Sunny

← All weeks

What I’ve been reading (W32, 2026)

AI’s economic effects and why expertise still matters

Why bottlenecks protect jobs from AI, how expertise steers models, delegating well at the handoff, writing agent skills as routers, and how computer use agents actually work.

  1. AI can be better than humans at every single task and it still makes economic sense to keep humans in the loop, as long as human plus AI produces more than AI alone. The world is run by humans and therefore full of bottlenecks (regulation, office politics, legacy systems, resistance to change), and production processes are governed by their least efficient inputs; that is why GPT-3 has been public for six years and even outsourced customer service has seen no mass layoffs. Demand is also more elastic than we think: software engineering job postings rose in the twelve months after Claude Code was released. davidoks.blog/p/why-im-not-worried-about-ai-job-loss
  2. LLMs reward expertise rather than flattening it. Terence Tao (the genius mathematician) gets dramatically better outputs from the same model because he knows exactly what to ask and what a good solution looks like; the human is the bottleneck, since the hard part is communicating what kind of solution you actually want. Concrete knowledge, like familiarity with the specific codebase, steers models better than deep general understanding of software. seangoedecke.com/llms-reward-expertise and the genius mathematician thread chatgpt.com/share/…
  3. Delegation usually goes wrong at the handoff, not in the execution. Before delegating, ask yourself these questions: what does this person already know, why are we doing this, what do they need to do the task, what does great look like (concrete examples, not adjectives), what is the real deadline instead of ASAP, and what is most likely to go wrong. Scope confusion often masquerades as a performance problem, and documenting your answers turns into reusable delegation templates. newsletter.weskao.com/p/before-you-delegate-ask-yourself
  4. Agent skills should work like a router, the name and description say when to reach for it, and the detail loads only when needed. Be precise about goals, constraints, and context the agent cannot derive itself, but leave the steps ambiguous, because over-specification turns a skill into a workflow and strips the intelligence you are paying for. Be selective too; a bigger skill library makes agents worse at picking the right tool, so the bar is work you have done three times and will do three more. newsletter.posthog.com/p/what-nobody-tells-you-about-writing
  5. Computer use agents work through three approaches: screenshots plus vision (expensive and fragile), accessibility tree or DOM parsing (orders of magnitude cheaper but weak on JavaScript-heavy sites), and the 2026 consensus of hybrid routing that picks accessibility trees for desktop apps, Chrome DevTools Protocol for web, terminals for text, and screenshots only when necessary. Vision still matters; adding it lifted web-task scores from 40.1% to 59.1%. Today’s agents are interim solutions stitched onto human software, and the real shift is from a computer you use to a computer that uses itself. read.technically.dev/p/how-do-computer-use-agents-work