Over the past decade-plus, GPU-centric digital computing has dominated AI. Bigger clusters, higher bandwidth, more powerful GPUs, and denser data centers have seemed to be the mainstream path toward the next generation of AI.

But as model parameter counts move toward the trillion scale, the industry has begun talking more often about energy consumption. A more fundamental question has followed: if AI keeps scaling the way it does today, where will the electricity come from?

AI’s electricity bill and energy use have clearly evolved from operating costs into a structural bottleneck for the entire industry.

Against this looming energy crunch, Naveen Rao, the former head of AI at Databricks and a well-known Silicon Valley founder, has stepped into the spotlight with his new hard-tech startup, Unconventional AI.

Today, Unconventional AI officially released its first model, Un-0, an image-generation model powered by an “analog coupled oscillator system.” It can be seen as an early example of an emerging physical computing foundation. On ImageNet 64x64, Un-0 achieved an FID of 6.74, bringing its quality close to where some mainstream conventional image-generation methods stood when they were first introduced.

Rao called it “the first large-scale generative model built with physics as a computational primitive.”

“This marks a ‘Hello World’ moment for physics-based models. We use the natural time-varying behavior of physical systems to do computation for us. The result is a fundamentally new way to build computers, with the potential for major gains in energy efficiency.”

In media interviews, Rao even set a more ambitious “small goal”: in the future, AI inference energy consumption could be reduced to one-thousandth of today’s systems.

Sample trajectories of Un-0’s generation process as it evolves over time. The color of each line corresponds to a similarly colored box, which labels the category and shows how images in that category are gradually generated over time.

The company published a blog post introducing Un-0. Here is a closer look.

Un-0’s Starting Point: Rebuilding AI Computing With Physical Systems

Unconventional AI says its goal is to build a new kind of computer that uses the laws of physics to compute, with the hope that modern AI can eventually run at far lower energy levels than today’s machines, targeting roughly a 1,000-fold reduction in energy use.

So the company posed a question: can a physical dynamical system be trained to generate images on tasks at scale?

Today’s strongest AI models are largely conventional deep networks, especially Transformer-based models. But outside the mainstream path, researchers have long explored whether the dynamics of physical systems can improve energy efficiency, including noise, time variation, voltage, and current in analog circuits. These approaches do not compute with traditional digital values; they use the evolution of the physical system itself.

Examples include neuromorphic computing, Hopfield networks, and reservoir computing, as well as more recent approaches such as Hamiltonian Networks, Liquid Networks, Neural Wave Machines, Thermodynamic Computing, and Kuramoto Oscillators.

Un-0 is a new attempt along these unconventional computing paths. The core difficulty is that, to use these alternative computing methods, AI tasks must be mapped effectively into the dynamics of a physical system. What Un-0 aims to test is whether modern AI workloads can run on a physical substrate and ultimately become more efficient than today’s hardware.

How Un-0 Works

The company says to imagine two metronomes ticking side by side, as shown below.

Each metronome has a “phase” at any given moment, meaning the current position of its arm within the swing cycle. If the two metronomes sit on the same table, they influence each other through the tabletop. Depending on the strength of that interaction, or coupling strength, they may gradually synchronize, or they may enter a synchronized state with opposite phases.

That is the basic idea of an oscillator: each oscillator has its own phase and tends to rotate at its own frequency, while also being affected by nearby oscillators.

When two oscillators are scaled up to thousands, the system becomes more interesting. Many oscillators, connected by couplings of varying strength, self-organize into patterns through their interactions, as shown below.

Un-0’s compute engine is exactly such a large-scale population of oscillators, with the coupling strengths between oscillators serving as the model’s main learnable parameters.

These coupled oscillators are usually modeled as Kuramoto oscillators.

More specifically, each oscillator follows a simple rule that applies continuously over time: it rotates at its own natural frequency while being pulled and shifted by all the other oscillators.

The following ordinary differential equation describes how these oscillators evolve over time:

Each oscillator i has a phase in [0, 2pi), where it represents the oscillator’s natural frequency. The matrix specifies the coupling strength, determining how strongly oscillator j pulls oscillator i toward synchronization, or pushes it away from synchronization.

What Un-0 needs to learn is the coupling matrix K and the natural frequencies omega. Together, these parameters define the physical system itself.

Unconventional AI gives two reasons for choosing oscillators:

The first comes from the brain. Rhythmic activity and synchronization are widespread in the brain, and researchers have long believed these phenomena may take part in computation, such as binding distributed features into a coherent percept, controlling communication between brain regions, and organizing the timing structure of neural spikes. Coupled oscillators are among the simplest mathematical models for describing this kind of behavior, making them a natural basic unit for neurally inspired computing models.

The second reason is more engineering-driven: oscillators can be implemented as a physical circuit primitive. Unconventional AI believes coupled oscillator systems can be implemented directly in CMOS or on other physical substrates, allowing the system’s physical behavior itself to compute the dynamical evolution.

The bet behind Un-0 is this: if the laws of physics can compute AI workloads directly, then the execution substrate of the future could look very different from today’s GPUs.

Un-0’s Model Architecture

Generating an image with Un-0 roughly involves five steps:

Random initialization: set the phases of all oscillators to random angles, similar to random noise in diffusion models.

Class-conditioned input: use a smaller set of “conditioning oscillators” to input class labels, such as “volcano” or “daisy,” guiding the main oscillator cluster to evolve in a specific direction.

