Sakana AI has inaugurated the "Sakana AI RSI Lab" within its Tokyo headquarters, a specialized unit dedicated to Recursive Self-Improvement (RSI) technology. The company advocates that systems capable of efficient, iterative self-improvement under resource constraints represent the next evolution in AI, contrasting this with approaches that rely solely on massive-scale computational inputs.

Over the past two years, the firm has developed "Agent Native Model" architectures to create a cycle where systems like "The AI Scientist" automate research to build superior models. Notable achievements include "LLM-Squared," where a language model discovered a new preference optimization algorithm; the "Darwin Gödel Machine," which autonomously doubled performance on SWE-bench, achieving a 30-point absolute improvement; and "ShinkaEvolve," which solved optimization problems in just 150 trials. Other developments include the competitive programming-focused "ALE-Agent," the "Digital Red Queen" system for co-evolving vulnerability identification, and "The AI Scientist" for automating research workflows.

The ultimate mission of RSI Lab is to achieve frontier-level performance in compute-constrained environments, thereby democratizing AI. The company's roadmap outlines a four-stage progression: starting from "Agent Native Model" technology, moving to knowledge expansion through "The AI Scientist," achieving full code-rewriting "RSI," and finally enabling widespread AI customization. This approach is aligned with Japan’s Sovereign AI strategy, which promotes domestic development and operation of AI systems.

Moving forward, RSI Lab plans to treat emergent research behaviors and safety challenges as engineering problems, prioritizing verifiable safety designs. The firm has begun recruiting research scientists and software engineers to optimize exploration pipelines, aiming to establish a team that addresses the advancement of computational intelligence as a rigorous engineering discipline.

Sakana AI has inaugurated the "Sakana AI RSI Lab" within its Tokyo headquarters, a specialized unit dedicated to Recursive Self-Improvement (RSI) technology. The company advocates that systems capable of efficient, iterative self-improvement under resource constraints represent the next evolution in AI, contrasting this with approaches that rely solely on massive-scale computational inputs.

Over the past two years, the firm has developed "Agent Native Model" architectures to create a cycle where systems like "The AI Scientist" automate research to build superior models. Notable achievements include "LLM-Squared," where a language model discovered a new preference optimization algorithm; the "Darwin Gödel Machine," which autonomously doubled performance on SWE-bench, achieving a 30-point absolute improvement; and "ShinkaEvolve," which solved optimization problems in just 150 trials. Other developments include the competitive programming-focused "ALE-Agent," the "Digital Red Queen" system for co-evolving vulnerability identification, and "The AI Scientist" for automating research workflows.

The ultimate mission of RSI Lab is to achieve frontier-level performance in compute-constrained environments, thereby democratizing AI. The company's roadmap outlines a four-stage progression: starting from "Agent Native Model" technology, moving to knowledge expansion through "The AI Scientist," achieving full code-rewriting "RSI," and finally enabling widespread AI customization. This approach is aligned with Japan’s Sovereign AI strategy, which promotes domestic development and operation of AI systems.

Moving forward, RSI Lab plans to treat emergent research behaviors and safety challenges as engineering problems, prioritizing verifiable safety designs. The firm has begun recruiting research scientists and software engineers to optimize exploration pipelines, aiming to establish a team that addresses the advancement of computational intelligence as a rigorous engineering discipline.