PK Wiki

overview

Index

· updated 2026-11-08

The catalog of everything in this wiki. The maintainer agent keeps this current on every ingest. On a query, this is the first page it reads to find what's relevant. Start at Overview for the narrative entry point.

Each entry is a wikilink plus a one-line summary, grouped by category. Pages marked seed are your own unverified hunches; #to-read stubs are reading leads. Jot raw ideas into Inbox for later processing. The same catalog, in five orders, is the front door's lenses (wiki.notabox.app).

Overview

  • Overview — top-level map of the wiki and its main threads
  • Method Reflection — the wiki examining its own method: live tensions (T1 maxims without knowledge, T2 filing by surface, T3 who holds the knowledge (sharpened by the science-usage split), T4 automation vs engagement, T5 the wiki as agent memory, now with a navigation-cost branch and a verdict: adopt the graph model, defer the database; T3 sharpened twice, plus a "shape claims for checking" counterweight from provable inference; T4 sharpened by "insight comes from the process," and a keep-the-failures tension on T5), auto-ingest review queue, method decisions that held up, Dataview views

Sources

Subscribed feed — Anthropic Research (auto-ingested; auto-ingested until reviewed):

  • Anthropic — The Components of Agent Memory — Anthropic Engineering (2026-11-06): memory as prediction, salience vs hoarding, "what not how" steering, offline dreaming, and an evaluation of agent-maintained wikis (wiki + search beats either alone; Borges bloat; verbatim > summary). Opens the agents & memory domain.
  • Anthropic — Investigating Vault Particles with Claude — Anthropic Life Sciences (2026-11-04): first 3D structure of a prokaryotic vault particle (phage protein L05, 96 copies, 9.5 MDa), solved in under four months by a team outside its expertise; Claude ran search, construct design, purification optimization and cryo-EM. Documents its errors too — the wiki's first worked case of AI-accelerated science.
  • Anthropic — How Scientists Use Claude — Anthropic Economics (2026-11-05): 207k research sessions; use tracks AI capability (top quintile 2–4× expected use); analysis and writing over-used, ideation under-used despite being ranked most important; acceleration ~2× today, up to ~15× if non-physical work is automated, capped by physical work and AI cost. Opens the AI & science thread.
  • Anthropic — Progress on Provable Inference for Large Language Models — Anthropic Frontier Red Team (2026-11-02): FlashProve, an open-source zkSNARK prototype proving a 400B MoE's outputs came from committed weights (~110 tok/s on one H100, ~19 MB proofs, ~17 s CPU verify, ~72–94× inference cost). Claude-built; Lean proofs cover the security claims; "trust the verifier, not the prover." Extends the software-security domain.
  • Anthropic — Teaching AI to Build Scientific Models — AI for Science grantees Macke, Gedon and Wahl (2026-11-06; guest post): build → constrain → interpret, through Hodgkin–Huxley; simulation-based inference (~100k simulations, inference in seconds); LLM agents now design model equations; many models fit equally well, so report the distribution and what failed; "the process of constructing it… provides the insight."

Daily-pass grounding (agent-fetched 2026-11-09 at the human's request):

  • Ashby — An Introduction to Cybernetics — W. Ross Ashby (1956), chapters 1 and 11 read: cybernetics as "ways of behaving"; the Law of Requisite Variety ("variety can destroy variety"); regulation interpreted as communication, "R's capacity as a regulator cannot exceed R's capacity as a channel of communication." Grounds (partially) the wiki's oldest seed; gives T3 and Verification Requires Domain Knowledge a measure.

