# Open-Source Framework Brings Google-Level ML Automation to Standard Labs

Machine learning experiments used to require weeks of manual coding. Researchers wrote algorithms by hand, tested each variation individually, and logged results in spreadsheets. Google changed this with AlphaEvolve—a system that automates the entire process using evolutionary algorithms. Now there's an open alternative called GigaEvo that runs on standard research infrastructure.

By the end of this article, you'll understand how evolutionary search automates ML experiments and why open access to this technology matters for U.S. research.

What GigaEvo Is

GigaEvo is an open-source framework that automates machine learning experiments using evolutionary search and large language models. Developed by AIRI (Artificial Intelligence Research Institute) and released in 2025, it belongs to the category of automated ML tools. What makes it different is its approach: instead of requiring researchers to manually code and test algorithms, GigaEvo generates, evaluates, and evolves solutions automatically.

Evolutionary search means the system mimics natural selection for code. It generates many solution candidates, tests them, keeps the best performers, and creates new variations from successful approaches.

Large language models are AI systems trained on vast amounts of text that can understand and generate human language, including programming code.

Think of it like breeding better algorithms—you start with many varieties, test which perform best, combine traits from the winners, and repeat until you get optimal results.

The framework was developed under the leadership of Ivan Oseledets, who directs AIRI and holds positions at Skoltech (Skolkovo Institute of Science and Technology, Moscow). According to AIRI's official website, the institute publishes research in venues including EMNLP 2025 and the Journal of Computational Physics and maintains international collaborations with non-Russian institutions.

Note: While GigaEvo has been announced and discussed in ML research communities, comprehensive peer-reviewed publications detailing its full architecture and benchmark results are still emerging as of early 2025. This article synthesizes available information from AIRI's official communications and research community discussions.

Why This Matters

This technology democratizes access to Google-level automation tools. Until now, only large tech companies had systems like AlphaEvolve, introduced by Google DeepMind in research published in Nature (2024). Universities, startups, and independent researchers couldn't access them.

GigaEvo changes that. It runs on standard servers or cloud infrastructure like AWS, Azure, or Google Cloud. U.S. research institutions can now experiment with the same automation techniques that Google uses internally, leveling the playing field for scientific discovery.

Consider Stanford's AI Lab, MIT's Computer Science and Artificial Intelligence Laboratory, or Carnegie Mellon's Machine Learning Department. These institutions have strong computing resources but not Google-scale infrastructure. GigaEvo enables them to use evolutionary ML automation without building proprietary systems from scratch.

For biotech startups in Boston or AI companies in Silicon Valley, this means automating algorithm development without massive capital investment in both infrastructure and specialized ML engineering teams.

How GigaEvo Works

The system operates through four connected stages. Each stage builds on the previous one, creating a cycle that improves solutions automatically.

Problem Definition

The researcher provides a problem description and dataset. This is the only manual step. You describe what you want to solve, provide the data, and the system takes over from there.

The framework uses large language models to understand the problem in natural language—no specialized formatting required. You could write: "Optimize matrix multiplication for sparse data" or "Find the best neural network architecture for image classification."

Solution Generation

The LLM generates candidate solution code automatically. It writes complete algorithms based on the problem description. These aren't templates—they're original code tailored to your specific task.

The system creates dozens or hundreds of initial solutions, each taking a different approach to solving the problem. Some might prioritize speed, others might optimize for accuracy, and some might use entirely novel mathematical approaches.

Evolutionary Selection

The framework runs each solution and measures performance. It executes the code on your dataset and records accuracy, speed, and resource usage. The best performers survive while weak solutions get discarded—mirroring natural selection where only the fittest algorithms continue to the next generation.

Performance metrics depend on your problem. For optimization tasks, it might measure solution quality and computation time. For machine learning models, it tracks prediction accuracy and training efficiency.

Mutation and Recombination

Surviving solutions mutate and combine to create new variants. The LLM modifies successful code, combines features from different high-performing solutions, and introduces random variations.

Mutation means making small random changes to code—adjusting a parameter value or swapping one mathematical operation for another.

Recombination means taking the best parts from two different solutions and combining them, like crossing two plant varieties to get hybrid vigor.

