Field notes on applied AI
Notes on agentic workflows, product design when software can run the work, and what still requires human judgment. Subscribe via RSS.
- AI safety without locking in the leadersPacing frontier AI could buy time for safety while entrenching the labs setting the pace. The rules need to survive cheaper, more widely available capabilities.
- If AI saves time, why does the workday keep expanding?AI makes more goals feel reachable, so ambitious people spend the saved time before they have it. Why that makes the workday heavier, and why it's America's edge in AI.
- 13 illustration styles, revisited: GPT Image 2 vs. Nano Banana 226 images, 13 styles, one shared scene: a side-by-side comparison of GPT Image 2 and Nano Banana 2, with notes on style, detail, and composition.
- The first hour and the next few monthsAI can build an app faster than you can understand it. The long part is learning what you have, so you know what to ask for next.
- Design AI workflows like access can breakAI access is a business dependency, and dependencies fail. Give every workflow a reduced mode that keeps running without the frontier model.
- The uneven days of building with agentsSome days agents make building feel light; others you end up lost in side alleys. What separates the two, and what still matters when code is cheap.
- What changes when an agent can do the boring part?A proposal for turning each failed loan transfer into a tested fix, and how I'd decide whether it's worth building.
- 11 illustration styles that work well with AI image modelsThe same scene recreated in 11 styles across GPT Image and Gemini, with the exact prompt blocks I used and notes on what held up.
- How Git worktrees workGit worktrees give each coding agent its own folder and branch without duplicating the repository.