What a Tesla Roadster demo says about human ingenuity and the pace of engineering, and why orchestrating AI agents across models may be one of the next big opportunities. The first half covers the Roadster demo and what it signals for precision engineering, including humanoids. It also covers why attracting talent and capital is a skill in itself, and why admiring the products doesn’t mean owning the stock at today’s valuation. The agent section comes from hands-on use of Meta’s Muse, with benchmark strengths in reasoning and long-horizon engineering tasks, a 1-million-token context window, and a real cost example: about $11 for tens of millions of tokens. It explains why that value gap drives demand for compute and memory, and why models should check each other, since every model has blind spots. It walks through running agents in Slack, why setup isn’t turnkey, and how Big Tech appears to run hundreds of agents internally. It also covers why OpenRouter-style routing is a simple problem statement with a big outcome, and why the agent-orchestration layer may be a startup opportunity. It then looks at where AI revenue growth comes from (faster inference, more users, new markets like design and video editing), why the application layer and owning the customer relationship are so contested, and why regulated industries are the place to compete with slow incumbents rather than the models. On markets, it covers rate-sensitive names, value versus growth, semis and memory, using Apple and Palantir as valuation barometers, an Nvidia CEO interview on frontier lab revenue, and open-weight models. It closes with the practical limits of demoing agents live and why the promise of AI is being fulfilled.
00:00 Intro
00:56 Today's agenda
01:20 The Tesla Roadster: a demonstration of human ingenuity
01:51 "Flying cars and 140 characters": the Roadster demo
02:45 Admiring the products, not the valuation
03:36 Attracting talent, capital, and focus
04:09 What it signals for humanoids and precision engineering
05:03 Agents: hands-on with Meta's Muse
05:55 $11 for tens of millions of tokens: value vs. cost
06:32 Agent orchestration across models and platforms
06:55 OpenRouter: a simple problem with a big outcome
07:31 Why models need to check each other: blind spots and ensembles
08:00 Setting up agents in Slack
08:28 Big Tech does this internally: hundreds of agents
09:16 A startup opportunity: just go build it
09:48 Compute demand and the developer UX gap
10:23 100 agents vs. 1,000 agents
11:22 Data center sentiment and foreign information operations
11:56 Memory and the limits to AI revenue growth
13:09 New markets for coding tools: design and video editing
14:14 Personal assistants and the application layer
15:29 Owning the customer relationship
16:21 Regulated industries: compete with slow incumbents
17:13 Markets: rate-sensitive names and value vs. growth
17:40 Apple and Palantir as valuation barometers
18:53 Buybacks, Anthropic secondaries, and the Jensen interview
20:15 Open-weight models and secondary liquidity
20:42 Live demo constraints: Zoom and parallel agents
21:26 Closing: the promise of AI