Unconventional AI

Unconventional AI is rethinking the physical substrate AI runs on, using physical systems like coupled oscillators rather than conventional digital chip architectures, aiming for what it calls "biology-scale efficiency" — brains do far more computation per watt than today's AI hardware. Its Un-0 model, an image generator built on this approach, achieved a competitive FID score of 6.74 on ImageNet 64x64 with weights and code released openly, and the company has also published DS-ISA, described as the first instruction set architecture designed specifically for dynamical-system hardware, alongside research claiming a 98.4% reduction in model connections without a performance hit. It runs an "Unconventional Grant" program funding outside research into new computing paradigms. The company was founded by Naveen Rao — who previously founded MosaicML (acquired by Databricks for $1.3B) and Nervana Systems (acquired by Intel) — alongside MIT professor Michael Carbin and Stanford's Sara Achour. It raised a $475M Seed round in December 2025 led by Lightspeed Venture Partners and Andreessen Horowitz.