Interview, Conference Presentation
The Chip That Could Unlock AGI.
- Company & Leadership: Naveen Rao, co-founder of Mosaic and former head of AI at Databricks (where he also co-founded Nirvana, sold to Intel), is the CEO of Unconventional AI, a hardware startup founded at NeurIPS.
- Core Mission: Unconventional AI is not a traditional chip company; it focuses on "first principles" to discover physical systems that replicate the efficiency of biological learning, aiming to break the 80-year-old digital computing paradigm.
- Energy Constraints: The U.S. data centers currently consume 4% of the national energy grid; projections suggest this could rise to 8-10%, necessitating new capacity to prevent grid instability.
- Capacity Gap: Estimates indicate a need for an additional 400 gigawatts of power generation capacity over the next decade to meet AI demand, a significant shortfall given current infrastructure constraints.
- Architectural Shift: Rao argues against the binary "digital vs. analog" debate, proposing that analog substrates are superior for stochastic, dynamic systems (like intelligence) while digital systems remain necessary for deterministic numerical tasks.
- Physical Computation: Unlike digital computers which simulate physics via bits and error estimation, Unconventional aims to use the actual physics of electrical circuits to perform computation, eliminating the "lossy" abstractions of OS and API layers found in brains.
- Target Workloads: Initial hardware targets will support energy-based models, diffusion models, and flow models, which are defined by ordinary differential equations and inherently possess time-dynamics compatible with physical substrates.
- AGI Trajectory: Rao believes that incorporating time, causality, and dynamic evolution into hardware substrates is a necessary step toward achieving Artificial General Intelligence (AGI), which he views as the next evolution of humanity.
- Industry Relationships:
- TSMC: Identified as a confirmed partner for manufacturing scalability and rapid prototyping.
- Google: Described as having internal hardware capabilities with a focus on incremental improvements (TPUs) rather than radical paradigm shifts.
- NVIDIA: Position is ambiguous; they built the current standard programming platform, but Unconventional aims to build a better substrate than matrix multiplication, leaving room for future collaboration or competition.
- Talent Strategy: The company seeks "crazy" people with high agency to bridge theory and engineering, specifically recruiting talent in analog/digital circuit design, dynamical systems theory, and energy-based models.
- Development Approach: The roadmap prioritizes "existence proofs" and open-ended research for the first few years, deferring manufacturing trade-offs until physical validation is achieved, with a first prototype planned to be one of the largest analog chips ever built.
- Career Philosophy: Rao advises early-career individuals to pursue diverse full-stack experience (hardware, software, applications) rather than deep specialization in a single domain, as broad competency fosters the ability to drive future technological change.
- Motivation: Rao views the project as a generational opportunity to make AI ubiquitous and efficient, arguing that the current digital paradigm cannot scale to a future where AI is integrated into daily life without massive infrastructure overhaul.