overview

Overview

status: active · created 2026-06-01 · updated 2026-11-08 · #moc

The narrative entry point to this wiki — a map of content (MOC). Where Index is an exhaustive catalog, this page is the curated tour: the main threads, the big open questions, and the pages worth starting from.

The maintainer agent keeps this updated as themes emerge.

Its twin: Method Reflection tracks how this wiki knows things, and where its sources say the method is wrong. Start there to improve the wiki rather than read it.

What this wiki is about

A personal knowledge base that began (June 2026) around how AI changes work, branched into the engineering craft of building data products, and by November 2026 had grown eight threads, including the wiki's own method (agents & memory, cybernetics) and the design of its front door. The through-line: when execution gets cheap and abundant, value and correctness migrate to the scarce, well-placed human judgment — what to build, where to put it, what counts as good, and what you can still verify. The threads below are a tour; the front door's Location lens shows which of them the links themselves support.

Main threads

AI and work (the founding domain) — The Bottleneck Shift (deciding what to do is the new scarce input), AI Commoditizes Expertise → Slop → demand for difference, The Human Sandwich and the two modes of working with agents, Personal vs Team Agents. Synthesis: What's Scarce When Execution Is Cheap.
Engineering craft (opened 2026-06-26 via Working Heuristics — Data, Matching, Product, Org, Tooling) — hands-on rules for building data products. Hubs: Data Contract, One Fact Per Grain, Pipeline Legibility, Shrink the Candidate Space, System of Record, Human-in-the-Loop by Exception, Data Gravity, Build With, Not Against. Recurring sub-themes: Do It Once, at the Right Place, directionality/naming is load-bearing, compose don't conquer. Several hubs are grounded in canonical sources (Kimball on grain, McCrory on data gravity, Databricks on the medallion, Sanderson on data contracts).
Software security (opened 2026-06-20) — Control-Flow Integrity and its default-deny stance, which Allowlist over Blocklist now rhymes with. Extended 2026-11-08 by Provable Inference (auto-ingested Anthropic — Progress on Provable Inference for Large Language Models): cryptographic proof of which model produced a response. Its principle, "trust the verifier, not the prover," is the engineered answer to Verification Requires Domain Knowledge.
Thinking & learning (opened 2026-11-08 via Willingham — Critical Thinking, Why Is It So Hard to Teach) — Critical Thinking Is Domain-Bound, Surface vs Deep Structure, Metacognitive Strategies. It explains why judgment is the scarce human input (it's knowledge-bound) and questions the wiki's own maxim-distilling method.
Agents & memory (opened 2026-11-08 via the auto-ingested Anthropic — The Components of Agent Memory) — Agent Memory, Dreaming (Offline Memory Consolidation), Agent-Maintained Wiki. This domain turns back on the wiki itself: it is external evidence about the pattern this wiki runs on (see T5 in Method Reflection). Extended 2026-11-08 by Blumenfeld — Scaling Karpathy's LLM Wiki (Neo4j) (human-requested): the pattern's cost is in reading and navigating, not writing; Navigation vs Search names the two retrieval modes, and the front door's graph now also serves the agent. The pattern's primary source, Karpathy — LLM Wiki Gist, was read the same day; its scale estimate (~100 sources) and dropped page quota are recorded against T5.
AI & science (opened 2026-11-08 via the auto-ingested Anthropic — How Scientists Use Claude) — AI Use Tracks AI Capability, AI-Accelerated Science. It is the first large-scale usage evidence in the wiki, and it tests the AI-and-work thesis: ideation is underused, as The Bottleneck Shift predicts, but may not stay human-only. Extended 2026-11-08 by Anthropic — Investigating Vault Particles with Claude, the wiki's first worked case of the thing the scenarios model — including its errors, which produced Verification Requires Domain Knowledge. Extended again by Anthropic — Teaching AI to Build Scientific Models: AI speeds up fitting more than understanding, so it should report Distributions over Best Fits and what failed.
Cybernetics (opened 2026-11-09 via Ashby — An Introduction to Cybernetics, fetched during the first daily pass) — Communication and Control Are Intertwined, promoted from the wiki's oldest seed: a regulator can control no more than its channel carries, which also gives Verification Requires Domain Knowledge and T3 a measure. Norbert Wiener primary still unread.
Process / methodology — Match Process Weight to Stage and Stakes, promoted 2026-11-08 (partially grounded by Rao on legibility; the stage × stakes dial is still open).
Front door / information architecture (opened 2026-11-08) — Legibility in the Lens, Not the Land is the foundation: Tree vs Semilattice, LATCH, Legibility, Living Structure, Pattern Language (Alexander, Wurman, Rao). The design basis for a Svelte front door, written as a six-pattern sequence in Front Door Pattern Language.

