Naveen Rao on Unconventional AI: Analog Computing for AGI

The Vision Behind Unconventional AI

• 💡 Naveen Rao, co-founder and CEO of Unconventional AI, is building an AI chip startup focused on analog computing systems.
• 🚀 The company aims to rethink computer architecture by exploring first principles of how learning works in physical systems, moving beyond 80 years of digital computing.
• 🧠 Rao's background includes leading AI at Databricks and founding Nervana (AI accelerators) and Mosaic (cloud computing), bringing a full-stack perspective to hardware and software.

Analog vs. Digital Computing

• 📊 Digital computers represent numbers with fixed bits, offering precision and generality for simulating anything expressible as numbers and arithmetic.
• ⚡ Analog computers use the physics of the underlying medium to perform computation, making them inherently more efficient but historically difficult to scale due to manufacturing variability.
• 💡 Intelligence, particularly neural networks, is seen as a stochastic and distributed process, making it a better fit for the dynamic, "fuzzier" nature of analog systems than highly precise digital ones.

Addressing the Energy Crisis in AI

• ⚠️ The US data centers consume 4% of the national energy grid, with demand for AI requiring an additional 400 gigawatts over the next decade, highlighting a critical energy shortfall.
• 🌍 The current computing paradigm is energy-bound at a global level, necessitating a fundamental rethinking of how AI computations are performed.
• ⚡ Analog approaches, especially for dynamical systems with inherent time and causality, offer a path to significantly more efficient AI.

Path to AGI and Causality

• 🧠 Naveen Rao believes that building AI out of elements with time evolution and causality will lead to systems closer to Artificial General Intelligence (AGI).
• 💡 Unlike current AI models that can make "stupid errors," a dynamic, physics-based approach could impart a more innate understanding of causation, similar to how human brains operate.
• 🎯 Unconventional AI's hardware is well-suited for transformers, diffusion models, and flow models, particularly those expressed as ordinary differential equations, by mapping their dynamics to physical systems.

Building the Future of AI

• 🤝 Unconventional AI seeks to collaborate with partners like TSMC for manufacturing scalability, while potentially offering a better substrate than matrix multiply compared to current leaders like Nvidia and Google.
• 🚀 The company is a "practical research lab" focused on solving hard problems, combining 40+ years of academic research with engineering to build scalable, efficient AI solutions.
• 🌟 Naveen's motivation stems from the opportunity to change the world of computing and make AI ubiquitous, believing it represents the next evolution of humanity.