In Episode 21, Tom Spencer and Cameron Rohn break down the current state of AI — from market hype to hardcore engineering practice.
Topics include:
- Michael Burry’s short on Nvidia and Palantir
- Is there really an AI bubble — or just a new kind of economy?
- Breaking Nvidia’s CUDA lock-in with modular AI
- Google’s “Nested Learning” and Anthropic’s interleaved thinking
- Building AI copilots and MCP servers
- LangSmith experiments, evaluators, and continuous optimization
- Microsoft’s Copilot Studio and enterprise automation
- What real AI engineering looks like in production
🎧 Subscribe for weekly deep dives into AI products, agent frameworks, and research.
00:00 – Airport stories, brisket, and warm-up banter
03:00 – MCP servers and Polygon data experiments
05:00 – Minimax and Anthropic’s interleaved thinking
07:00 – Google’s “Nested Learning” paper and continual optimization
08:30 – NeurIPS, AI research culture, and the VC invasion
09:30 – Is there an AI bubble? Michael Burry’s Nvidia short
11:00 – Palantir, Nvidia, and the tech bubble debate
14:00 – CapEx growth and the “AI money loop”
17:00 – Are AI companies actually profitable?
19:00 – Free users, monetization, and ChatGPT’s economics
20:30 – The real differences from the dot-com era
22:00 – Nvidia’s margins, chip efficiency, and modular AI challengers
25:00 – Breaking CUDA lock-in and the rise of hardware portability
27:00 – Local inference, hybrid models, and agentic operating systems
33:00 – Chrome OS, MCP in browsers, and local AI
34:00 – Anthropic Excel plugin and Kimi Thinking model benchmarks
37:00 – MCP server demos and architecture discussion
43:00 – Building an AI options trading copilot
46:00 – Visualizing strategies, composable components, and LangGraph
50:00 – How MCP connects data and trading logic
55:00 – Skill systems, consistency, and reproducibility in LLM apps
58:00 – LangChain documentation and developer experience
1:00:00 – Combining MCP data for richer insights
1:03:00 – Converting trading logic into agentic workflows
1:06:00 – Building autonomous trading systems on LangGraph
1:08:00 – LangSmith experiments, datasets, and evaluators
1:13:00 – Backtesting AI outputs and customer feedback optimization
1:20:00 – Comparing models and evaluators in LangSmith
1:24:00 – Microsoft Copilot Studio and Power Automate for enterprise AI
1:29:00 – Wrapping up: AI compliance, tooling, and what’s next
