Naveen Rao, former AI chief of Databricks, is leading a new company called Unconventional AI in an ambitious bid to reduce the energy consumption of AI systems by up to 1,000 times. The key to their approach lies in a novel oscillator-based architecture that could transform traditional computing.

Their first model, named Un0, demonstrates this technology’s potential through image generation processes similar to those used by existing models like Stable Diffusion and GPT Image 1. However, the real innovation is under the hood, where the company believes they can achieve massive energy savings.

Rao envisions a future where AI computations are performed using chips that run at just one-thousandth of the power required today. This could revolutionize not only data centres but also consumer devices and cloud services, making them far more sustainable.

While Un0 currently runs on software simulations, the company is planning to release chip designs soon. They aim to build an entire inference stack around these chips, potentially turning AI into a much greener technology.

The project's scale is daunting given its current small team of less than 50 employees, but it could be crucial as the demand for AI computations continues to grow exponentially. If successful, this might just help us navigate the looming energy crisis in the age of AI.