Human-requested ingest (agent-fetched 2026-11-08):

  • Blumenfeld — Scaling Karpathy's LLM Wiki (Neo4j) — Zach Blumenfeld, Neo4j (2026-08-31): the LLM-wiki pattern fails at scale on read/search/navigation, not writing; a deterministic graph index (documents, sections, links, backlinks) makes "a hop a pointer rather than a scan"; ki reference implementation; Leiden themes; NICD evidence (caveats noted). Grounds Navigation vs Search; 1 contradiction with Anthropic's scale result.
  • Karpathy — LLM Wiki Gist — Andrej Karpathy (2026-04-04): the recipe this wiki runs on (raw / wiki / schema; ingest, query, lint; index + log); "~100 sources, ~hundreds of pages" before search is needed; "10-15 pages" per source as an example, not a target. Read 2026-11-08; four lines mapped onto T3–T5.
  • Bush — As We May Think #to-read — the Memex (1945): associative trails, the ancestor of navigation over documents; flagged via Karpathy.

Front-door groundings (agent-fetched 2026-11-08):

Canonical groundings (agent-fetched 2026-06-26 under the auto-ground convention):

Reading leads (flagged, not yet ingested):

Entities

  • Every — AI-native publication/product/consultancy; the lead case study for agent-saturated work.
  • Dan Shipper — CEO of Every; author of "After Automation," the episode's primary text.
  • NLW (Nathaniel Whittemore) — host of The AI Daily Brief; author of the "infinite backlog" and "human premium" framings.
  • The AI Daily Brief — daily AI news/analysis podcast; a recurring source channel.
  • OpenClaw — early-2026 maximal-autonomy agent paradigm (Mac mini + Telegram + heartbeats); now seen as overshooting.
  • Codex — OpenAI's agentic harness; anchors the semi-synchronous middle of the autonomy spectrum.
  • Claude Code — Anthropic's agentic harness (also "Claude Cowork"); cited alongside Codex.
  • Dario Amodei — Anthropic CEO; "half of entry-level white-collar jobs" — the replacement camp's headline.
  • Ethan Mollick — Wharton professor; author of One Useful Thing and the Leadership/Lab/Crowd framework.
  • One Useful Thing — Ethan Mollick's newsletter on practical AI; a recurring source channel.
  • Manus — Claude-based autonomous agent Mollick uses as an agentic-capability benchmark.
  • Anthropic — AI lab; maker of Claude / Claude Code; model base for Manus; led by Dario Amodei.
  • OpenAI — AI lab; maker of ChatGPT and Codex.
  • Norbert Wiener — founder of cybernetics; in the wiki via a secondary excerpt (information/originality, → Clichés Carry Less Information); primary works still unread.
  • Jean E. Fox Tree #to-read — psycholinguist on conversational devices as coordination signals; reading lead for Communication and Control Are Intertwined.
  • Christopher Alexander — architect; pattern languages, tree vs semilattice, living structure.
  • Richard Saul Wurman — information architect, TED founder; LATCH. #to-read: Information Anxiety.
  • James C. Scott #to-read — Seeing Like a State; origin of legibility.
  • Daniel T. Willingham — UVA cognitive psychologist; "Ask the Cognitive Scientist" columnist; author of the critical-thinking source.
  • Venkatesh Rao — legibility essay read 2026-11-08; 2×2 method still #to-read. Was: Ribbonfarm writer; 2×2 quadrant frameworks + "legibility" (imposed process destroys value); reading lead for Match Process Weight to Stage and Stakes.

Engineering-craft domain (via Working Heuristics — Data, Matching, Product, Org, Tooling):

  • Medallion Architecture — layered data-refinement pattern (bronze→silver→gold); a view-only medallion is where the "lands vs lives" distinction bites.

Concepts

Engineering-craft domain (all via Working Heuristics — Data, Matching, Product, Org, Tooling, 2026-06-26):

Data modeling & pipelines:

  • Lands vs Lives — where data physically lands (the base table) vs. where it lives (the modeling layer); a debugging/ownership compass.
  • Data Contract — derive once at the boundary so every consumer inherits a correct value instead of re-deriving (and re-bugging) it.
  • One Fact Per Grain — keep a fact/dimension to its grain; model the negative case (a "didn't-happen" metric needs both planned + actual rows).
  • Pipeline Legibility — name tables for the flow direction, trust lineage you can read, design ingestion to the source's real export contract.

