Where AI Value Lands

log

60 min read

Log

Append-only chronological record of wiki operations. Entry format: ## [YYYY-MM-DD] {ingest|ask|lint|discover|edit} | Title. Convention: newest entry at the top (prepend).

[2026-05-27] ask | Crystallize the 2026-05-27 daily thought as the first page in a new 07_business-ideas/ folder

  • Origin: /ask invoked on the 2026-05-27 daily thought (raw/daily tougths/2026-05-27.md, Italian). The note proposed a business: take an open-source model (Gemma 4 / Qwen / DeepSeek 3.7), do RL to close a specific capability gap, keep weights closed, sell metered API access, target agent builders who plug it into their orchestrators via MCP. Two example domains: LinkedIn-post writing and visual defect detection.
  • Opinion delivered (in-conversation, not filed as a page): the idea re-derives H4_rl-specialization-value-pocket from the operator’s seat. The two named examples diverge — visual defect detection is H4-canonical; LinkedIn-post writer fails the data-moat condition and belongs to the ai-development-facilitator 2021 → 2026 vintage problem (software-as-temporary-artefacts). The proposed business model (closed, hosted, metered API) is Track A of H5_ai-as-operational-not-product, inheriting its capital-locked dynamics. The keeper is the orchestrator-composition pattern: agent builders would pay for a drop-in specialist callable via MCP for the one node in their graph where their orchestrator underperforms. The author confirmed the composition framing as the core thesis worth filing.
  • New folder: wiki/07_business-ideas/ — speculative business / company pages, distinct from 06_ideas/ (observed use-case patterns) and from 02_hypothesis/ (project-paper research claims). README installed with explicit page contract.
  • New page: specialist-subagent-for-orchestrators — first page in the folder. Status: speculative. Confidence: low.
    • Mechanism (wedge): buyer keeps their orchestrator; rents a specialist for the one node where accuracy matters and they have neither the data nor the RL stack to close the gap themselves.
    • Three required conditions: (1) proprietary data (industrial / contractually-exclusive / captured-trajectory) — without it the next foundation model eats you; (2) vertical narrow enough that frontier labs structurally do not horizontally absorb; (3) low-friction integration surface (MCP today).
    • Three GTM options: A — single industrial customer partnership (H4_rl-specialization-value-pocket-canonical; only cold-start path that solves the data-moat); B — closed-source metered API (Anthropic-style; Track A of H5_ai-as-operational-not-product; hosting alone is not defensible — see rl-apis); C — data-aggregator consortium → shared specialist (reifies exclusive-industry-data-partnerships as a company; long-game defensibility). Plausible sequence: A → B for additional same-vertical buyers → C when consortium-scale.
    • Bear cases: frontier-lab horizontal absorption; single-customer surplus clawback; MCP commodification + buyer self-train on rl-apis; operations cost dominance (ml-monitoring-quality-cycle 17-practice floor); orchestrator-vendor first-party specialist marketplace (largest 2026–2028 disruption vector — re-asserts distribution-moat on top of the specialist).
    • Five open questions: viable vertical candidates (H6_industrial-ai-rollup-captive-suppliers geography first); minimum viable specialist surface; buyer identity (agent-builder vs end-industrial firm); customer-acquisition motion (gultekin-pinarbasi-commercialisation-ai-2025 gap 7); frontier-lab in-product specialist-routing as gating risk.
  • Reverse-link patches (Related sections only — Referenced-by deferred to /lint): H4_rl-specialization-value-pocket (buyer-side composition framing); H5_ai-as-operational-not-product (Track A/B framing → GTM choice is track choice); ai-first-company-loop (Layer 3 slot in the buyer’s loop); software-as-temporary-artefacts (second category of surviving middle alongside primitives — closed by buyer’s missing data); rl-apis (competitive surface on hosting); exclusive-industry-data-partnerships (Option C’s structural template — got an inline Referenced-by entry too).
  • Index updated: new ## Business ideas — 07_business-ideas/“ section with one entry; new last-edit comment under # Wiki Index; previous last-edit demoted to Previous: line; last-updated frontmatter bumped to 2026-05-27.
  • What this does not do: long-tail Referenced-by patches for the ~6 additional pages the new page forward-links to (middle-layer-defensibility, distribution-moat, ai-charged-product-service-provider, ai-development-facilitator, ml-monitoring, ml-monitoring-quality-cycle) and the candidate-vertical ideas pages (predictive-maintenance, minerals-processing-optimization, autonomous-drone-inspection) and RL-tech pages (rl-from-verifiable-rewards, rl-data-preparation) — deferred to next /lint. Does not run /reflect on the 2026-05-27 note (no annotated sibling produced; the author may invoke it separately).

