(author) The model is learning directly from a dataset of (input bits, output bits). It is directly predicting output bits (index of which column to play next) from input bits (the state of the board right now). It is doing so using logic gates directly as compute primitives.
Thanks for responding! I understand if you don't want to give more detail, but can you say something about how you go from that learned model to an ASIC? i.e. what is organizing the logic gates to implement the model?