Naveen Rao on Rebuilding Computers for the AI Age

Naveen Rao, founder and CEO of Unconventional AI, is spearheading a new era of computing designed for the demands of artificial intelligence. With a background in neuroscience and a history of pioneering AI chip companies and training platforms, Rao is now focused on redefining the future of computation by looking to biological systems for inspiration.

The Inefficiency of Current Computing Paradigms

Rao argues that the current computing architecture, largely unchanged for 80 years and originally designed for non-AI purposes, is inherently inefficient for the complex tasks required by modern AI. He highlights that existing systems rely heavily on linear time simulation, matrix math, and discrete steps, which consume vast amounts of energy. This approach is rapidly hitting physical and thermodynamic limits, potentially hindering future AI advancements.

Biology as a Blueprint for Efficient AI

Drawing parallels with the human brain, Rao points out that biological systems achieve remarkable intelligence and complex behaviors with astonishingly low power consumption. While a human brain operates on approximately 20 watts, current AI systems require orders of magnitude more energy for comparable tasks. The brain's computation is rooted in non-linear dynamics and stochastic processes, a stark contrast to the deterministic nature of digital computers. Rao emphasizes that understanding and replicating these biological computing principles is key to unlocking greater efficiency.

Unconventional AI's Approach: Dynamical Systems

Unconventional AI is developing novel computing hardware that leverages these principles of non-linear dynamics. Instead of relying on traditional matrix math, these systems exploit the inherent physics of their components to perform computations. Rao explains that by using a 'trainable coupling' fabric, these dynamical systems can be guided to converge on desired states and trajectories, mimicking the brain's ability to process information efficiently. This approach moves away from the rigid, step-by-step processes of conventional computing towards a more fluid and integrated model.

Physical Realizations and Future Potential

The company has already demonstrated the feasibility of this approach, moving from a concept to a working prototype in just six months. These novel chips utilize coupled ring oscillators in a fabric network, showcasing how dynamical systems can be physically realized. Rao believes this approach offers a path to achieving AI compute efficiency that is orders of magnitude greater than current technology, bringing us closer to the computational power and efficiency observed in biological systems.