For 53 years, a cornerstone of modern computing remained stuck. In 1969, mathematician Volker Strassen discovered a clever shortcut for matrix multiplication that reduced a key operation from 8 to 7 steps. Nobody could do better. Until DeepMind built an AI that could.
The 53-Year Stalemate
Matrix multiplication is the single most important mathematical operation in computing. It powers AI training, scientific simulations, graphics rendering, cryptography, signal processing, and virtually every computational task on Earth. It runs trillions of times per second across every computer, phone, and data center.
Strassen's 1969 algorithm was a breakthrough — it showed that the naive approach (8 multiplications for 2x2 matrices) could be beaten. For decades, mathematicians tried to find even better algorithms for larger matrices. For the standard 4x4 case, the best known required 49 multiplications. Nobody could go lower.
AlphaTensor: 49 Becomes 47
DeepMind's AlphaTensor used reinforcement learning — the same approach that powered AlphaGo — but applied it to pure mathematics. The AI explored a search space of over 10^33 possibilities (more than atoms in the observable universe) and discovered algorithms that humans had missed for half a century.
For 4x4 matrix multiplication: 49 steps became 47.
Two steps may sound small. It is not. Here is why:
- A 4% efficiency gain at this fundamental level compounds into massive real-world speedups
- Every matrix multiplication saved is energy saved — at data center scale, this means millions of dollars and significant carbon reduction
- The improvement applies to every computation that uses matrix multiplication — which is essentially everything
Why This Matters Even More Today
Your code is already faster — and you did not even know it.
AlphaTensor's discoveries have been integrated into the standard mathematical libraries that power modern computing:
- NumPy — the backbone of scientific Python
- PyTorch — the dominant AI training framework
- TensorFlow — Google's machine learning platform
If you write code that touches any of these libraries (and most developers do), your computations are already benefiting from AlphaTensor's discoveries. No code changes needed.
Hardware-Specific Optimization
AlphaTensor does not just find one universal algorithm. It discovers custom-optimized algorithms for specific hardware — GPUs, TPUs, and different chip architectures each get their own tailored solution that maximizes performance on that specific silicon.
This is something human mathematicians rarely attempt because the search space for hardware-specific optimizations is astronomically large. For AI, it is just another exploration problem.
The Deeper Significance
This was not just a speed record. It was proof of concept that AI can advance fundamental mathematics — not just apply existing math faster, but discover genuinely new mathematical knowledge in abstract spaces where human intuition had stalled for decades.
The same approach is now being applied to other open problems in mathematics, physics, and computer science. AlphaTensor was the proof that AI can be a co-discoverer, not just a tool.
As researchers described it: "A monumental achievement — proof that AI can solve what humans had long abandoned."
Sources: Nature (original paper), Google DeepMind Blog