Two builder-operators break down the last two weeks in AI using 10 Tyler Cowen–style questions. We get into Sora 2’s cameo culture, whether “thinking” models are worth the latency, agents that actually help, model choice for client work, the energy/compute wave, and why open-weights like DeepSeek matter (or don’t) for practitioners.

What you’ll get

Practical takes from people shipping client projects

Where Claude 4.5 vs GPT shines (coding vs writing)

When to use “extended thinking/deep research” vs fast models

Real talk on agents, meeting schedulers, and workflow design

Energy, nuclear, and why AI ≈ infrastructure

Open-weights vs ecosystems: where the moat really is

Chapters
00:00 – Cold open & intro
00:26 – Who’s Tyler Cowen and why this format
02:00 – Q1: Sora 2, IP, and the “cameo economy”
06:44 – What we’re doing in this episode (format explainer)
08:11 – Q2: GPT apps & the VibeCoder value prop (workflow architect vs app builder)
17:03 – Q3: “30-hour agents” & autonomy myths (Claude, Replit Agent)
22:57 – Q4: When to use thinking models vs fast models (and deep research)
29:04 – Q5: SB-53 AI transparency—useful or compliance theater?
30:24 – Q6: Picking models for clients: capability, brand, or last best output?
37:01 – Q7: Agents that actually help (Lindy scheduling, weekly pain points)
41:34 – Q8: Compute, energy, and nuclear—should builders be optimistic?
46:50 – Q9: DeepSeek R1 costs & the real moat (ecosystems > raw perf)
49:33 – Wrap-up & feedback ask

Links & mentions (non-sponsored)

Tyler Cowen / Marginal Revolution

Anthropic Claude 4.5 (coding + writing)

OpenAI GPT-5 (auto/fast tasks), Deep Research modes

Lindy meeting agent

Replit Agent 3 (autonomous build experiments)