for simultaneous high-speed RL across a large block of GB300s. (We do use a little C++ tbh, but not much)