Google DeepMind AlphaEvolve
Available to academic researchers, the AI agent for algorithm design based on Gemini (a combination of Flash and Pro), which combines the creativity of large language models (LLMs) with automated evaluators using metrics for discovering and optimizing algorithms. It uses an evolutionary approach to improve the best ideas.
Where is it already used?
1. Google Data Center Optimization 🖥️
- AlphaEvolve found a more efficient algorithm for resource allocation in Borg (Google's data center management system).
- Result: +0.7% of Google's global computing resources are now used more efficiently.
2. Hardware Design 💻
- Optimized matrix multiplications in TPUs (Google's specialized chips for AI).
- Accelerated the operation of arithmetic circuits while maintaining correctness.
3. Accelerating AI Training ⚡
- Reduced Gemini training time by 1% by optimizing matrix operations.
- Accelerated FlashAttention (a core algorithm for transformers) by 32.5%.
Improved Strassen's algorithm (1969) for 4x4 matrices, reducing the number of operations. Improved the best solutions for 20% of open problems in mathematical analysis, geometry, and combinatorics.
Interestingly, AlphaEvolve was used to optimize components involved in training the Gemini models themselves. This raises questions about the potential for recursive AI self-improvement and the approach towards a "singularity".