[2026-05-26] visualize | personal-liability-exposure (revised)

  • Output: outputs/personal-liability-exposure.html (~65 KB, single self-contained file).
  • Revision after first pass: hero replaced (radiating-arrows liability surface was unclear). New hero is the five-layer loop with Layer 4 highlighted in a solid orange box; learning feedback curves back over the top to all four prior layers (with the author’s 2026-05-26 amendment baked in). New interactive See. section between hero and Build.: a clickable five-position autonomy slider with two live columns (your queue / running autonomously) showing eight action classes redistributing as the slider moves; irreversibles (pay-invoice, deal-commit) pinned to the queue regardless of slider position; a ”▶ simulate 90 days (auto-ratchet)” button walks the slider from tight → loose to demonstrate ratchet maturation over time.
  • Build. retained: snapshot of N per-class sliders after ~90 days — the architectural truth that it is not one slider but one per action class.
  • Tuned for arrivals from a LinkedIn post on the concept. Practical Ship. section gives copy-pasteable on-ramps for an operator’s own agent (CLAUDE.md ledger paragraph, .claude/settings.json PreToolUse + PostToolUse hook stub, a /ratchet sibling skill sketch, and an inverted “what not to do” card).
  • Nodes: 14 (1 self / 8 concept / 2 hypothesis / 3 source). Edges: 21.
  • Sections (verb-word identifiers, orange period): Frame · See · Build · Ship · Trace · Hold · Cite.

[2026-05-26] ask | Why fully autonomous AI isn’t here even when the five layers are — personal liability as the Layer 4 binding constraint

  • Question (author): given that the model can do what it needs to do, the data exists, and physical autonomy is not the blocker, what psychologically stops a founder / consultant / entrepreneur from completely delegating to AI with guardrails? Is it a missing outline, missing execution, or something else?
  • Synthesis answer: the wiki decomposes the resistance into three pieces — (1) the eval-real-world-gap is empirically real (jagged intelligence, anterograde amnesia), (2) liability allocation is not engineering reliability (autonomy-slider via deloitte-ai-dossier-eri: “someone must be on the hook”), and (3) verification capacity ≠ judgment (fusion-skills). Architecturally, what is missing in most operator-builder setups is not Layers 1/2/3/5 of ai-first-company-loop but Layer 4 with a learning hook — per-action-class quality gates whose tightness ratchets down automatically as the action class earns trust.
  • Crystallized: personal-liability-exposure — new concept page. Decomposes Layer 4 of ai-first-company-loop into:
    • Hooks — typed pre-/post-action regulation triggered by action class (send-external-message tone+allowlist, deal-commit threshold, publish-content brand-voice scan, write-to-customer-record schema+audit). Encoded once, executes at loop speed; right home for deloitte-ai-dossier-eri’s “escalation paths” applied to operator-builder stakes.
    • Per-action-class autonomy slider with auto-ratchet — no workflow gets delegated on first execution; default high gate tightness; each successful execution loosens one notch. Architectural claim: if the ratchet is not automatic, the founder accumulates effort instead of shedding it and the H5_ai-as-operational-not-product surplus collapses into supervision time. Slider must be UI-accessible per workflow (first-class operating surface, not buried YAML).
    • Verification surface (Karpathy Iron Man) — the third leg; without it the operator’s review cost stays high enough that the ratchet never trips.
  • Named the load-bearing variable: personal liability exposure. Every delegated action is taken in the operator’s name; for a solopreneur the action and the person are the same surface, so small-action-high-frequency aggregates into a career-defining surface. The resistance is not psychology alone — it is a rational response to a system whose only setting is “off” or “max.”
  • Patches (3):
  • Index updated: new last-edit header summarising the crystallization; one new entry in the Concepts catalog; frontmatter last-updated comment bumped to “fourth edit of the day.”
  • What this does not do: long-tail reverse-link patches for the other pages personal-liability-exposure forward-links to (eval-real-world-gap, fusion-skills, continual-learning-paradigm, karpathy-software-3, middle-layer-defensibility, software-as-temporary-artefacts, H6_industrial-ai-rollup-captive-suppliers, deloitte-ai-dossier-eri, sutskever-age-of-research, vertical-ai-orchestration) — defer to next /lint sweep.

