Bezos-Backed AI Tackles Datacenter Power Challenges

A New Approach to Sustainable AI

Startup founder Naveen Rao has a bold vision for the future of artificial intelligence. He believes that the key to creating more sustainable and efficient AI systems lies in looking to nature, specifically the brain. Rao, who previously founded AI companies that were acquired by Intel and Databricks, is now leading a new venture called Unconventional AI.

Unconventional AI recently announced a significant milestone: it secured $475 million in seed funding from major investors including Jeff Bezos, Andreessen Horowitz, and Lightspeed. This investment is aimed at tackling one of the biggest challenges in AI development — the energy consumption problem.

Rao explained that AI is deeply connected to hardware, and hardware is closely tied to power. "We can't scale beyond a certain number of inferences per unit time because of the energy problem," he said. "We can't produce that much more energy in the next 10 years."

According to Rao, current AI systems are using the wrong tools for the job. He pointed out that natural learning systems, such as the human brain, don't rely on numerical simulations. Instead, they use the intrinsic physics of their substrates to build learning systems. "We believe we can recapitulate that behavior in silicon," he said.

This concept isn't entirely new. Companies like IBM and Intel have been exploring similar ideas for years. The human brain, which operates on just 20 watts of bioelectric energy, offers a model for what could be achieved with more powerful computing systems. However, despite decades of research, neuromorphic computing — which aims to reverse-engineer the brain's architecture — has made limited progress.

"Slow progress doesn’t mean this approach is wrong," Rao noted. He referenced the history of neural networks, which were once considered a backwater field until compute power became more abundant. "Some of these things don't work until they do."

Unconventional AI isn't solely focused on neuromorphic computing. Rao emphasized that while the brain's design may offer useful insights, there's no need to replicate it exactly. "There's probably concepts from the brain that are useful in building such a [learning] system. That's the way we look at it. It's not that it must work like the brain," he said.

Instead, the company is exploring various approaches to improve the efficiency of machine learning accelerators. While details about the research remain confidential, Rao mentioned that the chips will likely be analog rather than digital. "These are nonlinear dynamics of circuits. That's inherently an analog thing," he explained.

Digital systems provide determinism, which is essential for tasks like accounting software where consistent results are crucial. However, machine learning often involves nondeterministic processes. Rao envisions a future where a combination of non-deterministic analog and deterministic digital logic could accelerate different aspects of machine learning workloads.

Certain models, such as diffusion models, flow models, and energy-based models, are particularly well-suited for the kinds of non-linear dynamics that Unconventional AI is targeting. Rao acknowledged that solving these challenges will take time. "We're not going to have a product in two years," he said. "This is largely a research effort for the next several years."

Despite the long-term focus, Rao plans to share findings along the way. "Over the next several months, we're going to start releasing things," he said. While the initial phase is centered on research, his ultimate goal is to build a systems company that can deliver transformative solutions.