Unlocking the Future of Intelligence with Geometric Deep Learning
We are building fundamentally new architectures that break neural scaling laws, making AI fast, efficient, and explainable.
Our Vision
Scale-is-all-you-need doesn’t scale
Solution
We see a future with AGI integrated into everyday life, unlocking ubiquitous robotics, autonomous vehicles, generative multimodal media, and fundamental scientific advancements. However, today’s data and power-hungry transformer-based AI models won’t get us there.
A single 100 word ChatGPT response consumes a bottle of water in cooling. Scaling these models requires power plants of energy, hundreds of millions of dollars, and months of training. Now, recent results cast doubt on whether scaling laws will continue to hold at all.
We’re building next-generation frontier AI built on new mathematical principles discovered in neuroscience and geometric deep learning. These structures will enable orders of magnitude better scaling than the transformer-based architectures.
0 / 0
Our Vision
We see a future with AGI integrated into everyday life, unlocking ubiquitous robotics, autonomous vehicles, generative multimodal media, and fundamental scientific advancements. However, today’s data and power-hungry transformer-based AI models won’t get us there.
Scale-is-all-you-need doesn’t scale
A single 100 word ChatGPT response consumes a bottle of water in cooling. Scaling these models requires power plants of energy, hundreds of millions of dollars, and months of training. Now, recent results cast doubt on whether scaling laws will continue to hold at all.
Solution
We’re building next-generation frontier AI built on new mathematical principles discovered in neuroscience and geometric deep learning. These structures will enable orders of magnitude better scaling than the transformer-based architectures.
New structures for perception and cognition
Natural Intelligence
Brains have evolved mathematical primitives for representing the world—efficiently discovering, representing, and reasoning about its underlying structures.
World Models
Structured world models exhibit strong generalization far out of their training distributions, forming a robust foundation for perception, generation, planning, semantic understanding, and reasoning.
We’re building it
New Theory is building state of the art, general-purpose architectures for multi-modal world models trained with a fraction of the data and compute.
