Building The Chip That Could Unlock AGI | Naveen Rao

This episode dives into the hidden economics of GPU scarcity—how AI and crypto are colliding over compute power, and what this means for investors.

The Genesis of Unconventional AI and the Vision for a New Computing Paradigm

Naveen Rao, a seasoned expert in AI hardware and software, introduces Unconventional AI, his latest venture. He clarifies it's not merely a "chip company" but a deep dive into the first principles of how learning functions within physical systems. Rao, known for co-founding successful companies like Nirvana (AI chip accelerators) and Mosaic (a cloud computing company), and serving as Head of AI at Databricks, is now tackling the fundamental architecture of computing. His motivation stems from a profound belief that current digital computing paradigms, largely unchanged for 80 years, are insufficient for AI's future.

Bridging Hardware and Software: A Full-Stack Philosophy

Naveen Rao's career trajectory, spanning early hardware development for wireless technology and video compression to a PhD in neuroscience, has cultivated a unique "full-stack" perspective. He defines an "OG full-stack engineer" as someone proficient across silicon design, computer architecture, low-level software, and applications, contrasting it with the modern definition focused on JavaScript and Python. This holistic understanding allows him to view hardware and software not as distinct boundaries but as fluid elements, enabling him to right-size solutions to problems. His diverse experience underpins Unconventional AI's approach to rethinking computing from the ground up.

Reimagining Computing: From Digital to Analog for AI

Rao explains his passion for building a fundamentally new computer, driven by the brain's exquisite efficiency—consuming only 20 watts while performing complex computations. He highlights that digital computers, which represent numbers with fixed bits and precision errors, became dominant due to their scalability, exemplified by ENIAC (Electronic Numerical Integrator and Computer), built in 1945 with 18,000 vacuum tubes. However, analog computers, which were among the first, were inherently more efficient by using physical systems to model quantities, but struggled with manufacturing variability and scaling. Unconventional AI aims to leverage analog principles for AI, where intelligence is seen as a stochastic and distributed process, not naturally suited for highly precise, deterministic digital substrates.

The Efficiency Imperative: Why Analog Computing is Resurfacing for AI

Digital computers simulate anything expressible as numbers and arithmetic, making them general-purpose machines, but they are not inherently efficient for all tasks. Analog computers, conversely, use the physics of the underlying medium for computation, making them intrinsically more efficient. Rao cites wind tunnels as an example of an analog computer, modeling fluid dynamics more accurately than computational simulations. For AI, Unconventional AI seeks to build silicon circuits that recapitulate neural network behaviors, arguing that intelligence, being a stochastic machine, is better suited for an analogous physical system than a deterministic digital one. This approach could unlock significant energy savings and performance gains for AI workloads.

Intelligence as Physics: Emulating Brain Dynamics for Efficiency

Naveen Rao posits that in biological brains, intelligence is not an abstraction but the direct manifestation of physical neural network dynamics, mediated by chemical diffusion and the physical properties of neurons. This direct physical implementation, without layers of abstraction like operating systems or APIs, is what makes brains exceptionally efficient, even at a fraction of a watt for smaller creatures. Unconventional AI aims to build systems where intelligence is the physics, creating an isomorphism in electrical circuits that can subserve intelligence. This approach seeks to harness the inherent efficiency of physical processes to build AI systems that are orders of magnitude more power-efficient than current digital architectures.

AI's Energy Footprint: The Urgent Need for Efficient Compute

The escalating energy demands of AI present a critical global challenge. Naveen Rao highlights that US data centers, representing 50% of the world's capacity, already consume 4% of the US energy grid. Projections indicate a need for an additional 400 gigawatts of capacity over the next decade to meet AI demand, far exceeding current power generation capabilities (approximately 4 GW per year). This energy crisis underscores the urgent need for a fundamental shift in computing architecture. For Crypto AI investors and researchers, this implies that energy-efficient hardware is not just an optimization but a necessity for the sustainable scaling of decentralized AI, blockchain infrastructure, and AI-powered crypto applications, directly impacting operational costs and environmental footprint.

