Sakana AI: Pioneers of the Evolutionary Approach to Artificial Intelligence
Sakana AI was founded in 2023 in Tokyo by a group of former researchers from a leading global technology company. The founding team was led by David Ha, who previously served as a research lead at Google Brain, and Llion Jones, one of the co-authors of the groundbreaking Transformer architecture that changed the world of natural language processing. The company’s name—“Sakana” (さかな)—means “fish” in Japanese and reflects deep inspiration from the collective intelligence of schools of fish in nature. “Our philosophy is based on the belief that the future of AI does not lie in constantly scaling up a single monolithic model, but in the collective intelligence of many smaller, specialized systems that can communicate and collaborate effectively,” David Ha explained in a recent interview with a technology magazine.
A Revolutionary Approach to AI Research
While many large technology companies are investing in ever-larger language models, Sakana AI has taken a different path. Its research is inspired by evolutionary processes in nature and focuses on:
- Multimodel architecture - instead of a single large model, it creates an ecosystem of smaller specialized models.
- Evolutionary algorithms - models gradually evolve and adapt to new tasks.
- Energy efficiency - optimizing performance while minimizing the consumption of computing resources.
- Transparency and interpretability - emphasizing an understanding of AI decision-making processes.
Continuous Thought Machines - the Latest Breakthrough
In May 2025, Sakana AI introduced its latest project - Continuous Thought Machines (CTMs). This groundbreaking architecture is inspired by the functioning of biological brains and their dynamic processing of information over time. “We designed CTMs to mimic the way the human brain synchronizes the activity of neurons and processes information in real time,” the company’s chief technology officer explained. “The models are capable of continuous reasoning, which is much more natural than the discrete, step-by-step processing of traditional systems.” This technology has potential applications in many areas:
- In healthcare for transparent diagnostic systems.
- In finance for risk analysis with clear justifications.
- In industrial automation for real-time performance monitoring.
CTMs were released as an open-source project with demonstration applications and training scripts available on GitHub, underscoring Sakana AI’s commitment to open science and collaboration.
The AI Scientist - an Autonomous Research System
One of the company’s most remarkable achievements is “The AI Scientist,” introduced in August 2024. This system represents the first fully autonomous platform for scientific research capable of: Formulating research questions and hypotheses, Designing and implementing experiments, Analyzing results and creating visualizations, Writing scientific publications in a standard format, Critically reviewing its own work. An improved version, “AI Scientist-v2,” achieved a significant milestone in April 2025 when it produced a scientific paper that was accepted for publication at the prestigious ICLR (International Conference on Learning Representations) conference following standard peer review. This paper was generated entirely by artificial intelligence without human intervention, apart from the submission process itself.
Achievements in National Security
In March 2025, Sakana AI received the Innovative Spirit Award at the US-Japan Global Innovation Challenge, a competition jointly organized by the United States Department of Defense (Defense Innovation Unit) and Japan’s defense technology agency (ATLA). Of 60 global participants, Sakana AI was the only company selected as a finalist in the biodefense and counter-disinformation categories. “In a changing geopolitical environment, it is essential for Japan to develop its own capabilities in defense technologies, including artificial intelligence,” David Ha emphasized. “We cannot rely solely on foreign technologies in critical areas of national security.”
Other innovative Sakana AI projects include:
- TinySwallow-1.5B
A compact Japanese language model developed using the TAID method that achieves state-of-the-art results in its size category. The model is optimized for local operation on ordinary devices, eliminating the need for cloud APIs or extensive infrastructure. - CycleQD
A system focused on lifelong learning for language-model-based agents. It uses evolutionary computing techniques that allow agents to gradually acquire new knowledge without forgetting previous skills—a key step toward creating artificial intelligence systems capable of continuous learning.
Partnerships and the Future
Sakana AI actively collaborates with Japanese corporations including Sony, NTT, KDDI, NEC, and Fujitsu. These collaborations focus on integrating innovative AI technologies into industrial applications and strengthening Japan’s capabilities in fundamental artificial intelligence research. The company has also established partnerships with global technology leaders such as NVIDIA to optimize its models on state-of-the-art hardware.
Sakana AI remains at the forefront of next-generation artificial intelligence research and actively promotes Japan’s sovereignty in critical technologies. The company plans to expand its activities into the public sector and strengthen cooperation with academic institutions. “Our vision is to create artificial intelligence that is not only powerful, but also sustainable, transparent, and aligned with human values,” says David Ha. “We believe that the future of AI does not lie in a single universal system, but in an ecosystem of specialized models that can collaborate harmoniously with one another—similar to organisms in nature.” With ongoing research into Continuous Thought Machines, autonomous scientific activity, and evolutionary algorithms, Sakana AI has the potential to fundamentally change the way we approach the development of artificial intelligence in the years to come.
