AlphaEvolve: How DeepMind's AI Just Broke 56 Years of Mathematical Stagnation

When an AI uses complex numbers to solve a problem mathematicians dismissed as impossible

In May 2025, Google DeepMind quietly announced AlphaEvolve. A year later, it's broken records that stood for decades — including a 56-year-old algorithm that mathematicians assumed was optimal.

This isn't AI replacing mathematicians. It's something more interesting: AI exploring solution spaces humans would never try.

The Strassen Record: 56 Years, Finally Beaten

In 1969, Volker Strassen proved something counterintuitive: you don't need 64 multiplications to multiply two 4×4 matrices. You can do it in 49.

This blew minds. The naive approach — multiply every element combination — seemed like the only way. Strassen showed otherwise, and his algorithm became a foundational result in computer science.

For 56 years, no one improved on it for general 4×4 matrices. Not for lack of trying. Matrix multiplication is everywhere — AI training, graphics, scientific simulations, signal processing. Shaving off even one operation, applied recursively to larger matrices, compounds into massive efficiency gains.

AlphaEvolve found a way to do it in 48 multiplications.

The kicker? It used complex numbers. An approach mathematicians had largely dismissed as unpromising for this problem. The AI didn't know it was supposed to be a dead end.

Google has already integrated similar optimizations into Gemini's training kernels. The recursive self-improvement loop is real: AlphaEvolve makes Gemini training faster, which makes AlphaEvolve better.

How AlphaEvolve Actually Works

Unlike a typical LLM that generates code once and hopes for the best, AlphaEvolve runs an evolutionary loop:

Define the problem: Human provides evaluation criteria and seed code
Generate mutations: Gemini Flash explores breadth (many ideas quickly), Gemini Pro explores depth (sophisticated structural changes)
Evaluate rigorously: Automated evaluators score each candidate
Select and iterate: Best solutions become parents for the next generation
Repeat: Often for hours or days

The insight is that Gemini isn't making random changes. It understands code structure and can propose intelligent modifications — the kind a mathematician might try if they had infinite patience and no preconceptions about what "should" work.

This is evolution with a reasoning engine as the mutator.

The Ramsey Numbers: 5 Records in One Paper

Ramsey theory is notoriously hard. The joke goes: if aliens threatened to destroy Earth unless we computed R(5,5), we should try to compute it. If they demanded R(6,6), we should attack the aliens.

These numbers describe the minimum size of structures that guarantee certain patterns emerge. Progress is measured in increments of 1, and records stand for years.

In March 2026, AlphaEvolve broke five lower bounds simultaneously:

Number	Old Bound	New Bound	Previous Record Held
R(3,13)	60	61	11 years
R(3,18)	99	100	20 years
R(4,13)	138	139	11 years
R(4,14)	147	148	11 years
R(4,15)	158	159	6 years

Each increment of 1 might look trivial. In Ramsey theory, it's a genuine breakthrough. These bounds required finding specific graph constructions that avoid certain patterns — exactly the kind of search AlphaEvolve excels at.

Terence Tao Is Using This

When the world's most famous living mathematician starts collaborating with an AI tool, pay attention.

Terence Tao has been working directly with AlphaEvolve on Erdős problems — a collection of open questions posed by the legendary Paul Erdős. In December 2025, on Problem #1026, AlphaEvolve generated upper bound constructions that revealed an underlying pattern. Tao then proved the pattern rigorously.

His take:

"Tools such as AlphaEvolve are giving mathematicians very useful new capabilities. For optimization problems in particular, we can now quickly test potential inequalities for counterexamples, or to confirm our beliefs in what the extremizers are, which greatly improves our intuition about these problems and allows us to find rigorous proofs more readily."

This is the key reframe. AlphaEvolve isn't proving theorems — it's generating conjectures, finding counterexamples, and exploring construction spaces. The mathematician still does the proving. But the exploration phase, which used to take months of intuition-building, now takes hours.

The Kissing Number: 300 Years Old, Still Getting Updates

How many non-overlapping spheres can touch a central sphere? In 3D, the answer is 12 — a problem that allegedly sparked a debate between Newton and Gregory in the 1690s.

In higher dimensions, we only have bounds. AlphaEvolve found a new configuration in 11 dimensions: 593 spheres touching the central sphere, establishing a new lower bound.

A 300-year-old problem, still yielding to computational exploration.

Already in Production at Google

This isn't vaporware research. AlphaEvolve has been deployed at Google for over a year:

Data center scheduling: A heuristic for Borg (Google's cluster manager) that recovers 0.7% of global compute resources. That's enormous at Google's scale.
TPU design: Circuit optimizations integrated into next-generation silicon. The proposal was "so counterintuitive yet efficient" according to Jeff Dean that it went directly into the chip.
Gemini training: 23% speedup on a key matrix multiplication kernel, 1% reduction in overall training time.
FlashAttention: Up to 32.5% speedup on GPU instructions.

The feedback loop is remarkable: AlphaEvolve optimizes the infrastructure that trains the models that power AlphaEvolve.

What This Means

We're watching a shift in how mathematical research happens.

The old model: brilliant humans stare at problems for years, occasionally having insights. The new model: AI exhaustively explores solution spaces that humans would dismiss, surfacing candidates for humans to verify and prove.

AlphaEvolve using complex numbers for Strassen's algorithm is the perfect example. Mathematicians had good reasons to think that approach wouldn't work. The AI didn't care about those reasons — it just tried everything and found what worked.

This doesn't diminish human mathematicians. It amplifies them. Tao isn't being replaced; he's getting a research partner that can explore thousands of construction variations while he focuses on the creative work of proof and generalization.

The 56-year Strassen record wasn't broken by a smarter human. It was broken by a system that doesn't know what's "supposed" to be impossible.

AlphaEvolve is currently available in private preview on Google Cloud. The paper detailing the matrix multiplication breakthrough is available on arXiv.