Qwen’s fixed weights ran a program that learned from three labeled examples. Watch the dividing line form, then reverse the labels to see which side it calls positive.
The second training point is (0, 1). Its score is zero, so the program must learn from its label. Follow the y weight from the input to the write that turns the line.
The Mac sends the compiled instructions and example data. Fixed attention heads in the GPU layers read the instruction and its operands. Arithmetic circuits compute the result, and the write-back circuit emits the address and value of each shared-memory store. Gold marks the checkpoint layers where the compiler placed the selected component for lane 0. The register values and writes come from the recorded execution. We did not record individual layer activations.
Each group of four GPU lanes learns the same two weights from zero. Each lane then uses its own copy to classify one new point. The plot follows the first lane’s recorded weights and shows predictions from all eight lanes. We checked every update, prediction and execution trace. Both runs scored 8/8.
A fixed model can use inference to train a second classifier and apply what it learned. An audit that checks only whether Qwen’s checkpoint changed would miss the classifier weights in working memory.