World View

Where this note starts: a stub. My operating beliefs, to be filled in as
I catch myself applying them — the goal is a page I can point people (and future me) at, grown from evidence in the garden rather than written in one sitting.

Work

• Learning by doing beats learning by reading — but thinking before doing is what separates doing from thrashing.
• Automation pays twice: once in the time saved, once in the errors never made.
• KISS — lean scripts, minimal moving parts; a tool you can hold in your head is a tool you can trust.
• Verify, don’t guess — check live state before configuring; answers grounded in sources with file:line refs, never vibes.

Tools & technology

• Use all the tools at your disposal — including AI. What matters is the outcome and whether you understand what the tool did.
• Open by default: share the process, work with the garage door up;
  knowledge compounds when it’s public.
• Data deserves presentation — an insight that isn’t well presented is an insight that doesn’t land.

Expertise

The roon doctrine (@tszzl, 2022, 11.2k❤, captured): “experts are fake, smart generalists rule the world, everything is designed by people no smarter than you, and courage is in shorter supply than genius.”

I agree with most of this — with one carve-out, verbatim from the take that anchored it here: experts are fake is generally true apart from true masters of their craft with deep knowledge and real experience. The distinction that holds both:

• Fake experts are credential- or confidence-shaped: authority without accountability to outcomes, consensus-repeaters, middle-of-the-Dunning-Krayer curve. Against them, the smart generalist wins — and “courage is in shorter supply than genius” is the operative half of the doctrine.
• True masters are legible by their artifacts and error rates, not their titles — the What matters in AI world taste lineage (Huntley’s “gives a lot of shits,” Rubin’s finishers), Titus Winters’ wisdom-over-intelligence. They are rare precisely because the fake kind inflates the base rate; that rarity is what makes the generalist heuristic generally right and the carve-out load-bearing when it applies.

The test this leaves me with: don’t ask “is this person an expert” — ask “what have they built/fixed/broken that I can inspect, and do they know the failure modes of their own craft.” Generalist courage by default; masters exist and are worth finding.

The corollary: asking too many people kills good ideas (2026-10-18)

USER-added at capture (mymind, image vision-transcribed): “THE FASTEST WAY TO KILL A GOOD IDEA IS ASKING TOO MANY PEOPLE WHAT THEY THINK ABOUT IT.”

Fits the Expertise section as its corollary: the roon doctrine says courage is scarcer than genius — this is why ideas die anyway. Poll enough people and the fake-expert base rate dominates the sample: each lukewarm take costs more than it teaches, consensus converges on the median opinion, and the idea gets averaged to death before it’s built. Consistent with the carve-out: a small number of true masters (legible artifacts, known failure modes) is worth asking; a crowd is not — the crowd’s answer is the population mean by construction (AI brain rot’s distribution argument, applied to feedback instead of output). Operational rule: before asking anyone, decide who counts — if the list is longer than the people who’d have to live with the consequences, it’s polling, not review.

How to grow this note

• When a decision, TIL, or bit reveals a belief underneath it, link it here and add the principle (one line, testable).
• Contradictions are welcome: date them and keep both — a worldview that never updates is a bug.

Related: How this garden works (the workflow this view produces),
BITS (evidence of the “rarely used but confusing” belief).

Tools & wielded AI (the counter-thesis)

yyhh.org’s reply to Rich Hickey’s “‘AI’ is Not a Tool” (2026-10-13, captured) lands on the position this note’s Tools section already implies, said better: the useful distinction is between wielding an instrument with skill and delegating work without adequate understanding or oversight — “and that gives us something more practical than declaring that an entire class of software cannot be a tool.”

The arguments worth keeping:

• “Picking up a saw does not guarantee a straight cut. Holding a chisel does not make you a sculptor.” Tools amplify poor judgment as readily as good; usefulness depends on the operator’s ability to direct, interpret feedback, recognize mistakes, correct. Predictability and bounded function are qualities we evaluate tools by, not prerequisites for membership in the category.
• The compiler/spreadsheet counter: automation can be a feature of a tool — “why should generating text automatically disqualify an instrument when translating code or calculating numbers automatically does not?”
• The LLM feedback loop is real when the human runs it: ask an LLM to explain unfamiliar code, challenge it, inspect the source, test the change, reject, refine — “that is a feedback loop directed by a human pursuing a purpose. The possibility of avoiding intellectual work does not establish that every use avoids it.”
• “AI is a mirror” (the author’s colleague; independently ghuntley: “LLMs are essentially mirrors… they mirror the skill of the operator,” captured): it reflects the curiosity, judgment, and habits brought to it — the questions asked and the standards applied to answers are the operator’s.
• The vocabulary warning: repeating “go rogue / escape containment / conspire” accepts anthropomorphic framing; unexpected outputs and instruction failures “are real engineering problems. Calling them ‘conspiracy’ adds human motives without establishing any.”

Kin: What matters in AI world (the operator-side traits this wielding requires — taste, review, ownership), Software engineering in AI era (the debate map’s continuity pole), Dipping my toes into clojure (Hickey lane).