A Google AI beat a 56-year-old maths record and freed up 0.7% of Google's computing power

AlphaEvolve is a Google DeepMind coding agent built on Gemini. It improves programs by trial and error: it proposes changes, scores them with an automatic checker and keeps what works.

Maths. Multiplying two 4×4 matrices (grids of numbers) takes many small multiplications. For complex-valued matrices, the previous best came from applying Strassen’s 1969 algorithm twice: 49. AlphaEvolve found a way with 48. On 50+ open problems in analysis, geometry and combinatorics, it matched the best known answer about 75% of the time and beat it about 20% of the time. One example is a new lower bound for the kissing number in 11 dimensions (how many equal spheres can touch a central one): 593 spheres. In a follow-up with Terence Tao and colleagues on 67 problems, it matched most known results and improved several.

Engineering (Google’s own figures). A scheduling rule it found for Borg, the system that runs jobs across Google’s data centres, recovers on average 0.7% of Google’s worldwide computing power. It also sped up a key piece of Gemini training code by 23%, cutting overall training time by 1%.

What it means

AI can now find small improvements in maths and code that people had missed for decades, and a computer can confirm each one. This only works where every attempt can be scored automatically, so it helps with searching for better examples, not with open-ended research. Most of the gains so far are small steps on known records, plus savings inside Google’s own data centres.

Caveats

• It only works where results can be scored automatically, and mostly finds better examples, not proofs that they are the best possible.
• The 48-multiplication method needs complex numbers. Within a month, human researchers found a 48-multiplication version that does not.
• Most gains are small improvements on known bounds.
• The data-centre and training-code figures are Google’s own and not independently audited.