which non-transformer neural networks are matching frontier performance using compute scale?