CHAPTERS
- 0:00 – 1:23
Why AI is the next evolution—and why Naveen Rao is betting on new compute paradigms
The episode opens with a high-level framing: AI as a step-change for humanity, and Naveen Rao as a long-time hardware/software builder now leading Unconventional AI. Matt sets the stage with Naveen’s track record and the central question: why take on another, even harder, bet now?
- •AI positioned as a transformative force for human collaboration and understanding
- •Naveen’s credibility anchored in Nervana, Mosaic, and leadership roles in AI
- •Unconventional introduced as an AI-hardware startup with an unconventional thesis
- •Core prompt teased: why start a new chip company now?
- 1:23 – 2:13
From early specialized hardware to neuroscience: a career built around efficiency
Naveen explains how his early work focused on shrinking algorithms into efficient hardware to enable real-time use cases like wireless and video compression. A later pivot to a neuroscience PhD reinforced his interest in how physical systems compute efficiently.
- •Early career: custom hardware to make compute-heavy workloads feasible in real time
- •Hardware as an enabler for new form factors and widespread adoption
- •Neuroscience training shapes his intuition about learning in physical systems
- •Efficiency as a throughline across domains
- 2:13 – 3:22
‘Full stack’ before it meant JavaScript: why hardware/software is an artificial boundary
Matt probes why Naveen built a software-heavy company after selling a chip company. Naveen argues hardware vs. software is a chosen abstraction line; real innovation comes from placing that line where it best fits the problem and the user.
- •Old-school ‘full stack’ spans silicon, architecture, low-level software, and applications
- •Hardware/software separation is a pragmatic choice, not a fundamental boundary
- •Good system design is about ‘rightsizing’ the solution to the consumption point
- •Cross-layer thinking enables new architectures
- 3:22 – 5:09
Why ‘not a chip company’: starting from first principles of learning in physical systems
Naveen reframes Unconventional as a theory-first effort: understanding how learning could be implemented directly in physical substrates. He argues we’ve built essentially the same digital computer for ~80 years, while biology demonstrates radically better energy efficiency and adaptability.
- •Unconventional starts with theory and physical first principles, not just chip design
- •Digital computing paradigm has remained largely consistent since the 1940s
- •Brains highlight a massive efficiency gap (power use, dynamism, adaptability)
- •Industry incentives matter—now the incentives align because energy is the bottleneck
- 5:09 – 8:38
Digital vs. analog computing: precision, scalability, and why analog disappeared
Naveen contrasts digital computation (numbers represented by fixed bits, general-purpose simulation) with analog computation (using physics to directly embody quantities). Analog was early and efficient, but digital won because it scaled better despite device variability—an echo of today’s GPU-scaling story.
- •Digital: numeric representation with quantization/precision trade-offs, broadly general
- •Analog: computation by physical analogy—letting physics do the ‘math’
- •Analog’s historical limitation: manufacturing variability made scaling difficult
- •Digital abstraction enabled systems like ENIAC; scaling challenges mirror modern AI training clusters
- 8:38 – 11:26
Why intelligence may fit analog/mixed-signal: stochastic brains vs. deterministic machines
The conversation turns to why analog-style substrates might better match intelligence workloads. Naveen argues neural networks are stochastic and distributed, yet we run them on deterministic, precision-optimized digital stacks; Unconventional seeks an electrical-circuit ‘isomorphism’ that better matches intelligence.
- •Intelligence differs from classical numeric tasks requiring determinism and exactness
- •Brains implement neural dynamics physically—no OS/API abstraction layer
- •Modern AI models are stochastic; digital precision may be mismatched overhead
- •Goal: find circuit-level mappings that naturally support intelligent computation
- 11:26 – 12:53
The macro constraint: AI’s energy demand and the looming infrastructure shortfall
Naveen and Matt quantify the energy and grid pressures created by AI data centers. The argument: even aggressive generation buildouts may be too slow, and transmission constraints compound the problem—making efficiency-driven compute redesign urgent.
- •Data centers consume a meaningful and growing share of grid capacity
- •Early signs (e.g., brownouts) indicate stress as AI demand accelerates
- •Estimates suggest hundreds of gigawatts of new capacity needed over a decade
- •Efficiency improvements in compute could be a faster lever than infrastructure buildout alone
- 12:53 – 15:23
Tradeoffs and the ‘intelligence substrate’: where analog helps and where digital stays
Naveen rejects a binary ‘digital vs. analog’ framing, arguing different workloads benefit from different substrates. Systems with dynamics (time evolution) may map better to analog approaches, while traditional numeric computation remains important; Unconventional targets the fuzzy, integrative intelligence regime.
