Episode #84: From World Models to Robot Orchestras: Inside the New Stack of Real-Time Intelligence

00:00 — CPU history and why mainframes never had a central processing unit
05:00 — Jensen Huang's five-layer cake and the slowdown in LLM training data
10:00 — Ring zero, operating systems, and the shift from mainframes to personal computers
15:00 — ARM architecture, Apple's chip transition, and the Wintel breakup
20:00 — RISC-V as an open-source ISA and China's play for chip sovereignty
25:00 — TSMC vs SMIC, the node gap, and Intel's foundry ambitions
30:00 — Real-time inference vs batch LLM training and what that means for AI
35:00 — Stewart Jr.'s coding agent setup and the chaos of managing planning agents in parallel
40:00 — Hallucinations, probabilistic vs deterministic systems, and staying in the loop
45:00 — Competitive landscape of LLMs and the race toward general world models
48:00 — Fei-Fei Li's World Labs, Waymo's driver model, and the robot orchestra idea in Buenos Aires

1. The CPU was never part of mainframe architecture — it was a concept born with the personal computer. Once Intel and Motorola introduced the first chips, everything from operating systems to software stacks got built outward from that core, and that architecture eventually swallowed the mainframe world entirely.
2. ARM's low-power RISC design wasn't engineered for mobile — it was just cheaper and more efficient. That accidental advantage locked Intel out of the smartphone race entirely, and now ARM's licensed architecture sits inside nearly every mobile chip on the planet.
3. RISC-V's real revolution was legal, not technical. By releasing an open-source ISA, Berkeley gave China a path to chip independence that doesn't require licensing from Western companies — turning an academic project into a geopolitical weapon.
4. TSMC's manufacturing lead is structural, not just numerical. SMIC is roughly three generations behind, and because TSMC keeps advancing, the gap doesn't close — it compounds. China can design chips but still can't build the most advanced ones at scale.
5. The shift from LLMs to world models is fundamentally about time. LLMs are batch processes with a months-long lag between training and deployment. World models operate in real time, which is what robots, autonomous vehicles, and physical AI actually require.
6. Real-time inference is the new battleground. Jensen Huang's move into CPUs signals that the most important compute is no longer about building the model — it's about reasoning fast enough to react to the physical world as it happens.
7. Stewart Jr.'s multi-agent setup reveals something important: even with powerful AI, humans still need to own the architecture. The agents hallucinate, gaslight, and lose context — so the orchestration layer, the judgment about where to look and what to trust, still has to be a person.