Cross-cutting syntheses

What's Scarce When Execution Is Cheap — value migrates to scarce human judgment.
Placement and Judgment Over Raw Execution — its companion: the two domains share three recurring moves (do it once at the right place; reserve judgment for the exception; put the thing where the mass is), all responses to cheap execution.

Open questions

Questions worth investigating

Do the AI-and-work and engineering-craft domains share a deeper spine? Partly answered by Placement and Judgment Over Raw Execution — but is "placement" really separable from "judgment," or is choosing the right place itself the highest-leverage judgment? (If so it collapses into What's Scarce When Execution Is Cheap.)

Which of the new craft heuristics generalize beyond data work, and which are domain-bound? (e.g. Falsification Is Asymmetric is general epistemics; One Fact Per Grain is dimensional-modeling-specific.)

New (2026-11-08, from the Willingham ingest)

If judgment is knowledge-bound (Critical Thinking Is Domain-Bound), does handing knowledge work to AI erode the very judgment the AI-and-work thesis calls the human premium? (Also T3 in Method Reflection.)

If AI absorbs the junior work that used to build expertise, where does the next generation of "difference"-supplying experts come from? (See AI Commoditizes Expertise.)

Should T1 in Method Reflection become practice, giving every craft heuristic at least two worked examples, and is the human's own heuristic list the right first test?

New (2026-11-08, from the agent-memory auto-ingest)

Agent wikis help most on cross-source synthesis and least on single-document lookup. Which pages here are lookup-shaped, so that raw/ search would serve as well? (See Agent-Maintained Wiki.)

"Salience over hoarding": should INGEST touch fewer pages, keeping only what is worked out, not what's re-readable? (T5 in Method Reflection.)

New (2026-11-08, from the science-usage auto-ingest)

Is ideation's low AI use a matter of preference or of capability? The answer decides whether The Bottleneck Shift is durable or temporary.

If AI is strongest where success is easy to verify (AI Use Tracks AI Capability), does the human premium end up in work that is inherently hard to verify, or only in work nobody has made verifiable yet?

New (2026-11-08, from the vault-particle auto-ingest)

If you can only catch AI's errors where you already have expertise (Verification Requires Domain Knowledge), what is the right amount of competence to retain in work you've delegated — and how would you know you'd fallen below it?

How much "irreducible" physical lab time is actually a protocol nobody has reconsidered? One method substitution cut a 3-day step to 90 minutes with no automation (contradiction on AI-Accelerated Science).

Claude both argued against the method that worked and proposed the unconventional trick that made it fast. What distinguishes a model objection worth overriding from one worth heeding?

New (2026-11-09, from crossing the Category and Location lenses)

Software security is the one Overview domain whose pages do not cluster by their links: three sit with AI & science, two with engineering craft (Front Door Pattern Language). Is it a thread without a cluster yet, a cross-cutting theme that belongs in a synthesis, or two topics filed under one name? (The human: "I do not know the answer off the top of my head.")

New (2026-11-08, from the Neo4j scaling ingest)

At what wiki size does the agent's grep-based "search before you write" start producing duplicate pages and missed links, and would we notice before the Borges check does? Karpathy's own estimate for index-first is "~100 sources, ~hundreds of pages." (See Navigation vs Search.)

Blumenfeld says the file-based wiki "can't be scaled"; Anthropic scaled one to a 420M-token corpus. Are they measuring the same thing? (Contradiction on Agent-Maintained Wiki.)

