Un-0: Coupled Oscillators Replace the Neural Net

@unconvAI just dropped something genuinely different: Un-0, an image generator whose backbone is a system of coupled Kuramoto oscillators instead of conventional neural network layers.

They trained it as a generative model where random initial phases evolve under learned coupling strengths. The dynamics of these "swings" (as @radamar nicely put it) do the heavy lifting. They scaled it to 16k oscillators / 322M parameters and reached an FID of ~6.74 on ImageNet 64×64 — to their knowledge, the strongest result so far for a model built on a simulation of a physical system.

Key claims & ablations (from their thread):

• The physics is doing real work: trained dynamics beat a decoder-only baseline and random oscillator reservoir.
• More integration steps during inference improves quality.
• They're open-sourcing weights, training scripts, and ablation code so the community can experiment with physics-based backbones.

#A few thoughts

This sits at the intersection of reservoir computing, analog/AI hardware, and unconventional computing. Instead of approximating everything with matrix multiplies, they're leaning into the natural dynamics of coupled oscillators — a direction that could eventually lead to dramatically lower energy use if moved to physical analog substrates.

@radamar's commentary captured the vibe well: "Forget diffusion, you only need a bunch of 'swings'." It's a refreshing swing (pun intended) at the dominant paradigm. Whether this scales to frontier performance or stays a niche efficiency play remains to be seen, but expanding the Pareto frontier for small generative models is valuable in itself.

The broader bet here is exciting: biology-scale efficiency may require moving beyond digital floating-point networks toward substrates that compute with physics natively. Un-0 is an early, tangible step down that path.

If you're into alternative computing paradigms, this thread (and their blog) is worth a deep dive.

Related reading: for another oscillator-based take on neural computation, see the Winfree Oscillatory Neural Network (WONN) — which builds learning dynamics on Winfree-model oscillators.