ألفا إيفولف من ديب مايند: ثورة في تصميم الخوارزميات بالذكاء الاصطناعي مدعومة بجيميني

DeepMind has unveiled AlphaEvolve, a groundbreaking AI agent that leverages the power of large language models (LLMs) to design and optimize algorithms for a wide range of applications, from mathematics to practical computing. By combining the creative problem-solving abilities of Gemini models with automated evaluation systems, AlphaEvolve represents a significant leap forward in automated algorithm discovery. This innovation has already demonstrated its potential by enhancing Google’s data center efficiency, accelerating chip design, and improving AI model training, showcasing its broad applicability and potential for transformative impact across various industries.

AlphaEvolve: The Evolutionary Coding Agent

• Core Concept: AlphaEvolve is an evolutionary coding agent powered by Gemini LLMs designed for general-purpose algorithm discovery and optimization.
• Methodology: It couples the creative problem-solving of Gemini models with automated evaluators that verify solutions and uses an evolutionary framework to improve promising ideas.
• LLM Ensemble: It utilizes both Gemini Flash for breadth of exploration and Gemini Pro for in-depth analysis.

Optimizing Google’s Computing Ecosystem

• Data Center Efficiency: AlphaEvolve discovered a heuristic that improved Borg’s orchestration of Google’s data centers, recovering an average of 0.7% of Google’s worldwide compute resources.
  • This translates to increased task completion on the same computational footprint.
  • The solution offers interpretability, debuggability, predictability, and ease of deployment due to its human-readable code.
• Hardware Design: AlphaEvolve proposed a Verilog rewrite for a matrix multiplication circuit that removed unnecessary bits, improving its efficiency.
  • The change has been integrated into an upcoming Tensor Processing Unit (TPU).
  • This fosters collaboration between AI and hardware engineers for faster chip design.
• AI Training and Inference:
  • Matrix multiplication kernel sped up by 23%, leading to a 1% reduction in Gemini’s training time.
  • Optimization of low-level GPU instructions (FlashAttention kernel) resulted in up to 32.5% speedup.
  • Reduces engineering time for kernel optimization, from weeks to days.

Advancing Frontiers in Mathematics and Algorithm Discovery

• Matrix Multiplication: AlphaEvolve discovered new algorithms for matrix multiplication, improving upon existing solutions.
  • It found an algorithm to multiply 4x4 complex-valued matrices using 48 scalar multiplications, surpassing Strassen’s 1969 algorithm.
• Open Mathematical Problems: Applied to over 50 open problems in mathematics, geometry, combinatorics, and number theory.
  • Rediscovered state-of-the-art solutions in approximately 75% of cases.
  • Improved the best-known solutions in 20% of cases.
• Kissing Number Problem: AlphaEvolve advanced the kissing number problem by discovering a configuration of 593 outer spheres and establishing a new lower bound in 11 dimensions.

Future Directions

• Expansion of Capabilities: DeepMind anticipates AlphaEvolve’s performance will continue to improve alongside advancements in LLMs, particularly in coding abilities.
• Early Access Program: They are developing a user-friendly interface and planning an Early Access Program for academic users.
• Broader Applications: DeepMind believes AlphaEvolve can be applied to any problem solvable by an algorithm, with potential uses in material science, drug discovery, sustainability, and other technological and business fields.

AlphaEvolve represents a significant milestone in AI-driven algorithm design. Its ability to improve upon existing solutions and discover entirely new approaches to complex problems positions it as a powerful tool for accelerating progress across a wide range of domains. By democratizing access to algorithm optimization, DeepMind’s AlphaEvolve has the potential to unlock significant advancements in science, technology, and beyond.

Source: DeepMind