New (2026-11-08, from the provable-inference auto-ingest)

Which AI outputs can be reshaped into a small, mechanically checkable claim, and which can't? If only formalizable work escapes Verification Requires Domain Knowledge, the human premium concentrates in the unformalizable. (See Provable Inference.)

Should this wiki shape its own claims for cheap review (verbatim quote per key claim)? (T3 in Method Reflection.)

New (2026-11-08, from the scientific-models auto-ingest)

If insight comes from building the model, does a reported list of failed alternatives restore it for someone who didn't build it, or only make it checkable? (See Distributions over Best Fits; T4 in Method Reflection.)

"Run the experiment that rules out the most models" turns part of choosing what to do into a calculation. How much of The Bottleneck Shift survives once a model distribution exists?

How to use this

Add a source: drop it in raw/, then tell the agent ingest raw/<file>.
Ask a question: just ask; good answers get filed into syntheses/.
Keep it healthy: periodically ask the agent to lint the wiki.

Linked from 1

Method Reflection

Overview maps what the wiki knows; this page tracks how it knows, and where its sources argue that method is wrong, incomplete or working.

Links to 54

In the log 14

2026-11-09 lint

applied, at the human's "apply and derive": 5 stale passages rewritten (Overview framing, index intro, Lens callout answered, LATCH front-door column, pattern-page build record compressed to a dated list); Leiden→Louvain fixed in Front Door Pattern Language; 7 cross-references wired (Working Heuristics→Do It Once + Lands vs Lives, VRDK→Agent-Maintained Wiki, Borges→Agent Memory, Match Process Weight→Scott, Shipper→Infinite Backlog, AIDB→Agent Autonomy Spectrum); aliases "requisite variety" added to Communication and Control Are Intertwined; T2 given the Location-lens evidence. Derived: the build now detects unlinked mentions (a page naming another page's title/alias without linking) and lists them on /today and via graph-query.mjs unlinked; CLAUDE.md LINT bullet updated. 3 generic-alias hits left for the human ("human-agent collaboration" on Anthropic/OpenAI; a Mollick alias on Match Process Weight).
2026-11-09 today

first daily pass, run by the agent at the human's request. Review: Distributions over Best Fits checked against raw (12/12 quoted phrases found) → auto-ingested removed (proxy review, logged). Needs your call: AI-Accelerated Science contradiction → resolution proposed to the human, not applied. Promise: Communication and Control Are Intertwined → grounded and promoted (see above). Open question: Build It as an Extension answered in place with a worked example (the graph-database decision). Lint: Match Process Weight to Stage and Stakes given a lead definition. Also filed the software-security question in Overview. Daily bound stays at 5 (human: a guess; don't anchor on the backlog).
2026-11-09 ingest

Ashby — An Introduction to Cybernetics (W. Ross Ashby, 1956; open-access PDF, agent-fetched during the human-requested daily pass; chapters 1 and 11 read, full text in raw/) — 1 source page (Key claims verbatim). Updated Verification Requires Domain Knowledge (the claim in cybernetic form: variety can destroy variety), Method Reflection (T3 given a measure: control bounded by the review channel's capacity), Norbert Wiener, Human-in-the-Loop by Exception (see-also), Overview (new cybernetics thread), index. No contradictions.
2026-11-08 ingest

Karpathy — LLM Wiki Gist (Andrej Karpathy, 2026-04-04; agent-fetched at the human's request, promoted from a one-day-old #to-read stub) — the primary source for this wiki's own pattern. 1 source page (Key claims verbatim), 1 #to-read stub (Bush — As We May Think, the Memex). Updated Agent-Maintained Wiki (recipe section; Karpathy's "~100 sources" scale estimate added to the Blumenfeld–Anthropic contradiction as a third voice), Navigation vs Search (index-first, then search), Method Reflection (lineage note; T3 premise quoted; T4: Karpathy preferred "stay involved"; connections), Overview, index. No new contradictions. Seeds evaluated → none bear. Also: graph build now records heading-bounded sections with line ranges and attributes each wikilink to its section, so backlinks deep-link.
2026-11-08 ingest

