Artificial intelligence requires enormous amounts of computation and electricity, but startup Unconventional AI is exploring a radically different way to run AI models. Its experimental image generator, Un-0, uses networks of coupled oscillators instead of relying entirely on conventional transistor-based computation, tells Live Science.
Un-0 is a proof of concept for physical computing, in which the natural behavior of physical systems performs calculations. Conventional processors manipulate billions of transistors that rapidly switch between electrical states. Un-0 instead draws on oscillators, devices that generate continuous waveforms and naturally synchronize when connected.
The approach is based on the Kuramoto model, which describes synchronization among interconnected oscillators. Un-0 represents image information through different oscillator phases. A smaller group configured for a particular image category influences thousands of connected oscillators until the network settles into a pattern. The resulting phase values are then decoded into image pixels.
The potential energy savings are significant. Unconventional AI aims to make its technology 1,000 times more energy efficient than conventional computing. Rather than repeatedly switching transistors, future hardware could allow electrical current to flow through oscillator-based closed loops, potentially reducing the energy required for AI computation.
For now, however, Un-0 remains a simulation running on conventional hardware. The company ultimately plans to develop dedicated oscillator-based chips.
Tests on CIFAR-10 and ImageNet 64×64 showed that performance improved as the number of oscillators increased. On ImageNet, a system with 16,384 oscillators achieved a Fréchet inception distance score of 6.74. The results are comparable with early image-generation systems such as BigGAN and iDDPM, although current generative models remain more capable.
Unconventional AI has released the model weights and testing tools publicly, allowing researchers to explore whether physical dynamics could eventually provide a more energy-efficient foundation for AI computing.