[2026-05-26] reflect | Supply-chain rollup (H6) and the self-improving loop (YC five-layer framework)

  • Daily note: raw/daily tougths/2026-05-26.md — two ideas: (1) SME captive-supplier supply-chain rollup with AI replacing the management layer; (2) software as temporary artefacts + the YC Root Access AI-first-company self-improving loop.
  • Annotated: raw/daily tougths/2026-05-26 annotated.md.
  • Memory marker: supply-chain-rollup-and-self-improving-loop.
  • New pages (3):
    • H6_industrial-ai-rollup-captive-suppliers — new working hypothesis (low confidence; speculative thesis with descriptive premises supported by wiki). Holding company acquires supplier-tier SMEs at 40–60%+ single-customer revenue concentration, preserves manufacturing labor and physical production, replaces the management layer with the ai-first-company-loop, amortises one centralised AI-operations team across N acquired suppliers. Natural geography: Italian industrial districts. The rollup variant of H5_ai-as-operational-not-product — H5 at single-firm scale, H6 at multi-firm holding-company scale; both Track B.
    • ai-first-company-loop — new concept page; five-layer firm-level operating model (sensor data / policy / tool / quality gates / learning) adapted from YC Root Access “How to Build a Self-Improving Company with AI” (https://www.youtube.com/watch?v=t-G67yKAHBQ). Firm-level cousin of ml-monitoring-quality-cycle; Protschky scaffolds production discipline for one ML application, the YC loop scaffolds the entire company. Borrowed-framework citation per author’s election; no full source page for the YC video. Authorial correction baked in: learning closes back into all four prior layers, including the quality gates themselves (gates are learnable too).
    • software-as-temporary-artefacts — new concept page; strategic-consequences page derived from Layer 3 (tool) of the loop. Agents in the loop synthesise bespoke tooling on demand; the middle of H2_u-curve-of-value is sharpened — not just squeezed but in large part abolished — the surviving middle re-classifies as primitives for synthesis (Cursor’s IDE+model surface, LangGraph, RL APIs, eval tooling), not finished workflow products. The squeeze on ai-development-facilitator tightens from a third direction (in-house substitution) on top of the two existing mechanisms (frontier-lab cannibalisation, commoditised models).
  • Patches (6 reverse-link patches):
  • Planning updates: 09_paper-portfolio — P2’s load-bearing hypothesis set expanded to H5+H6 with operator-side concept anchors ai-first-company-loop and software-as-temporary-artefacts. 05_open-questions — two new sections: 5 H6-related P2 sub-questions (newness vs structural; natural diffusion timeline; acquisition acceleration; competitive delta empirically; H5/H6 separation) and 2 loop/artefacts sub-questions (does artefact substitution tighten H2 into an L; is gate-design a defensibility surface).
  • Index updated: new last-edit summary header; three new entries (one in Hypotheses, two in Concepts).
  • Key takeaway: the two threads share a single mechanism. ai-first-company-loop gives “managed by AI” an operational definition; without it, H6’s prescriptive half is metaphor. software-as-temporary-artefacts is H6’s second amortisation engine (per-firm SaaS cost collapses with synthesis) and the wiki’s sharpest reading of where the U-curve middle actually dies.
  • What this does not do: long-tail reverse-link sweep (autonomy-slider, continual-learning-paradigm, karpathy-software-3, sme-ai-adoption-gap, sme-champion-vertical-ai, vertical-ai-orchestration, application-layer, foundation-model-layer, circular-ai-economy, jevons-paradox-ai, context-control, digital-core, fusion-skills, taker-shaper-maker, value-capture, ai-charged-product-service-provider, and several data pages) — H6 / ai-first-company-loop / software-as-temporary-artefacts forward-link to these but the reverse links are not yet applied. Defer to next /lint sweep.

