Sunny

← All weeks

What I’ve been reading (W35, 2026)

Practical savviness and staying calm at work

Things that drift while nobody is watching: your own skill, your temper at work, a trading firm’s risk, and the computers agents run on + software cheap enough to reshape per user

  1. Savviness is the practical shrewdness you only get by doing things yourself and making mistakes. Every problem we hand to an LLM skips that struggle, so we ship faster while quietly becoming less capable, and the provider then trains its next model on the expertise we do still have. Worth choosing which problems to keep solving by hand pgaleone.eu/ai/2026/…/losing-savviness
  2. Being angry at work makes people work around you instead of with you, which cuts you out of decisions, which makes you angrier. Famous angry engineers like Linus Torvalds (the creator of Linux) are not a counterexample; they have the authority that stops people from working around them, and you don’t. The way out is either perspective from a rich life outside work, or genuinely caring about what the organization cares about seangoedecke.com/you-should-never-be-angry-at-work
  3. Most software is built for the average user, and anyone with a slightly unusual need is out of luck. Custom code used to be too expensive to fix that; with LLMs it is nearly free, so products can let each user extend the software by describing what they want, handing their custom code only the one or two functions it is allowed to call instead of broad access filtered afterwards jeremymorrell.dev/blog/extensible-software-in-the-age-of-llms
  4. Jane Street (the trading firm) built its record on holding positions for very short periods, and the longer your holding period, the more chance that something you can’t forecast occurs. As it grew it started holding bigger and longer bets, including backing Situational Awareness, the leveraged AI fund and in July the unforecastable arrived: a $15bn loss, its first losing month in a decade, even in a year where it has already made over $40bn, more than all of 2025 archive.is/Z9t2z
  5. The difference between a sandbox and a computer is what happens when you stop maintaining: a computer just keeps existing, a sandbox needs active work. For AI agents to do real work they need actual computers: files that survive, programs that keep running, an address on the internet, and a middleman that holds the passwords and keys so the agent never sees them. Today most platforms leave the key inside the environment and hope the model never prints it out. fly.io/computers-compared