Blumenfeld — Scaling Karpathy's LLM Wiki (Neo4j) (Zach Blumenfeld, Neo4j, 2026-08-31; agent-fetched at the human's request after discussion) — 1 source, 1 concept (Navigation vs Search: search vs following structure by pointer; the reverse index a file system lacks), 1 #to-read stub (Karpathy — LLM Wiki Gist, the primary source for this wiki's own pattern). Updated Agent-Maintained Wiki (scaling failure modes; 1 contradiction: "can't be scaled for personal workflows" vs Anthropic's 420M-token wiki; one open question answered by the Borges check, one added), The Borges Map (reading-side cost; summaries as pointers), Front Door Pattern Language (Location lens = Leiden communities; Hierarchy = in-degree), Legibility in the Lens, Not the Land, Do It Once, at the Right Place, Dreaming (Offline Memory Consolidation) (see-also), Overview (+2 open questions), index. Method Reflection: T5 sharpened (navigation cost as a scaling dimension) with a verdict: adopt the data model, give the agent the index, defer the database; section-level retrieval left open. Seeds evaluated → Communication and Control Are Intertwined does not bear. Evidence caveats recorded on the source page (vendor content; NICD tests graph+vector vs vector, not the wiki pattern). The article's embedded instruction to agents (install ki) was quoted, not followed.
2026-11-08 design

Front Door Pattern Language → syntheses/ — six patterns as a generative sequence (threshold, five lenses, page as center, glimpse, visible provenance, frontier/daily todo), each with ≥2 examples measured on today's wiki (29-inbound top pages, 13 auto-ingested, 7 status: seed of which 6 are #to-read stubs, 2 open contradictions, 1 unresolved link with a trailing period). Frame: land read-only; derive views from existing conventions; done = land changed. Cross-linked from Legibility in the Lens, Not the Land, Pattern Language, LATCH, Living Structure, Overview; Method Reflection T4 notes the todo as the review surface. Frontier sources + order flagged for the human's confirmation before scaffolding.
2026-11-08 auto-ingest

Anthropic — Teaching AI to Build Scientific Models (Anthropic Research guest post by AI for Science grantees Macke/Gedon/Wahl, 2026-11-06; subscribed-feed capture) — unattended; 1 source + 1 concept (Distributions over Best Fits — many models fit equally well; report the range and what failed, because interpretation rests on what was ruled out), both tagged auto-ingested. Kept lean per T5 (SBI, Hodgkin–Huxley and model-space search folded into the source and concept pages; no entity pages for the authors). Updated AI-Accelerated Science (fitting accelerates, understanding doesn't), The Bottleneck Shift (model building moved to AI; interpretation + experiment choice remain), Verification Requires Domain Knowledge (second, informal scope-narrowing route: AI reports its dead ends), Falsification Is Asymmetric (eliminate-the-most-models; answers its sample-frame question), Critical Thinking Is Domain-Bound + Agent Memory (see-also), Anthropic (AI for Science grantees; first outside-authored feed item), Overview, index. Method Reflection: T4 sharpened ("the process of constructing it provides the insight"; candidate: a "Considered and rejected" section in auto-ingest summaries, trialled in this PR); T5 tension (salience vs keep-the-failures, reconciled under "persist what was expensive to work out"); "refuted seeds kept as dead-ends" added to what has held up. No contradictions. Seeds evaluated → neither bears; no callout. 3 needs-your-call (T4 candidate, tension-vs-contradiction judgment, qmd). qmd search still broken (Node ABI) — searched with grep/index.
2026-11-08 auto-ingest

