Sakana AI Launches Frontier Intelligence Group to Explore Beyond-Transformer AI

Sakana AI publicly introduced the Frontier Intelligence Group (FIG) on September 18, 2026, describing it as a research collective that had already been growing quietly inside the company. FIG is dedicated to exploring paradigms beyond current transformer-based AI, with a mandate to pursue long-term, speculative research rather than benchmark-driven product work.

Details

• Purpose: FIG holds regular internal discussions on unconventional ideas and collaborates with outside researchers, operating on the principle of “the freedom to fail” without pressure to hit benchmark targets
• Five guiding principles: research freedom, personally meaningful uniqueness, the idea that “greatness cannot be planned,” drawing inspiration from nature/biology, and studying how intelligence actually works rather than chasing performance scores
• Continuous Thought Machines (CTM): reintroduces biological neuron properties such as temporal dynamics and neural synchronization to produce more diverse, dynamic representations
• Augmented Lagrangian Predictive Coding (PC-ALM): a new variant that lets networks up to 1,000 layers deep be optimized with biologically plausible, local learning rules instead of standard backpropagation
• TwELL format with NVIDIA: a sparse data format and custom CUDA kernels co-developed with NVIDIA that Sakana says unlock speed, memory, and energy improvements when running billion-parameter LLMs
• Other projects highlighted: an “AI Picbreeder” experiment testing open-ended creativity in generative image evolution, and “Smart Cellular Bricks,” a physical system demonstrating self-organization and self-repair through local communication between blocks

What happened next

Sakana framed the launch around the industry’s persistent pain points — models that “confidently hallucinate,” struggle to generalize to novel situations, and consume large amounts of energy — arguing biological intelligence offers a template for continuous learning with far less data. FIG’s work sits apart from Sakana’s shipping product line (Fugu, Namazu, Marlin), positioning it as the company’s long-horizon research bet rather than a near-term commercial release; the NVIDIA-developed TwELL kernels are the most concrete technical deliverable disclosed so far, with the rest presented as ongoing exploratory research.