Matching & identity:

  • Shrink the Candidate Space — unreliable join key → reduce the pairs you consider (blocking), don't tune the matcher.
  • System of Record — a convenient source isn't authoritative; verify authority before promoting to SOT.
  • Human-in-the-Loop by Exception — automate the bulk, reserve humans for true exceptions, pay each decision once, make every confirmation teach the system.
  • Allowlist over Blocklist — name what you keep and drop the rest; fails closed, so sensitive data can't leak by omission.

Product & platform:

Org / collaboration & tooling:

  • Build With, Not Against — turf friction is an unowned boundary; draw the seam, concede where they're better, follow customer-best, make the durable layer their extension.
  • Scope Hygiene for Tickets and Tooling — split orthogonal scope, freeze contingent specs while upstream is open, keep directional links right, scope tools to their users.

Cross-cutting:

  • Do It Once, at the Right Place — meta-principle behind several of the above: do the work once, at the boundary/data/decision/time where it's correct for everyone, instead of re-deriving it. (Hub for Data Contract ⇄ Data Gravity ⇄ Human-in-the-Loop ⇄ freeze-and-rewrite-once.)

Thinking & learning domain (via Willingham — Critical Thinking, Why Is It So Hard to Teach, 2026-11-08):

  • Critical Thinking Is Domain-Bound — critical thinking = reasoning/judgment/problem-solving done effectively, novelly, self-directedly; not a portable skill, because it runs on domain knowledge + practice.
  • Surface vs Deep Structure — people retrieve by scenario, not principle, so solutions fail to transfer (19% → 35% with a hint); deep familiarity makes deep structure perceptible.
  • Metacognitive Strategies — maxims about how to think: cue, not capability; Willingham's staged teaching recipe; the human's "teach as we go" rule as a live instance.

Agents & memory domain (via Anthropic — The Components of Agent Memory, auto-ingested 2026-11-08):

  • Agent Memory — knowledge an agent keeps between sessions; every note is a prediction; salience over hoarding; injected top layer + searched depths.
  • Dreaming (Offline Memory Consolidation) — offline reflection over transcripts to correct, represent, promote, focus and gap-fill memory; this wiki's LINT is a dream pass.
  • Agent-Maintained Wiki — the LLM-wiki pattern (and this wiki's own): compounds, beats search on multi-doc synthesis, best with search; fails via bloat, summary loss, many-writer overwrites, and (Blumenfeld) read/navigation cost that grows with size. Now carries a contradiction on whether the file-based form "scales."
  • Navigation vs Search — two retrieval modes: search matches a phrase against everything; navigation follows links, backlinks, sections and communities by pointer. A wiki needs both (wiki+search 61% > either). The agent now gets the front door's graph through a CLI. Grounded by Blumenfeld + Anthropic + Karpathy, 2026-11-08.

AI & science (via Anthropic — How Scientists Use Claude, auto-ingested 2026-11-08):

  • AI Use Tracks AI Capability — people use AI where it works: use ratio rises with task capability, which tracks RL suitability (0.78) and cognitive work; physical, interpersonal and regulated tasks lag.
  • AI-Accelerated Science — research-acceleration scenarios (~2× → 8× → 15–16×); bottlenecks: physical work (~40% of bench time), AI cost (~16× ceiling), ideation, adoption lag. Now carries a worked case and a contradiction on whether physical bench time is really the cap.
  • Verification Requires Domain Knowledge — you can only catch an AI's mistakes in areas you already know, so the work you delegate for lack of expertise is the work whose errors cost most; the precondition "human in the loop" assumes.
  • Distributions over Best Fits — many parameter sets and model structures fit the same data; report the range, what all fits share, and which alternatives failed, because interpretation rests on what was ruled out. Auto-ingested 2026-11-08; reviewed 2026-11-09 (all quotes verified against raw).