[2026-05-26] edit | Added H5 (operator-side hypothesis) and 09_paper-portfolio (P1/P2 split)

  • Trigger: author articulated a new operator-side thesis in chat: AI captures durable value as an operational input to a non-AI business, not as a delivered product, for builders outside the VC-megafund race-in track (Cursor / Windsurf / Perplexity). Two distinct surfaces emerged — a falsifiable hypothesis and a paper-portfolio decision (wiki now feeds two papers).
  • New pages (2):
    • H5_ai-as-operational-not-product — working hypothesis. Frames two discrete tracks (Track A = VC-megafund race-in / AI-as-product; Track B = operator-builder / AI-in-operations / non-AI deliverable). Confidence medium overall (descriptive half well-supported by existing wiki composition; prescriptive operator-side half unvalidated). Bear case includes the “Cursor for local SEO” small-vertical orchestration counter-example author flagged explicitly. The first project hypothesis written from the operator’s seat rather than the analyst’s.
    • 09_paper-portfolio — paper-portfolio decision (P1 = existing descriptive paper Where Value Lands, audience = capital-rich operators; P2 = new prescriptive operator-paper What to Start Now, audience = capital-light operator-builders without a distribution prior). Shared wiki feeds both. P1 ships first; P2 drafted after H5 hardens under 2–3 ingest cycles.
  • Patches (16):
  • Key takeaway: the wiki now hosts a two-track operator/seller decomposition that H1–H4 implicitly assumed but never named. H5’s load-bearing claim is the binary Track A / Track B framing; the strongest bear case is the “Cursor for local SEO” archetype that, if it exists at sub-Series-B scale, would force a third track and revise H5.
  • What this does not do: H5’s ## Referenced by block is a stub (lists only 05_open-questions and 09_paper-portfolio); a future /lint sweep should fill it with the back-references from the 13 concept pages patched in this pass plus any others the body text mentions. Reverse links not yet applied to: sme-champion-vertical-ai, business-model-portfolio, ai-startup-business-archetypes-weber, ai-development-facilitator, ai-data-analytics-provider, ai-deep-tech-researcher, non-tech-digital-core-synthesis, wef-ai-in-action-2025, deloitte-ai-dossier-eri, where-value-lands-2026, several data pages — all of which H5 forward-links to. Add via /lint.
  • Hypothesis touches: H1 neutral (operator-side / seller-side split is orthogonal to substrate ladder). H2 narrowed (H5 explicitly claims top of U is occupied by distribution-fortified incumbents only; new entrants without distribution prior cannot enter top of U via AI-as-product alone). H3 sharpened (H5 names the operator-vs-seller axis as one of H3’s missing axes). H4 complemented (H4 is the seller-side, H5 is the operator-side; the two are siblings, not competitors — H4 vendors are H5 operators’ suppliers in the cleanest reading).