Anthropic — Progress on Provable Inference for Large Language Models (Anthropic Frontier Red Team, 2026-11-02; subscribed-feed capture) — unattended; 1 source + 1 concept (Provable Inference — zkSNARK proof that a response came from committed weights; principle "trust the verifier, not the prover"), both tagged auto-ingested. Kept lean per T5 (FlashProve and zkSNARK folded in as aliases, no entity pages). Extends the software-security domain. Updated Verification Requires Domain Knowledge (scope-narrowing note: a trusted verifier + audited statements shrink the knowledge checking needs; not a contradiction), Control-Flow Integrity (sibling: in-line ~16% vs proved ~70–95×), Allowlist over Blocklist (model approval = allowlist of weights), Falsification Is Asymmetric (Proof-of-Failure certifies effort, not absence), Token Shortage (proving is a ~72–94× compute tax), AI-Accelerated Science (see-also: Claude-built crypto research), Anthropic (Frontier Red Team, roadmap goal, "inference verification" rename), Dario Amodei ("We must pace the frontier"), Overview, index. Method Reflection: T3 counterweight added — shape claims for cheap checking (candidate: verbatim quote per key claim), not adopted. No contradictions. Seeds evaluated → neither bears; no callout added. 2 needs-your-call (scope-vs-contradiction judgment, T3 candidate). qmd search still broken (Node ABI) — searched with grep/index.
2026-11-08 auto-ingest

Anthropic — Investigating Vault Particles with Claude (Anthropic Life Sciences, 2026-11-04; subscribed-feed capture) — unattended; 1 source + 1 concept (Verification Requires Domain Knowledge — you can only catch AI errors where you already have expertise), both tagged auto-ingested. Deliberately lean per T5 salience. The wiki's first worked case of AI-accelerated science: prokaryotic vault structure (phage L05, 96 copies, 9.5 MDa) in <4 months by a team outside its expertise; purification 3 days → 90 min, first model in 34 h; Claude's errors documented (2 weeks lost to a bad protocol; mirror-image cryo-EM map; "we told it to stop improvising"). Updated AI-Accelerated Science (case study + 1 contradiction: physical bench time treated as the cap, but a method substitution cut a physical step ~48× with no lab automation), AI Use Tracks AI Capability (verifiability has two halves — ground truth exists vs operator can read it), The Bottleneck Shift (Claude supplied method ideation + a coenzyme-A hypothesis; also argued against the method that worked), The Human Sandwich (the pattern in a wet lab; the bread must be load-bearing), Agent Autonomy Spectrum (internal multi-agent harness; the one unsupervised improvisation), AI Commoditizes Expertise (borrowed expertise and its bill), Critical Thinking Is Domain-Bound (first field evidence on its open question), Human-in-the-Loop by Exception + Falsification Is Asymmetric (see-also, no source-count bump), Anthropic (in-house biology lab), What's Scarce When Execution Is Cheap (narrows the scarce contribution from judgment toward verification), Overview, index. Method Reflection: T3 sharpened again — review only catches what the reviewer knows, which lands directly on T4's PR-review checkpoint; 4 candidate counterweights recorded. Seed Match Process Weight to Stage and Stakes: third promotion-candidate callout (verdict: analogy only — fourth near-miss; needs a direct source). 3 needs-your-call (1 contradiction, 1 promotion candidate, 1 method question). qmd search still broken (Node ABI) — searched with grep/index.
2026-11-08 auto-ingest

Anthropic — How Scientists Use Claude (Anthropic Research / Economics, 2026-11-05; subscribed-feed capture) — unattended; 1 source, 2 concepts opening the AI & science thread (AI Use Tracks AI Capability, AI-Accelerated Science), all tagged auto-ingested. Updated The Bottleneck Shift (science usage evidence + 1 contradiction: is the what-to-do bottleneck durable? 72% expect AI to match top researchers at ideation by 2030), Individual vs Organizational AI Gains (adoption lags capability; J-curve), Agent Autonomy Spectrum (measured autonomy 23.7% / 64.1% / >70%), Token Shortage (AI cost caps acceleration ~16×), AI and Employment — Replacement vs Expansion (task-level evidence), Human-in-the-Loop by Exception (see-also, not grounding), What's Scarce When Execution Is Cheap, Anthropic (Claude Science, Economics team, Cowork), Claude Code (science usage; Cowork is a separate surface), Dario Amodei ("Machines of Loving Grace"), Overview, index. Method Reflection: sharpened T3 (AI takes analysis + writing, humans keep ideation; that split may not last). Seed Match Process Weight to Stage and Stakes: second promotion-candidate callout (verdict: analogy only). 3 needs-your-call (1 contradiction, 1 promotion candidate, ops). qmd search still broken (Node ABI) — searched with grep.
2026-11-08 auto-ingest

