Google DeepMind's AlphaEvolve automates code discovery and breaks new math records
Created on May 15|Last edited on May 15
Comment
In a major leap for automated scientific discovery, Google DeepMind has unveiled AlphaEvolve, an ambitious AI agent that autonomously invents, evolves, and optimizes algorithms. Leveraging the generative power of Gemini’s cutting-edge large language models, AlphaEvolve marks a turning point in how computers can tackle previously intractable problems in mathematics, computer science, and engineering.
From Writing Code to Inventing Algorithms
Previous AI code systems, including well-known coding assistants, could draft helpful snippets or solve limited textbook problems. AlphaEvolve aims higher: it uses an automated, evolutionary process not just to generate code, but to out-innovate known human solutions. “AlphaEvolve doesn’t stop at code suggestion—it launches thousands of code experiments, automatically tests them, and keeps mutating candidates until true breakthroughs emerge,” said a DeepMind spokesperson.
At the heart of AlphaEvolve is a large-scale automated search. The system begins with a working (if imperfect) program supplied by the user and a machine-testable evaluation function. It then runs a closed feedback loop: using Gemini Pro and Flash language models, it proposes changes—sometimes subtle, sometimes drastic—to the codebase. These “child” versions are executed and scored. AlphaEvolve tracks not just the best but also a diverse array of candidate programs, ensuring it explores new algorithmic ideas rather than getting stuck on familiar ground.
Technical Details: How AlphaEvolve Works
Using AlphaEvolve involves the following steps:
- Setup: Users provide any baseline code and a function to evaluate solutions quantitatively—such as accuracy or execution time. Code regions to be evolved are clearly marked.
- Mutation: The system maintains a database of all program variants. It feeds promising or diverse candidates to an ensemble of Gemini LLMs, which propose new code “diffs”—everything from single-line tweaks to large algorithmic rewrites.
- Evaluation: Child programs are synthesized and tested, with results instantly fed back into the pool.
- Evolution: The database balances exploiting high-performers and exploring novel solutions, using evolutionary strategies like MAP-Elites to drive genuine innovation at scale.
This entire loop runs asynchronously across distributed compute, allowing AlphaEvolve to test thousands of code variations per day with minimal human intervention.
Real-World Impact Already Delivered
AlphaEvolve isn’t just a lab prototype. It’s already improved the backbone of Google’s infrastructure, squeezing a 0.7% efficiency gain out of the Borg data center scheduler—a massive savings at Google scale. The system sped up Gemini’s own matrix multiplication kernel by 23%, directly cutting model training times, and proposed Verilog improvements for custom TPU chips. In some cases, AlphaEvolve has slashed GPU kernel times by up to 32%, delivering optimizations beyond what even expert engineers typically attempt.
Record-Breaking Results in Mathematics
Beyond industrial applications, AlphaEvolve is proving itself on pure mathematics problems—a traditional human stronghold. On benchmarks spanning 50+ algorithmic challenges, the system rediscovered or improved known solutions in 95% of cases. Most dramatically, it broke a 56-year-old record for multiplying 4×4 complex matrices—designing an algorithm faster than any previously published. AlphaEvolve also made progress on the classic kissing number problem, a staple of high-dimensional geometry.
A Glimpse of the Future
Experts at DeepMind envision AlphaEvolve as a general-purpose discovery engine for any field expressible in code. Early Access programs for outside researchers are in the works, letting scientists apply AlphaEvolve to fields like scientific simulation, materials discovery, or even pharmaceutical R&D—so long as solutions can be tested automatically.
By shifting algorithmic discovery from intuition and hand-tuned experiments to high-throughput, AI-driven evolution, AlphaEvolve may change what is computationally possible. As DeepMind ramps up access and Gemini’s models grow more capable, AlphaEvolve’s disruptive potential only stands to increase.
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.