Introducing our first model, Un-0! We trained an image generator powered by a backbone of coupled oscillators in place of a more traditional conventional neural network.

At the @theallinpod Summit, our cofounder and CEO @NaveenGRao shared new details on our first test chip tape out: "This is the first physical dynamical computer ever built. We did it in five months." @chamath joins him to discuss the path to product. piped.video/yAsrMA_ADPc

I made some public disclosures about our hardware at the @theallinpod conf this week and wanted to share the progress here. Earlier this summer we got hardware back! The execution was madness…we went from no team in Jan to a tape out in 5 months. AI enabled MUCH tighter loops of research and our execution speed shows the results. This is the first large-scale demonstration of a causal, physical dynamical system to do real compute. 🚀 We released the Un-0 model (see link) in June and it runs on this physical system; the images below come from the actual hardware. What’s more, this chip requires <900 nJ to generate an image; this is many orders of magnitude less energy than conventional machines. unconv.ai/blog/introducing-u…

Unconventional AI was founded on a suspicion: that more parameters and more attention is one path among several, and that the alternatives stay underexplored not because they are less promising but because they are harder to fund. The Unconventional Grant exists to fund those alternatives. We said plainly that we would rather back a well-posed risk than a safe increment. Today we are announcing our first cohort. Five projects, each receiving $100,000: - Thierry Tambe (Stanford) is treating data retention as an architected resource, so processors stop spending area and energy preserving information long after the computation stops needing it. - William Gilpin (UT Austin) is testing whether the slowdown in reasoning models is transient chaos, the same phenomenon that governs turbulent flows, and looking for the fractal traps in latent space that delay convergence. - Logan Wright (Yale) is asking what neural scaling laws become when the neural network is a physical analog machine rather than a digital simulation of one, noise, drift, and energy budgets included. - Yoon Kim (MIT) is pushing looped Transformers past the tens of loops where gains have appeared to saturate, into the hundreds and thousands, with primitives designed to keep parameters on-chip. - Wanyu Lei, Daniel Kunin, and Friedrich T. Sommer (UC Berkeley) are deriving recurrent architectures directly from the mathematics of state updates, on the premise that efficient state tracking is not one capability among many but a foundational design principle. To the five teams: congratulations. Each team describes its own project, in its own words, in the full announcement. unconv.ai/blog/2026-unconven…

We removed 93.25% of the connections in our Un-0 image model, fully expecting to pay for it in quality. But it got better. FID 7.15 on ImageNet 64x64, roughly 1.9 ahead of the dense baseline at matched size. Same family of model, a fraction of the couplings, a better score. Here is why that is not as strange as it sounds. Un-0 is a coupled oscillator model, and in the dense version every oscillator talks to every other one. That sounds like a strength, but it means the whole system can fall into catastrophic synchronization: everything locks into step, gradients go flat, and learning stalls. Sparser connectivity leaves room for coherent and incoherent activity to coexist. The dynamics stay alive, and the model keeps learning. Connectivity turns out to be a control knob, not a dial you turn up until it stops. Learn more here: unconv.ai/blog/less-is-more-…

Building a new memory means a twenty-year grind to drive the error rate toward zero. AI just made that grind optional. Because neural networks are remarkably tolerant of noise, we no longer have to chase punishing reliability targets. Training can settle into a good solution even when the hardware underneath is imperfect (which is exactly what the loss landscape below is showing). That single shift lets us stop optimizing for zero errors and start optimizing for what matters now: more bits per cell, higher density, lower energy. The universal memory never showed up, but AI may be the workload that tells each technology exactly which job it was born to do. Read the full breakdown by Giacomo Pedretti (MTS, AI Hardware Architectures) and Srenik Mehta (VP of Engineering): unconv.ai/blog/smoothing-the…

What if your AI chip didn't calculate the answer, but physically settled into it? That's the premise behind dynamical system hardware, and there's never been a standard way to program it. In an upcoming International Symposium on Computer Architecture (@ISCAConfOrg) 2026 paper, Unconventional AI MTS Prof. Ang Li and co-authors introduce DS-ISA: a minimalist 9-instruction architecture that bridges digital processors and continuous-time analog physics. It's a proof-of-concept, but it's the abstraction layer the field needs to build real compilers and software stacks for dynamics-based AI hardware. Full deep dive: unconv.ai/blog/speaking-the-…