AlphaEvolve: OpenEvolve
An AI for Automated Discovery
AlphaEvolve: An AI for Automated Discovery
DeepMind has unveiled AlphaEvolve, a new AI agent designed not to solve specific problems, but to autonomously discover entirely new and more efficient algorithms. Unlike previous models trained for a single task, AlphaEvolve acts as a general innovation engine.
How It Works
AlphaEvolve combines traditional evolutionary algorithms with modern large language models (LLMs) like Gemini. The process works as follows:
- Initialization: The system is given a goal, defined by an “evaluation function” (which scores the quality of a solution), and a basic “code template” to start with.
- Evolutionary Loop: AlphaEvolve enters a rapid, automated loop of trial and error:
- Generate: It uses LLMs to create a large population of code variations (hypotheses).
- Test & Score: Each new algorithm is tested for correctness and performance, and the results are saved.
- Select: Only the best-performing algorithms are kept to “parent” the next generation of code.
3. Evolve: This loop is repeated millions of times, allowing sophisticated and highly optimized solutions to emerge through a process of automated natural selection. This system learns through “in-context learning,” where the base AI model doesn’t change, but it learns from the results of its own suggestions within the process.
Key Breakthroughs and Results
AlphaEvolve has already produced significant real-world results by optimizing the entire technology stack:
- Algorithms: It discovered a faster algorithm for matrix multiplication, beating a 40-year-old standard and leading to a 1% speedup in training and running Google’s Gemini models.
- Hardware: It improved the circuit design of Google’s Tensor Processing Unit (TPU) chip, reducing its physical area and power consumption.
- Software: It optimized low-level compiler instructions for the FlashAttention kernel (a key part of transformer models), speeding it up by 30%.
- Commercial Operations: It found a new optimization for Google’s data center management system (Borg), resulting in a 1% reduction in Google’s massive cloud computing bill.
Limitations
- Not Fully Self-Improving: AlphaEvolve does not yet upgrade its own core intelligence; this still requires human-led updates to the underlying LLM.
- Requires a Score Function: It can only tackle problems where a “better” solution can be clearly defined and scored numerically.
An open-source implementation of the AlphaEvolve system described in the Google DeepMind paper “AlphaEvolve: A coding agent for scientific and algorithmic discovery” (2025).
Prompt Sampler: Creates context-rich prompts containing past programs, their scores, and problem descriptions
LLM Ensemble: Generates code modifications via an ensemble of language models
Evaluator Pool: Tests generated programs and assigns scores
Program Database: Stores programs and their evaluation metrics, guiding future evolution
The controller orchestrates interactions between these components in an asynchronous pipeline, maximizing throughput to evaluate as many candidate solutions as possible.
