For a decade, executing deep neural networks on GPUs has defined how artificial intelligence is built. A startup called Unconventional AI is making the case that the next leap in efficiency will require a fundamentally different kind of computer, one where the laws of physics do the computing directly.
On June 25, 2026, the company released Un-0, an image generator powered by a simulated system of coupled oscillators. Unconventional AI describes the project as a demonstration of an emerging "physical computing substrate,"" and it is releasing the model weights, training code, and ablation analysis openly. For more context on this story, see our ongoing artificial intelligence updates.
The release quickly drew attention in the developer community, reaching the front page of Hacker News and sparking discussion about whether alternatives to the GPU-and-transformer status quo can realistically compete.
Matching conventional image generation at its launch quality
Un-0 reaches FID 6.74 on class-conditional ImageNet at 64x64 resolution. FID, or Fréchet Inception Distance, is a standard measure of image-generation quality where lower is better. Unconventional AI says that score matches the quality of leading conventional image generation methods at the time they were first published.
The company is candid that there is still a gap to today's frontier. Un-0 does not yet close the performance gap as a function of parameter count relative to state-of-the-art diffusion models. But the team frames the result as a proof of concept that modern AI workloads can be mapped onto the dynamics of a physical system.
"To our knowledge, Un-0 is the most capable image generator to date to use a simulation of a physical dynamical system," the company wrote in its announcement.
How a population of oscillators draws a picture
Un-0's compute engine is a large population of oscillators, each coupled to the others, where the coupling strengths between every pair of oscillators are the primary learnable parameters of the model. The system is modeled using what are known as Kuramoto oscillators, in which each oscillator rotates at its own natural frequency while being nudged by the pull of every other oscillator.
The company offers an intuitive analogy. Picture two metronomes ticking side by side on a shared table. Depending on how strongly they influence one another, they fall into lockstep or settle into opposition. Scale that from two oscillators to thousands, and the population self-organizes into patterns.
Generating an image with Un-0 follows five steps. The system starts from randomness, setting every oscillator's phase to a random angle, which acts as the seed equivalent to the noise a diffusion model or GAN samples. A second, smaller group of oscillators encodes the requested image class, such as a daisy or a volcano, and is coupled into the main population. The system is then released, letting the oscillators pull on one another until they settle toward a state dictated by their learned coupling. Images are read out at a fixed time and passed through a decoder.
The 1,000x energy-efficiency thesis
The deeper motivation is energy. Unconventional AI's stated goal is to run modern AI on roughly 1,000 times less energy than today's machines require.
The argument is that an oscillator is a primitive physical circuit. A coupled-oscillator system can be implemented directly in CMOS or other physical substrates so that the physics of the device itself carries out the computation, rather than simulating it on a conventional processor. If the laws of physics can execute AI workloads, the company argues, the execution substrate can look radically different from today's GPU clusters.
The company points to a long lineage of physics-based computing ideas, including neuromorphic computing, Hopfield networks, reservoir computing, Hamiltonian and Liquid networks, Neural Wave Machines, thermodynamic computing, and prior work on Kuramoto oscillators. It also draws a biological parallel: rhythmic activity and synchronization are pervasive in the brain and have long been hypothesized to perform computational work, such as binding distributed features into coherent perceptions.
Early days, open code, and an invitation
Unconventional AI is clear that this is a beginning rather than a finished product. The physical primitive Un-0 explores is not new, but the company has scaled it to a larger generative benchmark, performed an ablated analysis of the dynamics, and offered an interpretive look at the model's behavior.
By releasing the weights and code, the team hopes to make it easier for others to experiment with models grounded in physical dynamics. The company says it believes researchers can quickly push beyond Un-0, and that the field is still early in its effort to reseat modern AI on physical dynamics and reach the roughly 1,000x energy-efficiency gains it is chasing.
For now, Un-0 is a research artifact rather than a drop-in replacement for production image models. But it arrives at a moment when the energy and cost of running large AI systems has become a central concern for enterprises and researchers alike, and it represents one of the more creative attempts to question whether the GPU-centric paradigm is the only way forward.
Source: "Introducing Un-0: Generating Images with Coupled Oscillators," Unconventional AI, June 25, 2026.---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
