No Priors• Jun 26, 2025• 42mInterview

No Priors Ep. 120 | With Google DeepMind’s Pushmeet Kohli and Matej Balog

From No Priors

Pushmeet Kohli and Matej Balog•Researchers, Google DeepMind

Executive Summary

AlphaEvolve is an AI coding agent from Google DeepMind that combines Gemini models with evolutionary search to discover novel, more efficient algorithms for scientific and computational problems.
The system has demonstrated practical value by discovering algorithms that have been deployed to improve efficiency across Google's infrastructure, from data centers to hardware design.
AlphaEvolve has been used to optimize its own training infrastructure, representing a concrete, early example of recursive self-improvement in AI, which could accelerate future AI development.
A key requirement for AlphaEvolve is a well-defined evaluation function to score potential solutions, though future research aims to relax this constraint using LLMs as evaluators.

9 quotes

Concerns Raised

The significant computational resources required to solve difficult problems may limit broad accessibility.
The current dependency on well-defined, strict evaluation functions constrains the types of problems the system can tackle.
It remains an open question whether AI self-improvement will result in continuous exponential gains or diminishing returns over time.

Opportunities Identified

Accelerating scientific discovery in foundational fields like mathematics and computer science.
Unlocking significant efficiency gains and cost savings in large-scale computational infrastructure.
Broadening the application of AI to more complex problems by developing LLM-based and simulation-based evaluators.
Using AI-generated algorithms to provide novel insights and deepen human understanding of complex systems.

Key Themes

AI for Algorithm Discovery

AlphaEvolve represents a shift from AI merely applying known methods to discovering fundamentally new ones. By combining the creative potential of LLMs with the systematic exploration of evolutionary search, it can navigate vast, non-intuitive search spaces to find solutions that have eluded human experts for decades, such as in matrix multiplication.

This capability allows AI to contribute to core computer science and mathematics, potentially accelerating scientific breakthroughs and creating significant economic value through more efficient computation.

Recursive Self-Improvement

The system has been successfully applied to optimize the efficiency of its own training infrastructure, creating a feedback loop where AI makes itself faster and better. This demonstrates a practical, albeit early, form of recursive self-improvement.

This is a critical step towards AI systems that can accelerate their own development. While it's an open question if this will lead to continuous gains or diminishing returns, it has profound implications for the future pace of AI progress.

The Critical Role of Evaluation

The success of AlphaEvolve is heavily dependent on having a precise, automated evaluation function that can score the quality of generated code. This is currently a major constraint, limiting its application to problems where such an evaluator can be defined.

This highlights the importance of investing in high-quality simulators and evaluation metrics. Future progress in relaxing this constraint, perhaps by using other LLMs as evaluators, will be key to broadening the applicability of this technology to more complex, open-ended problems.

Human-Interpretable AI Solutions

Unlike black-box neural networks, AlphaEvolve produces human-readable code. This allows engineers and scientists to inspect, verify, and understand the novel solutions it discovers. In some cases, the structure of the AI-generated algorithm itself provides new insights, as seen with the cap set problem.

Interpretability fosters trust and enables a collaborative paradigm where AI not only provides better solutions but also deepens human understanding, acting as a tool for both discovery and education.

Get started free

Topics

AlphaEvolve Algorithm Discovery Evolutionary Search Large Language Models (LLMs)Gemini Recursive Self-Improvement AI Coding Agents Computational Efficiency Google DeepMind AlphaTensor FunSearch Matrix Multiplication Scientific Discovery Automated Evaluation Human-Interpretable AI

Processed Mar 31, 2026 yt-dlp + mlx-whisper + Gemini