DeepMind’s paper also pointed out that AlphaTensor discovers a richer space of matrix multiplication algorithms than previously thought — up to thousands for each size.