Most AI systems today work by switching transistors on and off, billions of times per second, consuming vast amounts of energy. A startup founded by prominent AI researchers is proposing something fundamentally different: a chip that does not switch at all, but instead lets physical oscillators settle into synchronized patterns, and calls the result an image.
Unconventional AI, founded by Naveen Rao (formerly of Databricks, Nervana Systems, and Intel), Michael Carbin (MIT), Sara Achour (Stanford), and former Google engineer MeeLan Lee, released its first model in June. Called Un-0, it is an image generator built on a network of simulated coupled oscillators rather than conventional transistor logic. The company claims the architecture could be 1,000 times more energy efficient than conventional AI hardware.
The claim is extraordinary. But the physics behind it is not.
Synchronization as computation
The principle is drawn from the Kuramoto model, a mathematical description of how coupled oscillators behave. Two metronomes on a shared surface, ticking at slightly different rates, will gradually synchronize, eventually locking into the same rhythm. The same phenomenon appears in fireflies that flash in unison, pacemaker cells in the heart, and neurons in the brain.
In Un-0’s architecture, thousands of virtual oscillators are physically linked. Each oscillator has a phase, an angle describing where it is in its cycle. Different patterns of phase angles across the oscillator network encode different images. A “shoe” is represented by one phase configuration, a “train” by another.
To generate an image, a small subgroup of oscillators is preset to the target phase pattern, this is the “prompt.” When connected to a larger group of random-phase oscillators, the control group pulls the entire system toward synchronization with the target pattern. After convergence, the resulting grid of phase values is fed to a decoder that translates the numbers into pixel colors.
Where the efficiency comes from
In a conventional AI chip, every operation requires transistors to switch between on and off states, consuming energy each time. A single inference of a large model can involve trillions of such switches.
The oscillator-based approach replaces discrete switching with continuous physical dynamics. Current flows through closed loops that naturally oscillate. There is no forced switching, no energy wasted on holding transistors in position. The system reaches its answer through the physics of synchronization rather than through step-by-step logical computation.
“The system of coupled Kuramoto oscillators offers the promise of learning with physical dynamics at a scale that’s beyond what has been done before,” the company wrote in a blog post announcing Un-0.
What the benchmarks show
The current version of Un-0 runs on a software simulation of oscillators, not on physical hardware. The company’s stated goal is to build custom oscillator-based chips.
On standard image-generation benchmarks, Un-0 produced Fréchet Inception Distance (FID) scores comparable to early image-generation models such as Google’s BigGAN and OpenAI’s iDDPM. With 16,384 oscillators, it achieved an FID of 6.74 on ImageNet 64×64, a metric that measures how closely generated images match reference images, where lower is better.
These results do not match the current state of the art. “Conventional generators are still stronger on absolute quality and parameter efficiency, closing that gap with new algorithms and model architectures is the work ahead,” the company acknowledged.
The road ahead
Un-0 is a proof of concept, not a commercial product. The path from a software simulation of oscillators to a physical chip that realizes the 1,000x efficiency gain is long and uncertain. The company has raised $475 million in seed funding at a $4.5 billion valuation, backed by Lightspeed, Andreessen Horowitz, Sequoia, and Jeff Bezos, suggesting that investors are willing to wait.
The deeper significance of the work may be architectural rather than commercial. The standard assumption in AI hardware is that better chips mean faster transistors. Un-0 challenges that assumption by asking whether computation could look less like logic and more like physics, a network of coupled elements settling into a synchronized state that encodes the answer.