The 47-Step Solution

The 50-Year Problem

Strassen's Ghost

Matrix multiplication is the simplest computational problem you can state and the hardest to solve efficiently. Here is the straightforward approach: to multiply two n-by-n matrices, you perform n³ scalar multiplications. For a 4x4 matrix, that is 64 multiplications. It is the naive method every introductory computer science course teaches, and for decades, mathematicians believed it was optimal. Why would you expect anything faster?

In 1969, a German mathematician named Volker Strassen shattered that assumption. He published a paper titled "Gaussian Elimination is Not Optimal," proving that you could multiply two 2x2 matrices using only 7 multiplications instead of 8. This does not sound like a breakthrough. It sounds like rearranging deck chairs. But Strassen's insight was recursive: because you can decompose a large matrix into smaller blocks, you can apply his algorithm recursively. When you cascade the recursion, the efficiency multiplies.

For 4x4 matrices, Strassen's two-level algorithm yields 7 multiplications at the first level, then 7 multiplications at each of the seven sub-problems — giving 7 × 7 = 49 multiplications instead of 64. A 23% improvement. In the decades that followed, mathematicians and computer scientists tried endlessly to beat this result. They failed. Strassen's algorithm stood as the human record for 50 years.

Why? Because the search space is incomprehensibly vast. For 4x4 matrix multiplication, the number of possible algorithms is larger than the number of atoms in the observable universe. You cannot enumerate them. You cannot search them exhaustively with any amount of human time or computational resources we possessed before the age of deep learning. The space of alternatives was too large for any mind — human or classical computer — to navigate. The algorithm existed in the dark, waiting for the right kind of patient explorer.

The reason we care about multiplications specifically, not additions, is hardware physics. On modern CPUs and GPUs, a multiplication operation takes longer and consumes more energy than an addition. An algorithm that replaces multiplications with additions — even if it adds more additions overall — is typically faster in practice. Strassen's algorithm does exactly this: it uses 49 multiplications but more additions. The net result is faster computation.

The Tensor Game

Playing Mathematics Like Chess

In October 2022, Fawzi et al. published a paper in Nature titled "Discovering faster matrix multiplication algorithms with reinforcement learning." The core innovation was audacious: they reformulated the problem of finding faster matrix multiplication algorithms as a single-player board game, then trained AlphaTensor — a descendant of AlphaGo and AlphaZero — to play it.

The mathematical foundation rests on tensor decomposition. Matrix multiplication can be represented as a three-dimensional object called a tensor. The problem of finding a faster algorithm is mathematically equivalent to decomposing this tensor into a sum of rank-1 tensors (the simplest possible tensors). The number of rank-1 tensors in the decomposition determines the number of multiplications needed. Fewer components equals fewer multiplications.

The researchers called this the "tensor game." The game board is the tensor itself. Each move consists of selecting and subtracting a rank-1 tensor component from the board. The goal is to decompose the entire tensor — reduce it to zero — using as few moves as possible. It is a strikingly elegant reformulation. A hard mathematical optimization problem becomes a playable game.

AlphaTensor is built on the same reinforcement learning architecture that defeated world champions at Go and Chess. But instead of learning to play Go against human players, it learns to play the tensor game by playing millions of games against itself. The system rewards moves that lead toward complete tensor decomposition and penalizes inefficiency. Through this process — trial, error, self-play, and reward backpropagation — AlphaTensor learns to find decompositions that human mathematicians have never discovered. It navigates the search space with patience no human mind can muster.

The key distinction here is important: AlphaTensor did not have a mathematical insight. It did not perceive some clever algebraic manipulation that humans had missed. It had computational patience. It explored a search space systematically and found the answer. This is a fundamentally different kind of problem-solving than the human mathematical intuition Strassen deployed. Both are valuable. But increasingly, patience — the ability to explore vast spaces methodically — is the scarcer resource.

The Breakthrough

Faster Than Strassen

On October 5, 2022, AlphaTensor's discovery became public: it had found an algorithm for 4x4 matrix multiplication using 47 multiplications. Two fewer than Strassen's 50-year-old record of 49. The first improvement in the two-level Strassen algorithm since 1969. The response from the mathematics and computer science community was immediate and intense. Here was a concrete proof that a machine had found something a half-century of human effort had missed.

