Naveen Rao's Unconventional AI releases Un-0 to test physics-first image generation

Naveen Rao's Unconventional AI released Un-0, an open image-generation model built on simulated coupled oscillators, achieving FID 6.74 on ImageNet 64x64. The release is the first public test of the company's thesis that AI workloads can be mapped onto physical dynamics for far lower energy use, though it remains a software simulation without hardware or commercial deployment.

Naveen Rao @naveengrao https://x.com/naveengrao?ref=runtimewire and Unconventional AI https://unconv.ai/?ref=runtimewire released Un-0, an open image-generation model built around simulated coupled oscillators, giving the founder's new AI-hardware company its first public artifact beyond a financing story. In a June 25 model release https://unconv.ai/blog/introducing-un-0-generating-images-with-coupled-oscillators/?ref=runtimewire , Unconventional AI says Un-0 reaches FID 6.74 on class-conditional ImageNet 64x64 and is releasing model weights, training code, evaluation code, and ablation code. The accompanying GitHub repository https://github.com/unconv-ai/Un-0?ref=runtimewire describes the project as a plain-PyTorch reference implementation with separate CIFAR-10 and ImageNet-64 training pipelines, pretrained checkpoints, and a full training recipe. That matters because Rao is not pitching another image model company. He is trying to prove that modern AI work can be mapped onto physical dynamics, and eventually onto analog or oscillator-based hardware, with far lower energy use than conventional digital accelerators. Un-0 is a software simulation today. It is not evidence of a commercial chip, a deployed inference service, or measured 1000x energy savings on hardware. It is the first public test of whether the company can make its computing thesis behave like an AI model at all. Rao's third AI infrastructure bet Rao arrives with the kind of operating history that explains why investors funded a hardware-science company at a scale that would normally be reserved for later-stage revenue. Unconventional AI's author bio https://unconv.ai/blog/author/naveen-rao/?ref=runtimewire says he earned a PhD in neuroscience from Brown University and founded two AI companies that were acquired: Nervana Systems by Intel and MosaicML by Databricks. TechCrunch reported https://techcrunch.com/2025/12/09/unconventional-ai-confirms-its-massive-475m-seed-round/?ref=runtimewire that MosaicML sold to Databricks for $1.3 billion. The new company is a return to the same question Rao has been circling for more than a decade: what happens when AI progress becomes constrained not by algorithms alone, but by the machines that run them. Unconventional AI's December 8, 2025 introduction post https://unconv.ai/blog/introducing-unconventional-ai/?ref=runtimewire framed the company around an energy bottleneck, arguing that AI needs a more efficient computational substrate and that neural networks should run on the physics of silicon rather than on digital abstractions layered over analog circuits. The founding team reflects that full-stack ambition. MeeLan Lee @MeeLan Lee https://x.com/MeeLan Lee?ref=runtimewire , Unconventional AI's cofounder and VP of engineering, previously worked on analog and wireless chips, led high-volume WiFi chip work at Atheros, worked on chip design at Google, and helped put Google's Titan security chip into production, according to her company bio https://unconv.ai/blog/author/meelan-lee/?ref=runtimewire . Michael Carbin @mcarbin https://x.com/mcarbin?ref=runtimewire , a cofounder and senior fellow, is an MIT associate professor whose work spans AI systems, programming languages /article/racket-rhombus-1-0-release , and compilation for emerging systems, per his company bio https://unconv.ai/blog/author/michael-carbin/?ref=runtimewire . Sara Achour @sachour https://x.com/sachour ?ref=runtimewire , a cofounder and research fellow, is a Stanford assistant professor in computer science and electrical engineering whose company bio https://unconv.ai/blog/author/sara-achour/?ref=runtimewire describes work on software for emerging hardware technologies. That mix is the point. Unconventional AI is not trying to optimize Nvidia-style matrix math. It is trying to co-design the model, software, and eventual hardware around the dynamics of a physical system. What Un-0 actually does Un-0 uses a population of oscillators as the compute engine. The company's post describes the system in Kuramoto-oscillator terms: each oscillator has a phase, a natural frequency, and couplings to other oscillators. In generation, the model starts from random initial phases, injects class conditioning through a smaller group of oscillators, lets the dynamical system evolve, and then uses a conventional decoder to turn the final oscillator state into pixels. The largest ImageNet-64 model listed in the release uses 16,384 oscillators and around 322 million trainable parameters. That is the model the company reports at FID 6.74. The release also makes clear how much conventional infrastructure still sits underneath this early proof. Unconventional AI says it trained the ImageNet-64 models on 8 B200 GPUs, with