- •Hybrid view: match substrate to workload characteristics
- •Dynamical systems and time-based processes are natural fits for analog computation
- •Brains integrate many variable inputs yet achieve high precision in real-world conditions
- •Concept of an ‘intelligence substrate’ distinct from classic computation
- 15:23 – 16:52
Which models and modalities benefit: diffusion, flow, and energy-based dynamics
Naveen explains that Unconventional won’t discard today’s best models but will start from them—especially those explicitly defined by dynamics (ODEs). He suggests transformers succeeded partly because they map well to GPUs, but may be parameter-inefficient relative to other representations better suited to physical dynamics.
- •Start from practical SOTA: transformers and diffusion models
- •Focus interest: diffusion/flow/energy-based models due to explicit dynamics (ODE forms)
- •Hypothesis: map model dynamics onto physical system dynamics for efficiency gains
- •Transformers may be more a GPU-friendly construct than a ‘natural law’ of intelligence
- 16:52 – 19:09
Does this move toward AGI? The role of time, dynamics, and causality
Pressed on AGI, Naveen offers a candid, intuitive argument: systems grounded in time evolution may develop stronger causal understanding than static formulations. He notes current models are useful but still make obvious mistakes and don’t feel like collaborating with a person.
- •AGI discussion acknowledged as inherently speculative and ‘hand-wavy’
- •Claim: dynamics with time/causality may be a better foundation for intelligence
- •Causality framed as a key missing ingredient in today’s model behavior
- •Current systems are powerful tools but not human-like collaborators
- 19:09 – 21:00
Industry landscape: partners, platforms, and manufacturing scale as a hard requirement
Naveen places Unconventional among major incumbents, emphasizing manufacturing scalability as non-negotiable for solving the global energy constraint. He expects foundry partnership (TSMC) and sees a spectrum of relationships with Google/TPUs and NVIDIA/CUDA—from parallel efforts to potential collaboration—while aiming beyond matrix multiply as the core substrate.
- •Five-year goal: demonstrate the paradigm, then ensure manufacturable scalability
- •TSMC positioned as a likely key partner for prototyping and scale
- •Google seen as optimizing lower-risk internal hardware iterations (TPUs)
- •NVIDIA relationship uncertain; Unconventional aims for a substrate ‘better than matrix multiply’
- 21:00 – 23:07
Motivation: the unique dopamine of hardware—and a pro-ubiquity, anti-doomer AI worldview
Naveen describes why he keeps returning to hard hardware problems: turning on a new chip is uniquely thrilling. He frames Unconventional as enabling AI ubiquity—arguing that the current compute paradigm won’t get us there and that AI’s benefits far outweigh its downsides.
- •Hardware development provides rare, visceral ‘it works!’ moments
- •Naveen identifies as ‘the opposite of an AI doomer’
- •Belief: AI can deepen human understanding of each other and the world
- •Thesis: AI ubiquity requires changing the underlying computer paradigm
- 23:07 – 24:43
Why he thinks it can work: brains as proof, decades of research, and engineering integration
Asked about confidence, Naveen cites three pillars: biological existence proof, decades of academic prototypes, and growing theoretical understanding in dynamical systems and neuroscience. The challenge is engineering: combining partial advances across the stack into a manufacturable system, accepting iterative “sand it down and make it fit” reality.
- •Existence proof: biological brains demonstrate feasibility of efficient intelligence
- •40+ years of academic work suggests promising device and system concepts
- •Theory progress: neuroscience and dynamical systems provide design intuition
- •Engineering task: integrate disparate advances into a scalable product
- 24:43 – 30:08
Building the team and culture: a practical research lab, mixed-signal reality, and high-agency careers
Naveen outlines hiring needs across theory, algorithms-to-hardware mapping, architecture, and analog/digital circuit design—highlighting that pushing analog chips to unprecedented scale is inherently risky. He advises younger engineers that startups build breadth and adaptability, and he emphasizes a culture of exploration-first (existence proofs) with high agency and ownership.
- •Hiring across the stack: theory, AI systems mapping, architecture, analog + digital circuits
- •Expectations: first large-scale analog prototypes may behave unexpectedly
- •Career advice: early startup experience builds breadth that pays off amid change
- •Culture: seek existence proofs first; maximize agency so people can try, own, and learn fast
