A 1.2M-parameter transformer, trained from scratch, solves the cube subgoal by
subgoal — placing pieces one at a time, then finishing the last layer — emitting
one short macro of moves per subgoal. Every intermediate cube state along the way
is machine-verified.
connecting to solver…
hidden-state manifold
the model's internal state as it decides each move — colors cluster by the move it chose
drag a sticker turn that layer · drag empty space / right-drag orbit · scroll zoom solve animates the model's subgoal chain · points on the right are its hidden state per move