Google DeepMind AI systems have made great scientific advances in recent years, from accurately predicting the 3D structure of every protein known in the universe to forecasting the weather with greater accuracy than ever before.
Today, the UK-based lab unveiled its newest advancement: AlphaEvolve. This AI coding agent makes large language models like Gemini more effective at solving complex mathematical and computing problems.
AlphaEvolve uses the same models it is trying to improve. The agent uses Gemini to propose programs written in code that attempt to solve a problem. It runs each code fragment through automated tests to evaluate its accuracy, efficiency, and novelty. AlphaEvolve uses the best-performing code snippets as the basis of the next generation. This process “evolves”over many cycles, better and better solutions. It is, in essence, a self-evolving artificial intelligence.
DeepMind used AlphaEvolve already to tackle data center energy use, design improved chips, and accelerate AI training. Here are its five biggest achievements so far.
1. AlphaEvolve discovered new solutions to some the world’s most difficult maths problems. In 20% of the cases, it improved upon the best-known solution.
The of EU tech
The latest rumblings on the EU tech scene
A story from our wise old founder Boris and some questionable AI artwork. Every week, it’s in your inbox for free. Sign up today!
The of EU tech
The latest rumblings on the EU tech scene
A story from our wise old founder Boris and some questionable AI artwork. Every week, it’s in your inbox for free. Sign up today!
The 300-year-old Kissing Number Problemwas one of them. AlphaEvolve found a lower bound in 11-dimensional space with a configuration consisting of 593 spheres. This was a breakthrough that even mathematicians had not reached.
2. It made Google’s Data Centres more efficient
. The AI agent devised an effective way to manage power scheduling in Google’s Data Centres. This allowed the tech giant’s data centres to improve their energy efficiency by 0.7% in the last year.
3. It helped train Gemini quicker
AlphaEvolve enhanced the way matrix multiplications were split into subproblems. This is a core operation when training AI models such as Gemini. This optimisation increased the speed of the process by 23 percent, reducing Geminiās total training time to 1%. In the world generative AI, each percentage point can translate to cost and energy savings.
4. It co-designed a part of Google’s new AI chip
This agent uses its code-writing abilities to rewire physical objects. It rewrote an arithmetic-circuit portion in Verilog, a chip design language. This made it more efficient. This same logic is being used by Google to develop its future TPU (Tensor Processing Unit), a chip for machine-learning.
5. It beat a legendary 1969 algorithm
Strassen’s algorithm has been the gold standard in multiplying 4×4 complex matrixes for decades. AlphaEvolve discovered a more efficient way to solve the problem — by using fewer scalar multiplicities. This could lead to more advanced LLMs thatheavily rely on matrix multiplication in order to function.
DeepMind claims that these feats for AlphaEvolve are only the tip of the Iceberg. The lab envisions an agent that can solve a wide range of problems, including discovering new materials and drugs as well as streamlining business operations. AI’s evolutionis a hot topic for theTNW Conference,which takes place in Amsterdam on June 19-20. Tickets are now availableandat – use code TNWXMEDIA2025 to receive 30% off.
