Sakana AI RSI Lab Formalizes Autonomous AI Research Redesign

Sakana AI established its Recursive Self-Improvement Lab to shift AI development from human-led processes to autonomous, self-evolving systems using foundation models. LLM-Squared (2024), developed with Oxford and Cambridge, produced the DiscoPOP preference optimization algorithm via LLM-driven generational evolutionary loops. Darwin Gödel Machine (2025), in collaboration with UBC researchers, maintained evolving agent lineages that rewrote their own codebases and achieved a 30 percentage point gain on SWE-bench. ShinkaEvolve (2025) demonstrated 150-sample efficiency in discovering a novel MoE load-balancing loss, while ALE-Agent placed first among 804 entrants in AtCoder Heuristic Contest 058 through inference-time scaling.