Tailoring Compute: Analog's Niche in Dynamic and Stochastic AI Workloads

Rao emphasizes that the future isn't a binary choice between digital and analog but a strategic integration. Analog approaches are particularly well-suited for workloads that can be expressed as dynamical systems, where time and causality are inherent. Unlike numeric computing, which simulates time, analog systems can directly leverage physical dynamics. This is ideal for "fuzzier" intelligence tasks like retrieving and summarizing information, where brains excel at integrating diverse, noisy inputs to produce highly accurate behaviors, as seen in elite athletes. For Crypto AI, this suggests that analog systems could enhance AI models for complex, real-time decision-making in dynamic environments like DeFi, where precise yet adaptive responses to constantly changing market conditions are crucial.

Towards AGI: Causality and Dynamic Systems in Next-Gen AI Models

Unconventional AI is initially targeting state-of-the-art models like transformers and diffusion models, noting that diffusion and flow models inherently incorporate dynamics, often expressed as ordinary differential equations (ODEs). Rao believes that building AI out of dynamic elements that inherently understand time and causality will be a superior foundation for achieving Artificial General Intelligence (AGI). He argues that current AI, despite its utility, lacks a true sense of causality, leading to "stupid errors." By integrating causality through dynamic systems, Unconventional AI aims to develop machines that exhibit a more profound understanding of the world, moving closer to human-like intelligence. For Crypto AI researchers, this implies a potential for more robust, less error-prone AI agents capable of understanding complex, multi-step interactions within blockchain ecosystems.

Navigating the AI Hardware Landscape: Partnerships and Competition

Naveen Rao outlines Unconventional AI's strategic position within the broader industry, aiming to find an analogous paradigm for intelligence within five years and achieve scalable manufacturing. He views TSMC (Taiwan Semiconductor Manufacturing Company) as a crucial partner for prototyping and high-volume production. While Google (with its TPUs), Nvidia, and Microsoft are at the forefront of AI applications, Rao sees them as potential collaborators rather than just competitors. Unconventional AI's goal is to build a "better substrate than matrix multiply," the core operation for current GPUs. This suggests a long-term vision for a new foundational compute layer that could complement or even redefine existing AI hardware ecosystems.

The Visionary Drive: AI as Humanity's Next Evolution

Naveen Rao's personal motivation is deeply rooted in the transformative potential of AI. He rejects the "AI doomer" narrative, viewing AI as the "next evolution of humanity" that will enable deeper collaboration and understanding of the world. His passion for hardware, particularly the "dopamine hit" of seeing a new chip power on, fuels his ambitious pursuit. Rao believes that achieving AI ubiquity requires a fundamental change in computing, as the current paradigm, despite its advancements, cannot scale to that level. This vision underscores a commitment to long-term, foundational innovation that aims to leave a lasting impact on history.

Cultivating Innovation: The Unconventional AI Team and Culture

Unconventional AI operates as a "practical research lab," fostering an open-ended culture that prioritizes exploring what's possible before considering manufacturing constraints. Rao seeks a diverse team, including traditional AI systems engineers skilled in mapping algorithms to physical substrates, theorists focused on dynamical systems and neural networks, system architects bridging theory and buildability, and analog/digital circuit designers. He emphasizes hiring individuals with "high agency" who are passionate about tackling hard problems and are given the autonomy to experiment and learn from both successes and failures. This approach aims to attract top talent eager to push the boundaries of computing.

The episode highlights the critical need for energy-efficient AI compute, driven by an impending global energy crisis. Crypto AI investors and researchers should track analog computing advancements closely, as this paradigm shift could drastically reduce operational costs for decentralized AI and unlock new, causality-aware AI applications within blockchain ecosystems.