[2026-05-26] edit | Lint fixes from 2026-05-25 lint report

  • Trigger: apply all actionable fixes from the 2026-05-25 lint report. Two parallel subagents were dispatched but both were blocked by permission policy; all edits applied from the main session.
  • Tier 1 fixes:
  • Tier 2 fixes:
  • Tier 2 deferred: T2-B (page splits for H4, weber-taxonomy-2026-gaps, H2 labor-side mirror) — author elected to skip splits this pass; revisit in a dedicated /edit session. T2-D (GTM taxonomy / Strand B / gtm-business-archetypes) — awaits /discover pass for Tunguz / Janz / Bessemer-style sources before instantiation.
  • Index: added two one-line entries in the Concepts catalog for the new pages; refreshed last-edit header.
  • What this does not do: Two of the four bands of H2’s new Confidence section reference slugs the lint draft assumed existed but do not (cursor-revenue-trajectory, bloc-divergence). The wording was substituted but the underlying claim — that there is concrete revenue evidence at the top of the U and an explicit bloc-divergence concept page — would be sharper with those pages eventually materialised.

[2026-05-25] lint | Full wiki health check (130 content pages)

  • T1 findings: 3 issues — (1) broken wikilink [[exclusive-industry-data-partnerships]] in [[ai-charged-product-service-provider]] line 90; (2) log chronological ordering broken (2026-05-25 Gültekin/gaps entries at top, Westerveld/Protschky entries at bottom — opposite conventions); (3) H4 “Referenced by” block empty despite 26 inbound pages; plus 7 missing reverse links from Gültekin ingest to concept pages that link back to the source.
  • T2 findings: 4 — H2 missing ## Confidence body block; [[weber-taxonomy-2026-gaps]] (3,777w) and [[H4_rl-specialization-value-pocket]] (4,473w) ripe for splitting; “foundation-model-layer” concept missing despite 18-page mention frequency; GTM taxonomy (Strand B) still not created.
  • T3 findings: 3 — no orphans; 7 hubs ≥15 inbound (H2 at 79 is largest); H4 last-updated lags 2026-05-25 patch; H3 blocks absent by design (deferred stub).

[2026-05-25] ingest | Protschky et al. (2025) — What Gets Measured Gets Improved: Monitoring ML Applications in Their Production Environments

  • Source: raw/What_Gets_Measured_Gets_Improved_Monitoring_Machine_Learning_Applications_in_Their_Production_Environments.pdf (IEEE Access 13, 34518–34530; CC BY 4.0; 13 pages).
  • Per author instruction, ingested primarily into the tech subfolder; cross-linked into concepts and hypotheses where load-bearing.
  • New pages: ml-monitoring (tech anchor), ml-production-environment-characteristics (concept — C1–C5), ml-monitoring-quality-cycle (concept — 17 practices), ml-drift-types (concept — concept/data/virtual/adversarial), protschky-ml-monitoring-2025 (source).
  • Patches: eval-real-world-gap (C2 = operational decomposition), continual-learning-paradigm (Practice 16 = operational instance), H4_rl-specialization-value-pocket (monitoring as operational complement — strongly supportive), rl-testing-validation (production-time counterpart), vertical-ai-orchestration (monitoring as operational stickiness substrate), ai-development-facilitator (monitoring tooling as candidate surviving sub-pattern in squeezed middle — 15 tools reviewed, none complete), scaling-gap (operational root-cause reading), weber-taxonomy-2026-gaps (candidate 8th gap dimension), README (catalog).
  • Key takeaway: ML monitoring is the operational layer that closes the continual-learning-paradigm loop and makes H4_rl-specialization-value-pocket / vertical-ai-orchestration possible in production. Five production-environment characteristics (C1–C5) + 17 practices on a define / measure / assess / act / control cycle. The paper’s review of 15 monitoring tools found none implement the full surface — an empirical wedge under Weber’s “squeezed middle” thesis.
  • Hypothesis touches: H4 strongly supportive (supplies the operational vocabulary the H4 vendor’s product spec needs). H3 modestly supportive (monitoring sophistication is a 2026-era orthogonal axis Weber’s 2021 instrument does not capture — candidate 8th dimension on weber-taxonomy-2026-gaps). H2 modestly informative (operational anchoring for why the middle of the U is harder than tooling-vendor optimism suggests; one possible exception in monitoring-specific tooling). H1 neutral (orthogonal to substrate-ladder claim).
  • Contradictions / caveats surfaced: paper is normative and tool-light — 15 tools reviewed without naming them in main text (T01–T15 in Appendix C); claims under-served tooling space without firm market-share data. Single multivocal-DB (Medium) is a documented bias. Interview sample skews to data scientists / ML engineers (no procurement, no operations executives) — buyer-side perspective absent.
  • Open questions surfaced: paper’s own Table 4 (basic ML monitoring metrics; drift-type measurement; decision-support architecture for ML tradeoffs; non-ML-stakeholder communication; SE-process integration; roles / skills). Wiki-internal: does a monitoring-tool vendor implementing the full 17-practice surface escape Weber’s squeezed-middle fate, or is the only viable form factor embedded inside a vertical AI orchestrator? Is monitoring-stack under-investment the under-named operational root cause of a non-trivial share of the scaling-gap?
  • Source moved to raw/ingested/.
  • Next: /lint to verify reverse-link integrity; consider promoting the wiki-internal open questions into 05_open-questions; the paper’s “monitoring sophistication” dimension is a clean candidate for the 8th entry on weber-taxonomy-2026-gaps if the author wants to elevate it from candidate to confirmed.

