Google's AlphaEvolve finds new ways to calculate, and speeds up Gemini itself

LONDON, MAY 2025 - Google DeepMind has presented a system that invents new ways to calculate: AlphaEvolve. The system lets Gemini, Google's AI models, write computer programs and keeps improving them until they work better. With it, DeepMind found a faster way to multiply certain tables of numbers, a calculation step that appears everywhere in computing. It also helped Google speed up its own data centres and the training of Gemini.

How AlphaEvolve works

The method is similar to breeding. You take many variants, keep the best ones, make new variants from those, and repeat this again and again. It works in four steps:

1. One part of the system writes an instruction for the AI models, with an earlier program as the starting point.
2. Two Gemini models write a new program. Gemini Flash, the fastest model, comes up with many different ideas. Gemini Pro, the most powerful model, adds depth.
3. An automatic evaluator runs each program and gives it a score. This is a rule that measures objectively how well the program works, for example how fast or how accurate it is.
4. The programs and their scores go into a database. A calculation method inspired by evolution then picks which programs inspire the next instruction.

So the method only works for problems where the answer can be judged automatically. This is why it suits mathematics and computer science best. Its predecessor FunSearch, from 2023, still made single functions. AlphaEvolve works with complete programs.

What AlphaEvolve found

Multiplying tables. A table of numbers is called a matrix. Multiplying matrices is at the heart of much computer work, including AI. The ordinary method needs 64 multiplications for two 4 by 4 matrices. In 1969, Volker Strassen showed that 49 are enough. AlphaEvolve found a way with 48 multiplications for matrices with complex numbers. According to DeepMind, this improves on Strassen's method for this case.

Google's own systems. AlphaEvolve found a rule for Borg, the program that divides work across Google's data centres. This rule has been running for more than a year and saves on average 0.7 per cent of all of Google's computing power worldwide. It also improved a design for a computing part of a new TPU, the chip Google uses to train AI.

Training AI. AlphaEvolve found a smarter way to split up a large matrix multiplication. As a result, one part of Gemini became 23 per cent faster and the training of Gemini took 1 per cent less time. For one part of the software, FlashAttention, the gain was up to 32.5 per cent. Work that took experts weeks now took days of automatic trials.

Mathematics. On more than 50 open mathematical problems, AlphaEvolve matched the best known solution in about three quarters of the cases, and in about a fifth it went beyond it. For example, it found an arrangement of 593 spheres that touch one sphere in 11 dimensions, a new record.

What happens next

DeepMind is starting an early access programme for selected scientists and is working on a user-friendly interface. It sees possibilities in every field where a solution can be written as a program and checked automatically, such as materials science and drug development.

Looking back from 2026

What makes AlphaEvolve special is the loop. AI models improved software that they run themselves, right up to the training of a next model. One per cent less training time is a modest gain, but it is one of the first concrete examples of AI helping to build better AI. It concerns work that a human can check: every program gets a score from a calculation rule.

A month later, in June 2025, DeepMind followed with AlphaGenome, and in July 2025 a model from Google won gold at the International Mathematical Olympiad. Those are other fields, but the line is the same: AI that finds new solutions by itself in areas where experts can judge them.