- In the AI world, the default software moat over the last few decades, switching costs, is likely not sustainable. This may actually help you replace incumbents, in that it may be easier to migrate customers over, but it hurts your ability to remain sticky and defensibility shifts toward cornered resources, typically proprietary data, real network effects when your line of business supports them, and compounding process power. Being extremely AI-native is a good place to start. https://www.tanayj.com/p/moats-in-the-age-of-ai
- Apple doesn’t do budget. But it has gotten very good at doing value, and this week it set a new floor: Apple has decided that $599 USD is the new floor for a real Apple device. What matters this week isn’t any single product, but the coherent way Apple is moving to capture profits across the entire phone, tablet, and laptop market. It is refreshing its entire accessible tier at the same time, with phone, tablet, and laptop all at the same price point. https://om.co/2026/03/02/apple-does-value-week https://www.macrumors.com/2026/03/04/apple-unveiled-seven-products-this-week/
- When you delegate work to an AI agent, four things should be true before you share the output: it actually works and you’ve verified it does, the scope is small enough for someone to check quickly, there’s enough context on the goal for others to understand it, and the summary itself has been reviewed properly instead of just sounding polished. https://simonwillison.net/guides/agentic-engineering-patterns/anti-patterns/
- Good software knows the purpose it serves, does not try to do everything, and knows when to stop, what to improve, and what is really just another project or another tool. https://ogirardot.writizzy.com/p/good-software-knows-when-to-stop
- LLMs understand a codebase better without executing it, but you need to make them show their work: explore the repo (so it sees surrounding context, not just the snippet), state explicit premises, trace the real execution path (including imports, shadowed functions, and test-relevant branches), and prove the conclusion from code evidence instead of pattern-matching. https://arxiv.org/pdf/2603.01896
- The responsibility is the reward. https://world.hey.com/dhh/the-responsibility-is-the-reward-69e5b73f
What I’ve been reading (W10, 2026)
AI moats, Apple's value play, agentic delegation, software restraint, and teaching LLMs to show their work