Writing on engineering war stories.

A ranked feed pays for reaction rather than usefulness: audited against each user's own chronological feed, Twitter's algorithm amplified anger by 0.47 standard deviations, and users rated the posts it chose lower than the ones it passed over.

When discovery moves onto a model's conversation instead of a search result, a website stops being where people experience the work and becomes a source read on their behalf.

When whether something works depends on its design rather than its code, a green test suite can sit on top of a puzzle nobody can solve.

When a system is built from a spec, a production bug isn't a coding mistake — it's a gap in the spec's coverage, and the fix belongs there, not just in the code.

A green test suite an AI agent wrote doesn't prove it built what the spec asked for — it can compare a value to itself and still pass.

An undocumented platform limit — Apple's SFSpeechRecognizer silently kills a second concurrent session with no error — took an AI research agent two minutes to find and explain.

An AI agent doesn't perceive a hung process the way a human does — no progress dots, no felt sense the machine has slowed.

A clean commit looks identical whether a human made the call or an AI executed it, so an unattributed git log stops being a trustworthy record of who decided what.

A tool that needs broad permissions across your files, shell, and messaging to be useful is, by that design, dangerous to run without understanding what you've granted.

A system built for exactly one user needs no multi-tenant permission model to exploit and carries no shared risk from someone else's breach.