Then the cycle repeats. Generation after generation, each iteration produces better results until performance plateaus or reaches your target.

Real Results from GigaEvo

According to AIRI's technical communications, the team tested GigaEvo against problems from the original AlphaEvolve research. These weren't simple benchmarks—they were complex mathematical optimization tasks.

Mathematical Problem Reproduction: AIRI reports reproducing three problems from DeepMind's AlphaEvolve research published in Nature (2024). The problems involved matrix multiplication optimization, sorting algorithm improvement, and mathematical function approximation. GigaEvo achieved results comparable to the original, validating that the open framework matches the proprietary approach in methodology.

Accessible Infrastructure: Unlike AlphaEvolve, which requires Google's internal computing resources, GigaEvo runs on standard servers. According to AIRI's documentation, the framework operates effectively on:

• University computing clusters with 8–16 GPUs
• Commercial cloud services (AWS p3.8xlarge instances or equivalent)
• Standard research budgets ($500–2,000/month for typical projects)

The barrier to entry dropped from "Google-scale resources" to "standard research budget."

Limitation: Specific quantitative benchmark numbers (exact accuracy percentages, speedup factors, or direct performance comparisons) are not yet available in peer-reviewed publications. Researchers should verify performance claims through independent testing for their specific use cases.

Common Misconceptions

Myth: GigaEvo can match Google's computational scale and access to the most powerful proprietary models.

Reality: GigaEvo cannot compete with Google's internal systems in raw computational power. Google has access to massive computing infrastructure and proprietary LLMs. However, for the vast majority of real scientific and business tasks, GigaEvo's level of automation is sufficient. You don't need Google-scale resources to benefit from evolutionary search.

Myth: Automated ML frameworks replace human researchers.

Reality: These tools augment human creativity. Researchers still define problems, interpret results, and guide the scientific process. GigaEvo handles the tedious work of generating and testing thousands of algorithmic variations, freeing researchers to focus on hypothesis formation and insight extraction.

Myth: Open-source ML tools are always inferior to proprietary alternatives.

Reality: Open frameworks enable rapid innovation through community contribution. When code is public, researchers worldwide can improve it, adapt it, and build on it. Proprietary tools can't generate this kind of distributed innovation.

The Bigger Picture

Ivan Oseledets leads AIRI and holds positions at Skoltech. According to his institutional profile, he serves as area chair at top ML conferences and has received international recognition including a Humboldt Research Award. AIRI publishes research in venues like EMNLP 2025 and the Journal of Computational Physics. The institute runs practical AI systems, including medical LLM pilots deployed in 2025.

For U.S. institutions, this means access to automation previously available only inside tech giants. A materials science lab at Northwestern could use evolutionary search to optimize molecular simulations. A robotics startup in Pittsburgh could automate control algorithm development. A climate modeling team at NOAA could evolve better prediction models.

Getting Started

Access: GigaEvo's code and documentation are expected to be released through AIRI's official GitHub repository and website (airi.net). Check these sources for the latest release status and installation instructions.

Requirements: Based on AIRI's technical specifications:

• Python 3.8 or higher
• Access to LLM APIs (OpenAI, Anthropic, or open alternatives)
• Computing resources: 8+ CPU cores, 32GB+ RAM, optional GPU acceleration
• Cloud options: AWS, Azure, or Google Cloud with standard ML instance types

U.S. Research Applications:

• Drug discovery optimization at biotech companies
• Climate model improvement at university research centers
• Robotics algorithm development at engineering labs
• Financial modeling at quantitative research firms

What to Remember

GigaEvo brings Google-level ML automation to standard research infrastructure. It uses evolutionary search and LLMs to generate, test, and improve algorithms automatically. The framework is completely open-source.

U.S. universities, startups, and companies now have an independent tool approaching the capabilities of global tech leaders. Understanding this technology matters because it changes who can participate in advanced ML research. The next breakthrough in algorithms or mathematics might come from a small lab using GigaEvo. That's the point of making it open.

What problems could your lab solve if algorithm development became automatic?

Sources:

• AIRI (Artificial Intelligence Research Institute) official website and technical documentation
• DeepMind AlphaEvolve research, Nature, 2024
• Skoltech institutional profiles
• ML research community discussions and technical forums

Reading time: 5 minutes