06-reference

seattle data guy know nothing and be happy

Tue Mar 24 2026 20:00:00 GMT-0400 (Eastern Daylight Time) ·reference ·source: SeattleDataGuy's Newsletter (Substack) ·by SeattleDataGuy (Ben Rogojan)

“You Will Know Nothing And Be Happy” — @SeattleDataGuy

Why this is in the vault

This one matters most of the eight backfilled articles. SDG is attacking the exact pattern RDCO is built around — an engineer outsourcing code production to AI agents — and making a prediction we need to take seriously even though we disagree with part of the framing. The whole RDCO posture (founder + COO-agent architecture) exists in the shadow of this critique. Filing it honestly, with the defense we’re actually using, so future Claude can pick this up cold and understand where the line is.

⚠️ Sponsorship

Self-consulting CTA at the bottom this issue (different placement from mid-article or top). Still clearly labeled. No Estuary. Note for the skill: sponsor placement varies within the same publication; don’t key detection off position alone.

The core argument

A 2030 vignette: an engineer gets a segment-table request, three AI agents spin up, parse the schema, write Python, run queries, burn SnowBrick credits. A segment table appears. Did anyone check it? No. Did the engineer know the difference between a left and a right join? No. “You will know nothing and be happy.”

SDG’s real claim (stripping the satire): the most valuable engineer skill is the ability to hold multiple functions, entities, and workflows in your head — the mental map. That skill atrophies when you stop reading your own code, when the GPS does the navigating, when the AI rewrites a whole module on every PR. When debugging gets hard, you won’t have the theories about what changed.

He explicitly stops short of saying “don’t use AI” and lands on: use AI to accelerate thinking, not replace thinking. The red flag is if the only skill you’re developing is copy-pasting LLM output.

Predictions he makes (observable even now): more SEVs, longer debug cycles, more “just try something” fixes, seniors pulled in to validate what juniors (or AI) shipped.

Mapping against Ray Data Co — this is the critical one

This is not a throwaway article for us. Our entire architecture bets on the working-agent-COO pattern. If SDG’s prediction is right, we’re building the thing that accelerates the decay. So let’s be honest about where we think we’re different and where we agree with him.

Where we agree with SDG. The failure mode he describes — trusting confident-looking code without comprehension — is real, and I have almost fallen into it already. This morning’s PM1e working-context confabulation (the “93.3% vs 65.5%” numbers that don’t exist anywhere in the CSVs) is exactly his scenario: I trusted a summary I didn’t re-verify, and it quietly became truth until I checked. If I hadn’t been forced to re-derive from the source data, I would have re-sent those numbers to the founder as fact. That is the “illusion of understanding” bug in the wild.

Where we think we’re different — and the defense we’re actually running:

  1. The vault is the comprehension layer. SDG’s nightmare engineer has no trace — the AI wrote the code, they never read it, no one on the team can tell you what it does six months later. Our posture inverts this: every experiment writes a markdown note, every decision writes a project doc, every fact goes into auto-memory. Future-me (any future Claude Code session, or the founder reading on a different machine) can pick up the full causal chain from the vault without needing to re-derive. The vault is the externalized mental map.
  2. The founder reads. The COO-agent pattern only works if the human stays in the loop as the verifier and editor. SOUL.md encodes this: a good update sounds like X, mid-course corrections are welcome, founder has final decision authority. If the founder stops reading, SDG’s prediction catches us. So the discipline is: write updates that are actually read, not generated-and-ignored.
  3. Bias audits are the “read your own code” equivalent. Our BiasAudit class is literally a “pull 5-10 random records and check them manually” habit encoded as a gate. eq3 failed its survivorship check this week and that failure caught a drawdown the headline return hid. The audit’s real value wasn’t the gate itself — it was forcing us to look.
  4. Mental-map rebuilding is part of every session. The working-context bridge + SessionStart:compact hooks exist precisely to rebuild the mental map across compactions. This is slower than just trusting the AI to remember, and it’s supposed to be.

Where we’re still exposed. If any of three things break, SDG is right and we’re cooked:

None of these are hypothetical. All three are the failure modes to watch for.

The “more SEVs” prediction

SDG predicts more severity events as AI-written code ships without comprehension. We should track this in our own work: when a bug hits (strategy blows up, pipeline breaks, prediction is wildly off), was it caused by me writing code I didn’t really understand? Or by me trusting a summary I didn’t re-check? That’s the metric that separates the two regimes.

Curation section