UC Berkeley’s OpenEvolve AI Creates Algorithms 5x Faster Than Humans

In what could signal a fundamental shift in how computer algorithms are designed, researchers at UC Berkeley have demonstrated that artificial intelligence systems can now outperform human experts in creating optimized algorithms. According to their recently published preprint paper, the team used an AI coding agent called OpenEvolve to develop a load balancing algorithm that significantly beats prior human designs.

AI-Driven Research Breakthrough

The UC Berkeley team reports that their OpenEvolve implementation, which builds on Google DeepMind’s AlphaEvolve concept, achieved a remarkable 5x speedup for an Expert Parallelism Load Balancer (EPLB) algorithm. This particular algorithm plays a crucial role in large language models, where it routes tokens to specialized expert modules to reduce computational overhead.

What’s particularly striking is how the AI approached the problem differently than human programmers. Sources indicate that OpenEvolve replaced traditional looping structures with vectorized tensor operations and implemented what the researchers describe as a “zig-zag partitioning scheme.” The result? A runtime of just 3.7 milliseconds compared to 19.6 ms for an undisclosed frontier lab’s implementation and 540 ms for DeepSeek’s open-source version.

The Cost of Innovation

Perhaps most surprising is how little it cost to achieve these breakthroughs. Analysis of the research methodology shows that the team used a combination of Gemini 2.5 Flash and Gemini 2.5 Flash Lite models, spending less than $10 over just five hours of computation time. That’s an astonishingly efficient research process compared to traditional methods that might require weeks or months of human engineering effort.

Meanwhile, this isn’t the only success story emerging from AI-driven algorithm design. Google previously highlighted how its AlphaEvolve system improved data center orchestration and optimized matrix multiplication operations in TPU hardware. The timing is notable—just as the UC Berkeley paper emerges, Google DeepMind researchers published in Nature about an autonomous method for discovering reinforcement learning rules.

Redefining Research Roles

The implications for computer science research are profound. The Berkeley researchers argue that we’re entering an era of AI-Driven Research for Systems (ADRS), where AI models iteratively generate, evaluate, and refine solutions. Their paper suggests that human researchers will increasingly focus on problem formulation and strategic guidance rather than hands-on algorithm design.

When questioned about whether OpenEvolve is truly being creative or just brute-forcing solutions, co-author Audrey Cheng offered an insightful perspective. “As researchers, we know that we ‘stand on the shoulders of giants,'” she explained in correspondence. “Only by deeply understanding the ideas of others can we come up with ‘novel’ solutions. The creative process requires known data. OpenEvolve uses this data and applies it to new problems.”

Industry Adoption Already Underway

This isn’t just academic speculation—industry adoption appears to be already happening. Cheng pointed to Datadog’s recent engineering blog about self-optimizing systems as evidence that companies are beginning to embrace these approaches for performance tuning. The current focus on performance problems makes sense because they’re easier to verify objectively, but researchers expect the methodology to expand into security and fault tolerance once robust evaluation frameworks are developed.

The open-source nature of OpenEvolve means other research institutions and companies can build upon these findings. For organizations running systems at scale, the potential efficiency gains could be substantial. Cheng believes most large companies will eventually use some form of ADRS for performance optimization.

What remains unclear is how quickly this approach will spread beyond performance tuning to other complex computer science challenges. The researchers note that the current bottleneck isn’t the AI’s capability but rather having robust evaluation and validation frameworks for problems where correctness is harder to verify than pure performance metrics.

As one industry observer noted, we may be witnessing the early stages of a fundamental transformation in how algorithms are created—not through gradual human iteration, but through AI systems that can explore solution spaces in ways human minds cannot easily conceive.