A new explanation of the Berlekamp-Massey algorithm is given using a method that is not based on the usual description as shift register synthesis but in terms of matrices. It has the advantage of being more didactic and it reveals some properties that cannot be deduced so easily from Massey's interpretation.