AI Matrix Multiplication: Count Operations, Then Measure Time

An AI system finds a way to perform a calculation using fewer multiplications. That sounds like a straightforward speed improvement. Yet two questions remain: does the new method produce the right answer, and does it finish sooner on the machine that will actually run it? Neither question should disappear inside the word “efficient”.

For business owners and independent creators assessing AI research, this distinction matters beyond mathematics. A promising shortcut can be real while its practical benefit remains conditional. AlphaTensor offers a useful case study in reading an achievement carefully enough to appreciate what it establishes.

What AlphaTensor actually investigated

Published online on 5 October 2022, the original AlphaTensor paper by Alhussein Fawzi and colleagues used deep reinforcement learning to search for matrix multiplication algorithms. The search worked with tensor decompositions: representations from which multiplication procedures could be recovered. This is established research from 2022, rather than a new announcement in 2026.

The research distinguished improving an arithmetic operation count from optimising execution on particular hardware, including NVIDIA V100 GPUs and Google TPU v2 hardware. That distinction is central to the result. An algorithm can be assessed against a mathematical objective or against measured execution time; those objectives need not select the same procedure.

The accompanying research repository makes the separation visible. Its algorithms directory provides tensor factorisations, with separate files for standard arithmetic and arithmetic modulo two. Its benchmarking directory contains a script for evaluating actual speed on an NVIDIA V100 GPU. Having a factorisation and having a hardware measurement are different pieces of evidence.

A small example you can inspect

To see why counting multiplications is interesting, consider a much older method: Strassen’s algorithm, introduced in his 1969 paper, “Gaussian Elimination is not Optimal”. The example below illustrates its seven-product construction. It is an educational calculation, not an output obtained from running AlphaTensor.

Take matrix A with first row (2, 3) and second row (4, 1). Matrix B has first row (5, 2) and second row (1, 6). Ordinary row-by-column multiplication gives a result with first row (13, 22) and second row (21, 14). For example, the top-left entry is 2 × 5 + 3 × 1 = 13. Computing all four entries this way takes eight scalar multiplications.

Instead, calculate these seven intermediate products, labelled M1 to M7:

• M1 = (2 + 1) × (5 + 6) = 33.
• M2 = (4 + 1) × 5 = 25.
• M3 = 2 × (2 − 6) = −8.
• M4 = 1 × (1 − 5) = −4.
• M5 = (2 + 3) × 6 = 30.
• M6 = (4 − 2) × (5 + 2) = 14.
• M7 = (3 − 1) × (1 + 6) = 14.

Recombine them to obtain the four entries: top-left is M1 + M4 − M5 + M7 = 13; top-right is M3 + M5 = 22; bottom-left is M2 + M4 = 21; bottom-right is M1 − M2 + M3 + M6 = 14. The answer matches, using seven multiplications, alongside additional additions and subtractions.

For this article, a small independently written JavaScript check compared that construction with ordinary multiplication for every pair of 2 × 2 matrices whose entries came from −1, 0 and 1. All 6,561 input pairs matched. This finite check supports the implementation used in the example; it is not a proof for every possible input. No AlphaTensor training or GPU benchmark was run.

Correct arithmetic needs the right setting

The arithmetic domain is part of the claim. Under ordinary arithmetic, 1 + 1 equals 2. Modulo two, it equals 0. A procedure established in one setting cannot simply be advertised as an improvement in the other without checking that it applies. The repository’s separate factorisation files are a useful reminder to retain that qualification when discussing the work.

The same reading habit applies to less mathematical AI stories. Before repeating a result, identify the input, the permitted operations and the required output. A discovery becomes more informative when its boundaries travel with it. Removing those boundaries can turn a precise achievement into an unsupported general promise.

Counted work is only one performance measure

Our seven-product example establishes no speed advantage on a laptop, a server or a production application. We did not time it. Additional arithmetic, the implementation and the computing environment remain relevant to an actual comparison. Saving one kind of operation is a reason to investigate performance, not a substitute for measuring it.

Imagine a small creative studio commissioning an analytics tool. A supplier proposes an AI-discovered numerical routine. The studio should request a comparison using its intended workload, alongside a clear account of what was timed. Does the measurement cover the calculation alone, or the process from receiving input to returning usable output? This is a hypothetical procurement example, not a report of a product deployment.

A practical evaluation brief could ask for four things:

• The exact calculation, input sizes and arithmetic or numerical format being evaluated.
• A correctness check appropriate to the intended outputs.
• A named baseline, with the machine and software conditions recorded.
• Repeated timing measurements with a clear start and finish, including any preparation excluded from the timed section.

Those requests help distinguish a faster numerical component from a faster service. They also make a negative result useful: a method might satisfy its mathematical objective while offering no measured benefit in the particular workflow being considered.

Keep the discovery and the evidence together

There is a related lesson in separating proposed constructions from deductions in AI geometry proofs: record what the system generated and how the result was checked. Here, preserve the algorithm, the arithmetic setting, the correctness evidence and the timing conditions.

AlphaTensor is interesting because algorithm discovery itself can be a research target. Readers do not need to turn every such discovery into a universal acceleration claim to recognise its value. Count the operations when that is the objective. Measure elapsed time when speed is the promise. Ask for both when a decision depends on both.

Sources

• Fawzi and colleagues, “Discovering faster matrix multiplication algorithms with reinforcement learning”, Nature; published online 5 October 2022.
• Google DeepMind’s AlphaTensor research repository, algorithms and benchmarking documentation; consulted 30 October 2026.
• Volker Strassen, “Gaussian Elimination is not Optimal”, Numerische Mathematik, 1969; original paper hosted by the University of Illinois.