Let the physics run: release the system and allow the oscillators to pull on one another, evolve under the physical dynamics, and eventually settle.

Capture a snapshot: at a specific time T, record the phases of all oscillators to form a digital latent grid.

Render pixels: use a conventional decoder, which accounts for less than 13% of the model’s parameters, to turn the phase grid into the final image pixels.

The coupled oscillators evolve over time under the learned coupling relationships. A one-way low-rank class-conditioning matrix connects the conditioning oscillators to the main oscillator pool, injecting class information. At time T, the system reads the oscillator states through a decoder and generates an image. By sampling different initial conditions multiple times, it can generate the corresponding image distribution.

During training, the model mainly learns three types of parameters: how oscillators are coupled, namely the matrix K; each oscillator’s natural frequency; and the decoder weights. Overall, the oscillator system takes on computation that might otherwise be handled by conventional neural network layers.

Unconventional AI says it chose this architecture to give the dynamical system itself maximum freedom to perform computation.

In the forward pass during training, the model only needs to set the coupling matrix, oscillator frequencies, and initial phases, then let the dynamical system evolve and finally read out the image latent variables.

This differs from dynamic generative methods such as diffusion models and Flow Matching. Diffusion and Flow Matching typically guide explicitly how the dynamical system should evolve during training, while Un-0’s approach is more like looking only at the final generated samples and then using the loss function to optimize the entire dynamical system backward.

The cost is that it requires a more complex loss function, because the training signal comes mainly from the generated samples themselves.

How Was Un-0 Trained?

Unconventional AI trained three model sizes each on CIFAR-10 and ImageNet 64x64, with the following results:

Training results on CIFAR-10

Training results on ImageNet 64x64

The results show that as the number of oscillators increases, the model’s FID score continues to improve. The largest ImageNet 64x64 model uses 16,384 oscillators, has about 322 million parameters, and reaches an FID of 6.74.

For training, the company used a newly proposed “Drifting Loss” function, along with a DINOv2 feature extractor and the AdamW optimizer, for end-to-end training.

For evaluation, CIFAR-10 used 50,000 generated samples, compared with CIFAR-10 reference statistics using standard packages and evaluation procedures. ImageNet 64x64 also used 50,000 generated samples, with FID calculated through the ADM evaluation suite.

In terms of computing power, all CIFAR-10 models were trained on one B200 GPU, while all ImageNet 64x64 models were trained on eight B200 GPUs. The largest CIFAR-10 model consumed 20 B200-hours of training, and the largest ImageNet 64x64 model consumed 640 B200-hours.

The company says the main training bottleneck comes from computing the “Drifting Loss” function, because it requires a conventional image feature extractor and calculations across multiple feature views.

Where Does Un-0 Stand in Image Generation?

To better show Un-0’s performance, Unconventional AI plotted it on a “generation quality versus parameter count” curve and compared it with conventional and unconventional models.

Relationship between parameter count and FID on the CIFAR-10 dataset

Relationship between parameter count and FID for 64x64 images

The conclusion: Un-0’s quality is already comparable to, and in some comparisons better than, several early conventional generators, including NCSN, DCGAN-TTUR, WGAN-GP, BigGAN, iDDPM, Consistency Models, and TRACT. But it still lags later high-performance conventional models such as EDM and GDD.

In other words, Un-0 is not the strongest image-generation model today. It is more like the starting point for a new path: its performance is already close to where several classic generative models stood when they were first proposed, but catching up with the latest frontier of the conventional track will require continued optimization across algorithms, architectures, and physical primitives.

Taken as a whole, Un-0 demonstrates the feasibility of using physical dynamical systems for large-scale modern AI image generation. Although its performance in software simulation has not yet reached the peak of conventional AI, it opens a promising path toward future “unconventional AI hardware” with a thousandfold improvement in energy efficiency.

Rao also stressed that the emergence of Un-0 shows that “computation is not an invention unique to humans.” It exists throughout nature and the physical world. The physical processes of all physical entities include a time dimension, but today’s computing systems do not truly make use of it.

“What we are developing is precisely this time dimension.”

The link to energy efficiency is that, in today’s von Neumann machines, most energy is spent moving information between memory and compute units. Dynamical systems merge computation and memory into the same entity. More importantly, dynamical systems can tolerate noise, further opening new opportunities to save communication energy.

Un-0 represents an important first step in shifting the computing paradigm toward dynamical systems. “With this model release, we are connecting intelligence and dynamics.” For AI computing, dynamics is a natural framework for representation. Neural networks themselves can essentially be viewed as dynamical systems, making the mapping between the two more direct.

“There is no abstraction called linear algebra in the brain, so in a sense, we are bypassing the middle layer.”

Many users in the comments also expressed anticipation.

“In fact, this kind of performance-efficiency improvement is enormous. If the technology can be widely adopted, many applications running locally could become feasible.”

“If this technology can make it to market, it would be an extremely advanced piece of brain technology.”

What do you think of this research? Join the discussion in the comments.

References

https://x.com/NaveenGRao/status/2070184079199494583

https://unconv.ai/blog/introducing-un-0-generating-images-with-coupled-oscillators/

https://techcrunch.com/2026/06/25/databricks-former-ai-chief-thinks-he-can-cut-ais-power-bill-by-1000x/

© THE END

Please contact this public account for authorization before reposting.

For pitches or coverage requests: liyazhou@jiqizhixin.com