Un-0: The First Large‑Scale Generative Model Built on Physical Computation, Targeting 1000× AI Energy Savings

Background and Motivation

For more than a decade, AI has been dominated by GPU‑centric digital computation, scaling up clusters, bandwidth, and model parameters while energy consumption has become a structural bottleneck. As models approach the trillion‑parameter scale, the industry questions where the electricity will come from.

Physical Computing as a New Primitive

Former Databricks AI leader Naveen Rao and his startup Unconventional AI introduce Un‑0, the first large‑scale generative model that treats a physical system—specifically a network of coupled oscillators—as the basic computational primitive. By letting the natural time‑evolution of a physical dynamical system perform the computation, they aim to cut AI inference energy by roughly a thousandfold.

Model Architecture

Un‑0’s engine is a large ensemble of Kuramoto oscillators whose coupling strengths form a learnable matrix

. Each oscillator i has a phase θ_i∈[0,2π) and a natural frequency ω_i. The dynamics follow the Kuramoto ODE (shown in the article) where every oscillator is pulled toward synchrony by the weighted sum of phase differences from all other oscillators. The learnable parameters are the coupling matrix K and the natural frequencies ω, together defining the physical system.

The generation process consists of five steps:

Randomly initialize all oscillator phases (analogous to noise in diffusion models).

Inject class information via a small set of “condition oscillators”.

Let the physical system evolve freely under its dynamics.

Capture a snapshot of all phases at a chosen time T, forming a latent grid.

Decode the latent grid into pixels with a lightweight decoder that accounts for less than 13 % of total parameters.

Training Procedure

Training is performed on CIFAR‑10 and ImageNet‑64×64 using a newly proposed “Drifting Loss” together with a DINOv2 feature extractor and AdamW optimizer. The loss is computed from generated samples themselves, requiring multiple feature‑view evaluations and making it the primary computational bottleneck.

All CIFAR‑10 models train on a single B200 GPU; ImageNet models train on eight B200 GPUs. The largest ImageNet model (16 384 oscillators, ~322 M parameters) consumes 640 B200‑GPU‑hours.

Results and Scaling

FID scores improve consistently as the number of oscillators grows. The top ImageNet‑64 model reaches FID 6.74, comparable to early releases of mainstream generators. Compared against a range of baselines—NCSN, DCGAN‑TTUR, WGAN‑GP, BigGAN, iDDPM, Consistency Models, TRACT—Un‑0 matches or exceeds many early models but still trails state‑of‑the‑art diffusion models such as EDM and GDD.

Implications and Future Outlook

Un‑0 proves that large‑scale image generation can be realized on a physics‑driven dynamical system, opening a pathway toward hardware that merges computation and memory in a single physical substrate. Because the dynamics inherently tolerate noise, they may also reduce communication energy that dominates von Neumann architectures. Rao emphasizes that computation is not uniquely human; it exists throughout the physical world, and exploiting the time dimension of physical processes could unlock the promised thousand‑fold energy gains.

While current software simulations do not yet surpass the performance of conventional AI, Un‑0 serves as a “Hello World” for non‑traditional AI hardware and suggests a promising research direction for future energy‑efficient AI systems.