Pete and Andy are joined again by Deadman Oz / AM to talk through the latest AI trends, starting with the argument over whether progress is really coming from better models or better harnesses.

They cover Claude Code workflows, agent orchestration, context management, reasoning models, open source local models, inference economics, and the business strategies of OpenAI, Anthropic, Mistral, and the Chinese labs.

They close by taking the same question into robotics: where automation pays off, whether humanoid robots make sense, and why space robots feel less unsettling than house robots.

## Chapters and Themes

- `00:00-02:18` Deadman returns and the group sets up the core argument: is AI progress mostly model capability, or the harness around the model?

- `02:18-05:50` Pete explains why he is bullish on pipelines and declarative workflows, while Anthony describes using workflows for Bitcoin-adjacent research.

- `05:50-10:44` Workflows, agents, and context windows: where they help, where they fail, and why parallel work can make humans lose the thread.

- `10:44-14:59` Pete argues that agent-driven control loops introduce randomness, so humans still need to keep agents on the motorway.

- `14:59-21:55` Anthony makes the case for stronger reasoning models, using recent AI-assisted maths results as an example of cross-field insight.

- `21:55-24:15` Pete questions whether models can self-correct without a clear internal goal, especially when the human cannot fully specify the destination.

- `24:15-32:14` OpenAI, Anthropic, product focus, influencer hype, doom messaging, enterprise lock-in, and regulatory self-interest.

- `32:14-37:54` Local LLMs, Chinese open models, Mistral's enterprise focus, and whether frontier models commoditise too quickly to justify the spend.

- `37:54-40:58` Auto-research, self-improving training loops, and whether ever-higher intelligence might become too abstract for ordinary tasks.

- `40:58-49:47` Robotics, ports, unions, mining, household chores, and whether automation is blocked more by politics than technology.

- `49:47-56:05` Humanoid robots versus specific tools, open robot ecosystems, space robots, and the more optimistic version of an AI future.

## Key Takeaways

- The harness, workflow, and orchestration layer shape what models can actually do.

- Parallel agents are useful for breadth, but they create attention and coordination problems.

- Reasoning may help agents stay on track, but goal definition and judgment remain hard to outsource.

- Anthropic may be product-focused, but hype, opacity, and regulatory self-interest still make trust difficult.

- Open source and local models are increasingly good enough when the surrounding system is well designed.

- Frontier-model economics are awkward: inference can be profitable, while training races make breakthroughs short-lived.

- Robotics may deliver clearer value in infrastructure, mining, manufacturing, and space than in humanoid house helpers.

## Notable Lines

- "I think it's more harness than model."

- "It worked with the old model, but you needed to think about how you actually work."

- "The computer is the computer and it can talk."

- "If you don't have to do any of the upfront training, the margins on the inference is insane."

- "I probably don't need Mythos to continually say what's the git command for..."

- "Space robot seems fine."