[2026-05-25] ingest | Westerveld et al. (2023) — The Business Model Portfolio as a Strategic Tool

  • Source: raw/The-business-model-portfolio-as-a-strategic-tool_2023_The-Journal-of-Strategy recap.md (structured recap, 63 lines).
  • Pages touched: business-model-portfolio (new concept), westerveld-business-model-portfolio-2023 (new source), weber-taxonomy-2026-gaps (patched — new “Prerequisite — the unit of analysis” section near the top + See-also entries), ai-startup-business-archetypes-weber (patched — “Unit of analysis” note + See-also), weber-ai-startup-business-models (patched — See-also), H3_orthogonal-axes-under-priced (patched — “Structural reinforcement from the BMP frame” subsection), karpathy-software-3 (patched — reverse link: Huang ↔ Complementing, Steinberger ↔ Cannibalization), eval-real-world-gap (patched — reverse link: Sutskever’s “human reward hacking” ↔ Dogmatic paradox), digital-core (patched — reverse link: functional architecture as connective tissue).
  • Key takeaway: the Business Model Portfolio (BMP) frame supplies the prerequisite scoping rule for the Weber classifier — the BM, not the firm, is the unit of strategic analysis. A startup running multiple BMs is classified per-BM, not forced into one Weber pattern. Three strategic intents (Diversifying / Sensing / Complementing) and four permanent paradoxes (Horizon / Cannibalization / Digital / Dogmatic) now in the wiki as diagnostic vocabulary.
  • Hypothesis touches: H3 structurally reinforced — BMP is the structural reason a single firm occupies multiple positions on every orthogonal axis, sharpening H3 as “best plays at axis intersections, evaluated per BM.” H2 modestly clarified — a single firm can occupy multiple U positions (one BM at top, another at shoulder). H1, H4 neutral.
  • Contradictions / caveats surfaced: Westerveld’s 38 cases are airlines / media / retail — applying BMP to AI startups is an extrapolation this wiki is making, not a finding of the source. Original literature window closes 2021; the AI-era bridges (Karpathy, Sutskever, Steinberger, Huang, WEF, Deloitte) are recap-author interpretive overlays, not paper findings.
  • Open questions surfaced: Which of the four paradoxes is most salient for AI startups specifically? Does the Complementing intent dominate 2026 AI, or is the mix still Sensing-heavy? Does the BMP frame transfer cleanly to small / early-stage firms?
  • Source moved to raw/ingested/.
  • Next: /lint to check link integrity across the BMP additions; consider promoting BMP-induced open questions into 05_open-questions; treat the BMP frame as the default scoping rule on next AI-startup ingest (declare weber-pattern: per BM in source frontmatter).