But why do two fewer multiplications matter? In isolation, on a single operation, the difference is marginal. Yet matrix multiplication is not isolated. It is the atomic operation of computing. A 4% improvement compounds. In cloud computing infrastructure processing billions of matrix multiplications per second, a 4% speedup translates to proportional reductions in latency, power consumption, and cost. More importantly, the specific structure of AlphaTensor's algorithm proved to be well-suited to particular hardware architectures — field-programmable gate arrays (FPGAs), specialized GPUs, and tensor processing units (TPUs). The algorithm was not just mathematically superior; it mapped naturally onto the parallel processing structures of modern silicon.

AlphaTensor's discoveries extended far beyond 4x4 matrices. The same approach discovered improvements for many other matrix sizes — 3x3, 5x5, 6x6 — and for matrix multiplication in modular arithmetic, the finite-field mathematics used in cryptography and error-correcting codes. Some of these results were completely novel; no previous improvement had been found for certain problem sizes. Others were unexpected in their structure. The breadth of discovery suggested that AlphaTensor was not finding isolated anomalies but rather mapping out an entirely new landscape of mathematical possibilities.

After the paper's publication, mathematicians Kauers and Moosbauer conducted an independent verification. They confirmed AlphaTensor's 47-step result and went further, discovering additional 47-step algorithms through entirely different methods. This confirmation was crucial: it proved the result was not a unique artifact of AlphaTensor's particular search but rather a genuine mathematical fact that could be reached via multiple routes. The world's best human mathematicians could now, with effort, find what AlphaTensor had found. But they needed the machine to show them the way first.

The Implications

The Algorithm That Accelerates Itself

There is a peculiar irony worth pausing on. AlphaTensor is an artificial neural network, built entirely on matrix multiplication and tensor operations. It runs on GPUs optimized for matrix operations. It was trained using techniques that, at their core, involve matrix multiplication. And what did it discover? A faster way to multiply matrices. The tool made itself faster. The system used to solve the problem improved the fundamental operation underlying that very system.

This recursive self-improvement is not a harbinger of runaway exponential acceleration or artificial general intelligence. AlphaTensor's discovery is a one-time mathematical fact — the algorithm either exists or it does not. But it does illustrate something important: AI systems trained on general reasoning can find improvements to specific mathematical primitives that AI systems depend on. The feedback is not automatic or continuous. It is a discrete jump: a new algorithm exists, hardware manufacturers implement it, the next generation of neural networks runs slightly faster, and so on. But the principle is real.

The practical implications are already in motion. GPU and TPU manufacturers have shown interest in implementing AlphaTensor's decomposition structures. The specific structure of the 47-step algorithm maps well onto the parallel architecture of graphics processors — the way the multiplications can be grouped and scheduled aligns with how modern silicon can distribute work across thousands of processing cores. Unlike some algorithmic improvements that are elegant in theory but clumsy in practice, AlphaTensor's discoveries are hardware-friendly.

The broader research direction is where the real significance lies. AlphaTensor is not a closed achievement. It is a demonstration of methodology: formulate a mathematical optimization problem as a game, train a reinforcement learning agent to play it, and use the solutions to advance the state of the art. The same approach could be applied to other mathematical primitives — tensor contractions, convolution operations, Fourier transforms, and sparse matrix operations. Each of these operations is foundational to some domain of computing. Each one might harbor algorithmic improvements that human intuition has missed. The methodology could spawn a generation of algorithms discovered not by insight but by systematic exploration.

The philosophical point deserves emphasis. Strassen's 1969 discovery was an insight — a clever observation about algebra that flowed from human mathematical intuition. AlphaTensor found its improvement by systematically exploring a search space too vast for human intuition to navigate. It did not have the insight. It had the patience. That distinction is crucial. And in the age of exponential data growth and computational complexity, patience — systematic, exhaustive search through possibility spaces — is increasingly the scarcer resource.

Volker Strassen's record stood for 50 years. It was an extraordinary human achievement. Not because it was simple — it required deep mathematical insight — but because it was correct and it endured. In October 2022, a machine broke that record not through insight but through the patience to explore what humans could not. The 47-step solution was not invented. It was discovered. It existed in the space of all possible algorithms, waiting for a mind patient enough to find it.