TLDR: OpenEvolve, an AI coding agent, has successfully developed a more efficient algorithm, demonstrating a 5x speedup for an Expert Parallelism Load Balancer (EPLB) used in large language models. This achievement, detailed by UC Berkeley researchers, highlights the transformative potential of AI in systems research.
Computer scientists at UC Berkeley have announced a significant breakthrough in algorithmic optimization, utilizing an AI coding agent named OpenEvolve to develop a more efficient load balancing algorithm. This development, detailed in a preprint paper titled “Barbarians at the Gate: How AI is Upending Systems Research” by 17 UC Berkeley researchers, showcases the profound promise of AI models in discovering and optimizing complex algorithms.
OpenEvolve, an open-source implementation of Google DeepMind’s AlphaEvolve, was specifically employed to enhance an Expert Parallelism Load Balancer (EPLB) algorithm. This EPLB algorithm is crucial for large language models, where it routes tokens to specialized expert modules, thereby improving efficiency by reducing the number of processed parameters. The AI agent achieved an impressive 5x speedup for this algorithm, significantly outperforming previous human-designed versions.
The researchers initiated their work with DeepSeek’s open-source EPLB implementation, noting its inherent slowness due to being written in Python and relying on a linear search for optimal GPU allocation. OpenEvolve’s ability to optimize this process underscores the potential of AI-Driven Research for Systems (ADRS), an iterative approach where AI models generate, evaluate, and refine solutions.
This achievement aligns with earlier discussions from Google in May 2025, where AlphaEvolve was highlighted as an “evolutionary coding agent” capable of improving the efficiency of Google’s data center orchestration, optimizing matrix multiplication operations in its Tensor Processing Unit (TPU) hardware, and refining FlashAttention kernel implementations in Transformer-based AI models. AlphaEvolve, powered by Gemini models, combines creative problem-solving with automated evaluators to verify and improve ideas through an evolutionary framework.
The UC Berkeley team’s success with OpenEvolve further validates the value of AI-based optimization. While previous automated approaches from DeepMind had not yet surpassed human-designed reinforcement learning systems, this new work demonstrates AI’s capability to devise more efficient solutions for critical computing tasks.
Also Read:
- AI Agent Systems Mimic Human Scientists to Tackle Complex Design Challenges
- AI Models Now Design Better Optimization Algorithms for Software Tuning
“Our results highlight both the disruptive potential and the urgent need to adapt systems research practices in the age of AI,” stated the researchers. This sentiment reflects a growing recognition that AI is poised to transform not just how code is written, but how fundamental algorithmic challenges are approached and solved, potentially making parts of computer science once exclusive to top human researchers accessible to AI.
