Sakana AI
An AI research company that develops systems for automating scientific research workflows, including The AI Scientist prototype, which targets the automation of hypothesis generation, literature review, and experimental design.
Core Facts
Timeline (last 90 days)
AI Scientist v2 breaks away from reliance on human-written templates, introduces agentic tree search, and generalizes to multiple ML domains, at the cost of a lower per-run success rate; one of its workshop papers scored 6.33 at the ICLR 2025 ICBINB workshop, but Sakana acknowledged that none of the three generated papers met the threshold for main-conference publication.
Jurgen Schmidhuber reiterated his 1987 priority claim on recursive self-improvement and disclosed that he has joined Sakana AI as chief scientific advisor.
AI Scientist v3 was released alongside the MASS paper, which describes a framework in which multiple agent instances collaborate, critique each other, and then converge on a research direction; v3 shifts its positioning from a single researcher executing a linear pipeline to multi-agent team collaboration.
Sakana AI achieves recursive self-improvement of models through collective intelligence, building a population of models that search for better weights through a simulated evolutionary process similar to a genetic algorithm.
The research team constructed a test set called the Contradiction Benchmark, containing 1,164 deliberately designed error cases.
In a paper published in TMLR (Transactions on Machine Learning Research), Sakana AI proposed a mechanism called Multi-Layered Review (MLR).
AI Scientist is an autonomous AI scientist project launched by Sakana AI in 2024.
All Facts (10)
Sakana AI's 'AI Scientist' project attempts to have AI complete the full process from ideation to paper writing.
80%Unverified2024年8月Sakana AI联合多所大学发布了"The AI Scientist"论文，展示了能自主产生研究想法、编写代码、运行实验并撰写论文的系统，每篇论文生成成本约15美元
85%UnverifiedAI Scientist v2 breaks away from reliance on human-written templates, introduces agentic tree search, and generalizes to multiple ML domains, at the cost of a lower per-run success rate; one of its workshop papers scored 6.33 at the ICLR 2025 ICBINB workshop, but Sakana acknowledged that none of the three generated papers met the threshold for main-conference publication.
50%UnverifiedSakana AI achieves recursive self-improvement of models through collective intelligence, building a population of models that search for better weights through a simulated evolutionary process similar to a genetic algorithm.
50%UnverifiedThe research team constructed a test set called the Contradiction Benchmark, containing 1,164 deliberately designed error cases.
50%UnverifiedAI Scientist is an autonomous AI scientist project launched by Sakana AI in 2024.
50%UnverifiedAI Scientist v3 was released alongside the MASS paper, which describes a framework in which multiple agent instances collaborate, critique each other, and then converge on a research direction; v3 shifts its positioning from a single researcher executing a linear pipeline to multi-agent team collaboration.
50%UnverifiedJurgen Schmidhuber reiterated his 1987 priority claim on recursive self-improvement and disclosed that he has joined Sakana AI as chief scientific advisor.
50%UnverifiedIn a paper published in TMLR (Transactions on Machine Learning Research), Sakana AI proposed a mechanism called Multi-Layered Review (MLR).
50%Unverified2024年Sakana AI发布的《The AI Scientist》研究首次展示了端到端自动生成机器学习研究论文的可能性
50%