.@GoogleDeepMind presented AlphaEvolve, a coding agent for math,
computing and general-purpose algorithm discovery.
It's powered by:
- Ensemble of top Gemini models to enhance creativity and
- Automated evaluators to verify solutions.
AlphaEvolve is an evolutionary system - its elements work in a loop, constantly trying out changes to the code, checking what works better, and keeping the best options to optimize algorithms.
Details 🧵
1. The working process:
- Firstly, you provide AlphaEvolve an initial version of the code and an evaluation function.
- AlphaEvolve uses Gemini models to suggest code improvements. Gemini Flash maximizes the breadth of ideas, and Gemini Pro gives critical depth to suggestions.
- Every new version of the code is scored by running it through the evaluation function. The best-performing ones are saved and reused for future suggestions.
- The process repeats and the code gets better and better.
- Firstly, you provide AlphaEvolve an initial version of the code and an evaluation function.
- AlphaEvolve uses Gemini models to suggest code improvements. Gemini Flash maximizes the breadth of ideas, and Gemini Pro gives critical depth to suggestions.
- Every new version of the code is scored by running it through the evaluation function. The best-performing ones are saved and reused for future suggestions.
- The process repeats and the code gets better and better.
2. Here's what AlphaEvolve has already achieved:
- It created a better scheduling system for Google’s data centers.
- Simplified computer hardware design, removing unnecessary bits and designing faster arithmetic circuit for matrix multiplication. This is applied in upcoming Google's Tensor Processing Unit (TPU).
- Helped speed up training of Gemini models.
- Achieved up to a 32.5% speedup for the FlashAttention kernel implementation.
- Explored new solutions to open math problems👇
- It created a better scheduling system for Google’s data centers.
- Simplified computer hardware design, removing unnecessary bits and designing faster arithmetic circuit for matrix multiplication. This is applied in upcoming Google's Tensor Processing Unit (TPU).
- Helped speed up training of Gemini models.
- Achieved up to a 32.5% speedup for the FlashAttention kernel implementation.
- Explored new solutions to open math problems👇
Generated by Thread Navigator
Press ⌘ + S to quick-export