the largest ImageNet-64 model using 640 B200-hours. In other words, Un-0 is not escaping the GPU stack at training time. It is using that stack to search for a model form that might later map to physical hardware. That distinction keeps the release from being overread. Unconventional AI's claim is not that it has displaced diffusion models or commercial image systems. In the release, the company situates Un-0 alongside earlier conventional generators and short of the latest frontier, and notes that quality improves with scale more slowly than conventional baselines. That is a sober placement: a credible research artifact, not a state-of-the-art product launch. The ablation is the story The most useful part of the release is not the sample mosaic. It is the ablation work designed to answer whether the oscillator dynamics are doing anything beyond handing random features to a decoder. Unconventional AI tests decoder-only models and variants that use frozen random reservoirs versus learned dynamics. The company reports that trained dynamics bring clear benefits, indicating the oscillators are contributing useful computation beyond the decoder. That is exactly the claim the company needed to make testable. The ablations do not prove the eventual hardware thesis, but they narrow the question. The next test is whether those dynamics can be implemented in silicon or another physical substrate without losing the behavior that made the benchmark work in simulation. A $475 million seed round now has code attached Unconventional AI announced on December 8, 2025 that it had raised $475 million in seed funding at a $4.5 billion valuation in its introduction post https://unconv.ai/blog/introducing-unconventional-ai/?ref=runtimewire . The company said the round was led by Lightspeed https://lsvp.com/?ref=runtimewire and Andreessen Horowitz https://a16z.com/?ref=runtimewire , with participation from Sequoia https://www.sequoiacap.com/?ref=runtimewire , Lux Capital https://www.luxcapital.com/?ref=runtimewire , DCVC https://www.dcvc.com/?ref=runtimewire , Future Ventures https://future.ventures/?ref=runtimewire , Jeff Bezos, and other investors. It also said Rao would personally invest $10 million. TechCrunch https://techcrunch.com/2025/12/09/unconventional-ai-confirms-its-massive-475m-seed-round/?ref=runtimewire later reported that the $475 million was a first installment toward a possible round of up to $1 billion, a framing also noted by Data Center Dynamics https://www.datacenterdynamics.com/en/news/neuromorphic-compute-startup-unconventional-ai-raises-475m-in-seed-funding/?ref=runtimewire . Those reports should not be read as a completed $1 billion raise. The verified company announcement is the $475 million seed round at a $4.5 billion valuation. That financing bought Unconventional AI time to attempt something most startups cannot afford: a new compute substrate, not merely a new model architecture. It also put pressure on Rao's team to show something concrete early. Un-0 is that first concrete thing. It is open code, a reproducible training recipe, and a benchmark result that outside researchers can inspect. TechCrunch reported on June 25 https://techcrunch.com/2026/06/25/databricks-former-ai-chief-thinks-he-can-cut-ais-power-bill-by-1000x/?ref=runtimewire that the current Un-0 runs in software simulation and that the company plans to release schematics for an actual chip, eventually supplying compute capacity. Rao framed the release as a hello world for a new kind of computer. The hard part comes next The competitive set for Unconventional AI is broader than image-generation labs. Nvidia, AMD, Google, Cerebras, Groq, and a long list of custom-silicon companies are all pursuing more efficient AI compute by improving digital accelerators or systems around them. The closer technical neighborhood includes neuromorphic, analog, optical, thermodynamic, and other physical-computing approaches, many of which have produced striking demos but struggled to become mainstream AI infrastructure. Unconventional AI's opening move is to avoid starting with a chip demo alone. It starts with a model and a benchmark, then works backward toward the hardware. That sequencing is founder logic as much as technical logic. Rao's career has been about the boundary between model software and compute systems, and Un-0 is built to make the boundary visible. The unanswered question is whether the mapping survives contact with real hardware. Software simulation lets researchers train and evaluate the idea under controlled conditions. Physical oscillators bring noise, manufacturing variation, coupling constraints, I/O overhead, and the economics of building systems that customers can actually use. Until Unconventional AI publishes hardware measurements, the 1000x figure remains the company's goal, not a demonstrated result. Still, the release changes the burden of proof. Before Un-0, Unconventional AI was a heavily funded thesis about energy and physics. After Un-0, it is a thesis with code, weights, benchmarks, and ablations. That is the right first artifact for a company trying to convince the market that the next AI computer may not look like the last one.