AlphaTensor

AlphaTensor is a reinforcement learning system from Google DeepMind that discovers novel, efficient and provably correct algorithms for fundamental computations such as matrix multiplication. It reframes algorithm discovery as a single-player game: the board is a three-dimensional tensor, and moves correspond to building a low-rank tensor decomposition. The system builds directly on the AlphaZero approach that mastered Go, chess and shogi, adapted to a far larger search space.

In its headline result, AlphaTensor found an algorithm for multiplying 4x4 matrices over a finite field using 47 multiplications, beating the recursive application of Strassen's algorithm (49 operations) for the first time since Strassen's 1969 discovery. It also re-discovered the best-known algorithms for many matrix sizes and produced hardware-specialized routines that DeepMind reported ran faster on tested GPUs and TPUs. The paper was published in Nature on October 6, 2022, and DeepMind open-sourced the discovered algorithms and code via a public GitHub repository under Apache 2.0. Note that this is an algorithm-discovery agent rather than a downloadable foundation model: openWeights is false because there is no released model checkpoint, even though the findings and code are public.