@dboing #17 ok, i did not understand this part, please elaborate:
"PGN can be split into sequences of FENs, which can be expanded into basic input vector into NNue. or a chessboard with pieces on it.
some people prefer to define a position by hiding the maximum parallel information from the 2D board, assume initial position known, and only give clues through the move sequence, other might prefer the intermediate FEN, and me i prefer the board picture (or doing statistics on a database of basic input vectors, if the right ones could be available %). Compression bias from CS perhaps. Or bigger short term memory. or coding decoding talented individuals.
I must be spoiled."