Sakana AI and UC Berkeley researchers introduced a recursive self-improvement method where a single model acts as solver, judge, and optimizer, though it remains prone to shared blind spots.
Sakana AI has introduced an agentic review system that identifies 73% of core-claim errors in research papers at a cost of $0.47 per review.
MIT and Sakana AI have developed SIFT, a framework that reduces the computational cost of verifying self-improving coding agents by using a language model as a referee.
Sakana AI researchers have developed PC-ALM, a layer-local training method that enables the training of 1000-layer networks by using augmented Lagrangian optimization.
Sakana AI has launched Fugu Max and Fugu Ultra v2, new models for multi-agent orchestration that aim to reduce costs and vendor dependency.
Sakana AI and MIT researchers used GPT-4.1 mini and evolutionary algorithms to create AI warriors that defeated 317 human-coded programs in Core War.
