What makes the AI of Japan’s Sakana unique is that they developed an AI that can hold its own against the most powerful Anthropic’s models without being one itself. This AI system is called Fugu, and it works through coordinating a group of available AI models instead of using one giant model. Rather than seeing it as a model, you should see it as a conductor that chooses the correct combination of models to perform a specific task and produces one final output through the coordination. This concept was called “multi-agent system as a model” by Sakana, and the numbers prove how impressive it can be on paper.
Also read: Kunal Shah as WhatsApp chief: Why Meta and Zuckerberg picked CRED co-founder?
When it comes to some important tests such as SWE Bench Pro (software engineering), LiveCodeBench (coding), and GPQA-Diamond (graduate scientific reasoning), Fugu Ultra, the highest level of this system, rivals Anthropic’s Fable 5 and locked Mythos Preview models.
In this case, it turns out that Fable 5 and Mythos Preview do not belong to Fugu’s agent pool. According to Sakana, it was done intentionally because these models are not publicly accessible, meaning that the system’s benchmark numbers are achieved exclusively through models that are freely accessible to anyone. The point is that a Japanese startup has managed to come up with something similar to Anthropic’s system developed for many years in secrecy.
Also read: The Mac Mini is the best on-device AI computer you can buy: Here’s why
Of course, how this benchmark parity will manifest itself in practical terms remains to be seen. Multi-agent systems tend to be more difficult to evaluate compared to single-model systems, and their orchestration increases latency, complexity and costs which may not always be reflected in headline figures. However, Fugu’s design is great for structured tasks where you can properly route your requests between models.
However, the main thing remains the same. When a multi-agent architecture can reach the level of performance of a frontier model using exclusively publicly available pieces, it redefines the idea of cutting-edge AI. It’s not necessary to have the best model; sometimes, you only need to combine the existing models properly.
Founded by Google Brain alumni David Ha and Llion Jones, Sakana AI has been moving in the direction of nature-inspired and evolutionary AI for quite some time already, since 2023. This is one of its most direct attempts to engage in the frontier lab discussion, but probably not the last one.
Also read: Can a thin and light gaming laptop really handle AAA games? I found out