“We also adapt the algorithm discovery procedure to finite fields, and improve over Strassen’s two-level algorithm for multiplying 4 × 4 matrices for the first time, to our knowledge, since its inception in 1969.”
DeepMind turned matrix multiplication into a game and let AlphaTensor play it. It found a way to multiply 4 × 4 matrices with 47 multiplications instead of Strassen’s 49. The catch is that the 47 only works in modular arithmetic, not on the real numbers your GPU uses. It’s a real result. The headline sells it as a faster computer.