Developed by DeepMind, the tool uses Google’s Gemini models to propose, test and iteratively improve algorithms for everything from matrix multiplication to data centre orchestration
Google has unveiled a new AI agent that it claims is reshaping how algorithms are discovered, optimised, and applied in real-world computing systems. Dubbed AlphaEvolve, the agent blends the creative potential of large language models with automated testing systems to evolve code that is not only functional, but also efficient and scalable.
Developed by DeepMind, the tool uses Google’s Gemini models to propose, test and iteratively improve algorithms for everything from matrix multiplication to data centre orchestration. The company says the system has already delivered measurable performance gains across its infrastructure, including improvements to chip design and AI model training.
While large language models (LLMs) have been used for tasks like summarisation and code generation, AlphaEvolve goes further. Rather than generating single functions, it can evolve entire algorithms and adapt them to complex environments. Google describes it as an “evolutionary coding agent” — one that uses the speed of its Gemini Flash model to explore diverse ideas, and the depth of its Gemini Pro model to refine the best ones.
Claims Of Efficiency Gains
According to Google, one of AlphaEvolve’s early successes was in enhancing the scheduling system for Borg, its internal data centre orchestration platform. The AI reportedly found a simple but effective heuristic that has led to a 0.7% recovery in compute resources globally — a seemingly modest figure that the company argues adds up to substantial operational gains at scale.
In hardware design, AlphaEvolve suggested a revision to a key arithmetic circuit used in its Tensor Processing Units. The suggested changes — which had to pass rigorous verification — removed unnecessary logic and were integrated into upcoming chips. This points to a collaborative future where AI may assist hardware engineers by proposing human-readable, verified optimisations in standard design languages.
Boosting AI training and model performance
Perhaps more significantly, AlphaEvolve has been used to accelerate the training of Google’s own AI models. In one case, the system discovered more efficient ways to handle matrix multiplication — a computational bottleneck in many machine learning systems — resulting in a reported 1% reduction in training time for Google’s Gemini models. It also sped up FlashAttention, a component of transformer-based AI models, by over 30% in some scenarios.
While such figures sound impressive, experts may remain cautious. Gains in GPU kernel optimisation, for instance, often require highly specialised knowledge and are already the focus of extensive manual tuning. Whether AlphaEvolve’s suggestions will generalise across applications or hold up under broader scrutiny remains to be seen.
Pushing Boundaries Of Mathematics?
Beyond engineering, Google also claims AlphaEvolve has been used to develop new algorithms for matrix multiplication — a foundational operation in both theoretical and applied computing. The tool reportedly built on minimal code scaffolds to propose new gradient-based optimisation procedures, offering potential insights into long-standing mathematical problems.
Yet while the prospect of AI contributing to open mathematical challenges is tantalising, it’s unclear how these contributions compare to traditional research methods or if the findings have been independently validated by the broader academic community.
Between Automation & Oversight
As AI systems increasingly intervene in technical decision-making — from chip architecture to low-level code — concerns around oversight, explainability, and error propagation will likely grow. While Google stresses the interpretability of AlphaEvolve’s output, the sheer complexity of the systems it interacts with could make robust human oversight challenging.
Nonetheless, AlphaEvolve marks a noteworthy step in the automation of algorithm design, suggesting a future where AI agents don’t just assist with coding tasks, but actively contribute to core scientific and engineering breakthroughs.
