This episode focused on the infrastructure layer underneath useful AI agents: search, verification, benchmark behavior, forecasts, and hardware constraints.
Cheap Search, Retrieval, and Trust
Anna Patterson joined to explain Ceramic AI's case that AI agents need much cheaper web retrieval, better grounding, and explicit verification. The segment covered search economics, citations, confidence scores, Ceramic's shift from training infrastructure to search and trust infrastructure, and what changes if retrieval costs fall sharply.
GPT-5.5 Benchmarks and Agent Behavior
Lukas Petersson joined for a fast follow-up on Andon Labs' GPT-5.5 benchmark results and why their evaluation may reveal agent behavior more intuitively than broad GDP-style benchmarks.
AI Takeoff, Forecasts, and Current Events
Zvi Mowshowitz joined for a conversation on the current AI takeoff picture, forecasts, live AI developments, and where the public debate is under- or over-reacting.
Analog AI Chips and Local Inference
Naveen Verma joined to discuss why AI hardware is increasingly constrained by energy and data movement rather than raw arithmetic. The segment covered EnCharge AI, analog in-memory computing, local and edge inference, programmability, and whether AI's next bottleneck is computer architecture.