AI making AI: Sakana AI launches "RSI Lab"

2026-06-04 · Source: Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Emerging Technologies & Innovation · Depth: Expert, medium

Summary

Sakana AI has launched its dedicated "Sakana AI RSI Lab" in Tokyo, focusing on Recursive Self-Improvement (RSI) technology to redesign the AI development process. This initiative aims to move beyond the current paradigm of massive models reliant on extensive data and computational resources, instead pursuing open-ended systems that collectively self-improve under constraints. The lab builds upon two years of foundational research, including projects like LLM-Squared (2024), which developed the DiscoPOP preference optimization algorithm, and The Darwin Gödel Machine (2025), which doubled SWE-bench performance. Other key contributions include ShinkaEvolve (2025) for sample-efficient program evolution, ALE-Agent (2025) which won a competitive programming contest, and The AI Scientist (2024-2026), a system that automates the entire research process and published in *Nature* on March 26, 2026. Sakana AI envisions a four-stage trajectory from Agent Native Models to Democratized AI, emphasizing computationally efficient self-improvement to enable broader access to frontier AI. The lab is actively recruiting Research Scientists and Software Engineers to advance responsible RSI with verifiable safety measures.

Key takeaway

For AI Scientists and ML Engineers evaluating future AI development strategies, Sakana AI's RSI Lab suggests shifting focus from raw computational scale to computationally efficient self-improvement. Your efforts should prioritize designing AI systems that can autonomously evolve and refine their own architectures, even under resource constraints. Consider integrating mechanisms for recursive self-improvement and verifiable safety from the outset to foster sustainable, democratized AI innovation. This approach could significantly reduce reliance on hyperscaler infrastructure.

Key insights

AI's future lies in open-ended, self-improving systems that evolve efficiently under computational constraints.

Principles

• Innovation thrives under severe constraints, not abundant resources.
• AI progress should prioritize novel mechanisms over computational scale.
• Responsible self-improvement integrates verifiable safety measures proactively.

Method

A recursive cycle where Agent Native Models develop AI Scientists, which in turn generate improved AI models, leveraging evolutionary processes and adaptive sampling for efficiency.

In practice

• Apply LLM-discovered preference optimization algorithms like DiscoPOP.
• Utilize program evolution frameworks for sample-efficient scientific discovery.
• Explore adversarial co-evolution for cybersecurity and automated red teaming.

Topics

• Recursive Self-Improvement
• AI Development Process
• Agent Native Models
• Computational Efficiency
• Evolutionary Algorithms
• AI Democratization

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• Introducing Sakana AI’s Recursive Self-Improvement (RSI) Lab — Sakana AI's RSI Lab aims to achieve exponential AI self-improvement through sample-efficient, adaptive architectures, moving beyond brute-force scaling.
• Transforming financial operations with AI agents: Sakana AI, Software Engineer Interview — Integrating AI agents into finance requires robust quality standards, security, and human-AI collaboration for practical, enterprise-grade solutions.
• Sakana AI bets AI that improves itself can break the compute arms race of frontier labs — Recursive self-improvement allows AI to iteratively enhance its own development, potentially bypassing the compute arms race.
• AI Is Starting to Build Better AI — AI systems are increasingly demonstrating recursive self-improvement capabilities, though full autonomy remains a future challenge.
• Recursive Superintelligence emerges from stealth with $650M raise — Recursive self-improvement is posited as the fastest path to achieving superintelligence.
• Anthropic confronts the RSI clock — AI systems are already significantly accelerating their own development, raising concerns about future autonomous self-improvement.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Blog.