Sakana Fugu
sakana.ai
June 22, 2026
4 min read
🔥🔥🔥🔥🔥
51/100
Summary
Sakana Fugu is a multi-agent system that autonomously enhances a small GPT's training recipe using AutoResearch, which iteratively edits training code and conducts experiments. The AI agent completed 123 experiments over approximately 14 hours on a single H100 GPU, tracking improvements in validation bits-per-byte (BPB).
Key Takeaways
- The AI agent Fugu-Ultra autonomously improved a small GPT's training recipe, achieving the best mean bits-per-byte (BPB) of 0.9774 across 123 experiments on a single H100 GPU.
- Fugu-Ultra outperformed three frontier models in a reading order task for classical Japanese kana, scoring 0.80 on the NED metric, while the best frontier model scored only 0.24.
- In a benchmark for writing a Rubik's Cube solver in Python, Fugu-Ultra successfully solved all 300 cubes, while two other models produced code that crashed without valid solutions.
- The results suggest that orchestrating multiple strong models can lead to superior performance in agentic machine learning research compared to individual frontier models.
Community Sentiment
Positives
- Sakana's approach to model orchestration could streamline AI usage, allowing users to select the best model for their needs without deep technical knowledge.
- The integration of multiple models checking each other is seen as a promising strategy, potentially leading to better performance and user outcomes.
- The team behind Sakana is perceived as intelligent and capable, which raises expectations for their product's success.
Concerns
- Concerns about Sakana's involvement in military contracts may deter potential users who prioritize ethical considerations in AI development.
- The reliance on commercial models accessed via API instead of open-source alternatives limits the potential for innovation and accessibility in AI applications.
- High subscription costs for multiple AI services are seen as unsustainable, leading to fears of a price race to the bottom.