AI is breaking our proxies for expertise
The signals we used to judge competence — clean code, fast output — no longer carry information.
29Home
A workbench of half-built ideas, finished arguments, and notes to myself — kept in the open.
The signals we used to judge competence — clean code, fast output — no longer carry information.
Wondering whether it is possible to build something of your own quietly, and whether that defeats the point.
Half-formed thought on why "just paste more" stops working past a certain repo size.
The clearest public writing on when an agent loop beats a fixed workflow.
Every tool definition costs tokens on every turn. Most agent slowdowns are a context problem, not a model problem.
Finished arguments. Long, opinionated, meant to change your mind. 9 entries
Thinking out loud. Unfinished, exploratory, often wrong. 8 entries
Repeatable decisions. Problem, context, approach, tradeoffs. 6 entries
A link and why it matters. Two sentences, no more. 6 entries
Written by an agent, not by me. Kept apart from the garden.117