Sunny

← All weeks

What I’ve been reading (W21, 2026)

The capex bet on AI, designing tools for the model, framing the real question, attention as the curriculum, and why depth still beats speed

  1. Everyone is investing in AI capacity, on the conviction that the risk of under-investing is greater than the risk of over-investing. US data centre construction has overtaken office construction, and the big four alone are guiding ~$700bn capex for 2026. Coding is the standout proof case so far, lots of training data plus scalable verification (you can just run the code and check) took coding cost to near 0, something that was previously near impossible to make cheap. https://www.ben-evans.com/presentations
  2. A nice checklist for designing MCPs. The core point is to do the boring work in the schema and tool descriptions, the model will happily do the clever bit if you stop making it guess. Pick a small set of core verbs with consistent naming, make discovery a tool that returns options as data not prose, embed next-step hints in every response, give every mutating tool a dry-run mode, and keep repeat calls safe because models will retry. Your tools should be obvious for the model, not the other way around. https://taoofmac.com/space/blog/2026/04/29/2341
  3. People rise to whatever level of abstraction the question opens up. “Wide vs deep”, “platform vs point”, “CAC vs LTV” all sound strategic but they let everyone in the room sound smart without doing the actual hard thinking. The truth is always one level down. Refuse the binary at altitude and ask the harder question instead, do you actually know what your customers need, can you conceive real differentiation in the face of stiff competition, and can you build that differentiation creatively (and quickly). Everything else is theater. https://x.com/shreyas/status/2055689118248763561
  4. Treasure your attention, use it to do hard things (people are antifragile, not fragile), and spend it on real people in real life. What you pay attention to shapes what you care about, and what you care about shapes who you become. https://archive.is/Y1JmD
  5. Sometimes speed is just impatience disguised as ambition. Good speed is strategic velocity, taking fast signals seriously and pivoting from stagnant ideas. Bad speed is novelty over depth, never sitting long enough in the unglamorous stretch where the real edge is formed. AI’s pace can make constant reorientation look like progress, but in a world where AI compresses execution, compounding matters more, not less. The bottleneck shifts to knowing what to build and why, and that judgment comes from years of accumulation, not a weekend with a new model. https://x.com/adityaag/status/2054976557702066520