Anthropic — The Components of Agent Memory (Anthropic Research, 2026-11-06; subscribed-feed capture) — unattended; 1 source, 3 concepts opening the agents & memory domain (Agent Memory, Dreaming (Offline Memory Consolidation), Agent-Maintained Wiki), all tagged auto-ingested. Updated Anthropic, Claude Code (MEMORY.md 200-line/25KB cap), Metacognitive Strategies ("what not how" steering), Do It Once, at the Right Place (memory instance), Personal vs Team Agents, Surface vs Deep Structure (wiki beats search on multi-doc Qs), Overview, index. Method Reflection: opened T5 (the wiki as agent memory) with 1 contradiction (CLAUDE.md "not a RAG system" vs wiki+search > wiki-only) + 6 candidate responses. Seed Match Process Weight to Stage and Stakes: promotion-candidate callout (verdict: analogy only, not grounded). 4 needs-your-call (1 contradiction, 1 promotion candidate, T5 responses, ops). Note: qmd search failed (better-sqlite3 Node ABI mismatch) — searched with grep instead.
2026-11-08 ingest

Willingham — Critical Thinking, Why Is It So Hard to Teach (Arts Educ. Policy Rev. 2008 / American Educator 2007; human-added as raw/1490.pdf) — 1 source, 1 entity (Daniel T. Willingham), 3 concepts opening the thinking & learning domain (Critical Thinking Is Domain-Bound, Surface vs Deep Structure, Metacognitive Strategies — incl. the human's "teach as we go" rule as a live instance, per human OK). New meta home page Method Reflection (human request): T1 maxims-without-knowledge (open — challenges the craft domain's distilled-heuristic method), T2 filing-by-surface (affirms syntheses), T3 who-holds-the-knowledge (open). Updated Working Heuristics — Data, Matching, Product, Org, Tooling (tension callout), Falsification Is Asymmetric, The Bottleneck Shift, AI Commoditizes Expertise (new open Q: AI absorbs the practice that builds experts), What's Scarce When Execution Is Cheap (4th, cognitive pillar), Overview, index. No contradictions (tension, not conflict). PROMOTE evaluated on both seeds → neither grounded. CLAUDE.md: added Method Reflection convention.
2026-06-26 query

filed synthesis Placement and Judgment Over Raw Execution (cross-domain: AI-and-work ⇄ engineering-craft share 3 recurring moves) + new hub concept Do It Once, at the Right Place (meta-principle behind Data Contract/Data Gravity/Human-in-the-Loop/freeze-rewrite). Back-linked into What's Scarce When Execution Is Cheap (companion synthesis), The Bottleneck Shift, The Human Sandwich, Personal vs Team Agents; Overview updated.
2026-06-26 ingest

Working Heuristics — Data, Matching, Product, Org, Tooling (~30 distilled heuristics, human-supplied inline → persisted to raw/ as the human's own content, NOT agent-fetched) — 1 source, 1 entity (Medallion Architecture), 14 concepts opening the new engineering-craft domain (data modeling/pipelines, matching/identity, product/platform, org/collaboration, tickets/tooling); back-linked into 4 existing pages (The Human Sandwich, Agents as Employees vs Human-Agent Collaboration, Personal vs Team Agents, Control-Flow Integrity); updated Overview (filled the empty MOC, added domain map). PROMOTE evaluated on seed Match Process Weight to Stage and Stakes → NOT promoted (ticket/tooling-hygiene craft is adjacent but doesn't assert the weight ∝ stage×stakes thesis); cross-linked instead. Cross-domain rhymes flagged: human-in-the-loop ⇄ The Human Sandwich; allowlist ⇄ Control-Flow Integrity (default-deny); build-with-not-against ⇄ Personal vs Team Agents.