Introducing Sakana AI's Frontier Intelligence Group (FIG)

Sakana AI has set up the Frontier Intelligence Group, or FIG, as an in-house effort to look past the current Transformer-heavy playbook. The company says today’s systems are impressive, but still fall short in familiar ways: they hallucinate, they stumble on novel situations, and they burn a lot of energy doing it.

That’s the real argument here. Sakana AI isn’t saying scale has failed. It’s saying scale may not be the whole story. Llion Jones, the company’s CTO and one of the inventors of the Transformer, has already pushed this question in the Transformer vs Post-Transformer debate and again in a 2025 TED Talk: if the present paradigm is so strong, it can also crowd out the kind of research freedom that produced it in the first place.

FIG is meant to push against that drift. The group holds regular meetings on ideas that run counter to the mainstream, brings in outside researchers from different backgrounds, and gives people room to work on speculative projects without the usual pressure to keep benchmark numbers moving. Sakana AI says the point is not just better scores. It is to study intelligence itself, while also looking for systems that can learn with something closer to the data and energy efficiency of the human brain.

The company leans on nature for a comparison. Biological intelligence learns continuously, generalises from far less data, and explores with open-ended curiosity. By contrast, current neural nets are still a long way from the brain in how they are built and updated. Sakana AI says that gap may be telling us something important, even if nobody thinks a literal brain copy is the answer.

FIG’s research agenda is broad on purpose. The group says it spans new architectures, optimisers, and objectives, from individual neurons to groups of agents. Its guiding ideas are research freedom, uniqueness, a belief that greatness cannot be planned, a nature-inspired outlook, and a focus on understanding intelligence directly. The company says the project began as a casual weekly chat between two researchers and has grown into a larger collective looking for the next paradigm shift.