Front-door foundations (agent-fetched groundings, 2026-11-08):

  • Tree vs Semilattice — living structure overlaps; minds and planners flatten it into trees. The wiki is a semilattice, its folders a tree (T2).
  • Living Structure — Alexander: centers built of centers; structure-preserving unfolding ("edit over create").
  • Pattern Language — generative sequences, not idea catalogs; the test the heuristics list fails (T1).
  • Legibility — Scott/Rao: simplifying a reality for the central eye kills what made it work.
  • LATCH — Wurman: Location, Alphabet, Time, Category, Hierarchy (= magnitude). Five lenses.
  • Match Process Weight to Stage and Stakes — promoted from seed 2026-11-08: legibility grounds the "high ceremony, low value" cell; the stage × stakes dial is still open.
  • The Borges Map — a wiki as large as its sources abstracts nothing; scripts/lint/borges.py measures it.

Cybernetics (opened 2026-11-09 via Ashby — An Introduction to Cybernetics):

  • Communication and Control Are Intertwined — promoted from seed 2026-11-09: a regulator can control no more than its channel carries (requisite variety); the converse and the conversational half stay open. Gives T3 and verification a measure.

Syntheses

  • What's Scarce When Execution Is Cheap — Wiener ⇄ Shipper ⇄ Mollick converge: when default output is free, value migrates to the scarce, high-information, high-judgment human contribution.

  • Placement and Judgment Over Raw Execution — companion to the above: the AI-and-work and engineering-craft domains keep landing on the same three moves (do it once at the right place; reserve judgment for the exception; put the thing where the mass already is) — all responses to execution getting cheap.

  • Legibility in the Lens, Not the Land — front-door foundation: organize the views (LATCH trees) freely; never reorganize the knowledge (the semilattice) to suit a view. Wurman ⇄ Rao ⇄ Alexander.

  • Front Door Pattern Language — the front door as a six-pattern generative sequence (threshold → five lenses → page as center → glimpse → visible provenance → frontier as daily todo), each with concrete examples from this wiki; "the land is read-only," "done means the land changed." Thin slice = patterns 1, 3, 6 (+5).

Seeds

  • Communication and Control Are Intertwined → promoted 2026-11-09 to Concepts (partially confirmed by Ashby). Was: hunch that communication and control are two faces of one phenomenon (captured 2026-06-01; evaluated 2026-06-02 against the Wiener excerpt → not promoted).

No open seeds. Capture a hunch and one will appear here.

  • Match Process Weight to Stage and Stakes → promoted 2026-11-08 to Concepts (partially confirmed by Rao). Was: hunch that process weight should be dialed to project stage and stakes: lightweight discovery loops early, ceremony added only where a specific pain appears, heavy SDLC only when coordination/correctness risk exceeds the ritual's cost (captured 2026-06-20; sharpened 2026-06-20 with a value × ceremony 2×2 — "high value, low ceremony" sweet spot, ala Venkatesh Rao). Opens a new process/methodology domain; no source yet. Grounding candidates: Boehm & Turner, Cynefin, Cockburn.

Dataview (live tables)

Once the Dataview plugin is installed, these blocks render automatically. Until then they show as code.

All sources, newest first:

TABLE author, source-type, date-published AS published, updated
FROM "wiki/sources"
SORT updated DESC

Pages flagged stale, superseded, or unsourced (lint targets):

LIST
FROM "wiki"
WHERE status = "stale" OR status = "superseded" OR status = "draft" OR contains(tags, "unsourced")

Unverified seeds and reading leads (your hunches awaiting a source):

LIST
FROM "wiki"
WHERE status = "seed" OR contains(tags, "hypothesis") OR contains(tags, "to-read")
The wiki is the source. This door reads it at build time and never writes to it.