[2026-05-25] ingest (partial) | Gültekin et al. (2025) — Commercialisation of AI in Türkiye

  • Source: Gültekin, D. G., Pinarbasi, F., Yazici, M., & Adiguzel, Z. (2025). Commercialisation of artificial intelligence: a research on entrepreneurial companies with challenges and opportunities. Business Process Management Journal, 31(2), 605–630. DOI: 10.1108/BPMJ-10-2023-0836. Located in raw/. PDF → marker → 500-line markdown.
  • Author instruction: “we don’t need full paper ingestion, but mainly concepts related to 04_concepts/weber-taxonomy-2026-gaps.md”. Honored — only material relevant to the six gaps was promoted.
  • What this paper is: qualitative interview study of 10 Turkish AI-as-core-product startups (Istanbul / Antalya technoparks, Nov 2022 — effectively pre-ChatGPT operationally). Sample: 5 computer vision, 4 predictive data analytics, 1 language processing. Adopts Weber’s 4-pattern classifier as-is. 9/10 firms classified as ai-charged-product-service-provider or ai-data-analytics-provider; 1 as ai-development-facilitator; zero ai-deep-tech-researcher.
  • Pages touched:
    • New: wiki/05_sources/gultekin-pinarbasi-commercialisation-ai-2025.md — partial source page (status: ingested-partial). Explicit note that the paper’s main thrust (Turkish ecosystem, BAH vs Claessens process models) was not materialised.
    • Patched: wiki/04_concepts/weber-taxonomy-2026-gaps.md — new “First independent evidence” section with per-gap table; new See-also entry.
    • Patched: wiki/05_sources/weber-ai-startup-business-models.md — new “Independent adoption” subsection + See-also entry.
    • Patched: wiki/04_concepts/ai-startup-business-archetypes-weber.md — See-also entry only.
    • Index: new source line; refreshed last-edit header.
  • First pass net effect: weak / narrow / asymmetry-confirming support for gaps 1, 4, 6; silent on 2, 3, 5. Author flagged that the first pass under-extracted the “commercialisation approach” material, which is in fact the paper’s most decisive contribution.
  • Revised net effect on gaps diagnostic:
    • Added gap 7 — Commercialisation motion / GTM posture — empirically anchored in §5 of the paper (BAH 1982 linear / Claessens 2015 marketing-led / fuzzy emergent modal in practice) and §5.1 (internal-vs-external resource spectrum). This is the first gap-page dimension derived from an ingested source rather than from wiki-internal triangulation. Maps to the “acquisition motion” dimension in the Strand B planning notes.
    • Split original gap 4 into pure channel geometry (now narrower) and process motion (new gap 7). The empirical signal (LP1 self-serve > 100k customers vs nine B2B firms < 50 each) cannot separate the channel effect from the motion effect on n=10, but it directionally aligns with distribution-moat.
    • Strong support for new gap 7; weak / narrow / asymmetry-confirming support for gaps 1, 6; gaps 2, 3, 5 unaddressed (pre-ChatGPT sample). No dimension retired. Agentic-workflow fifth archetype not validated.
  • Meta-finding worth flagging: the paper’s most useful contribution is methodological — Weber’s morphological method on Crunchbase operating-status snapshots cannot encode process dimensions; only qualitative-interview methods can. This is a structural reason to expect Strand A (a Nickerson-method follow-up taxonomy) to still miss gap 7 unless the researchers explicitly ask founders about commercialisation process. Worth carrying into Strand B source selection (Tunguz / Janz are interview- and observation-based, so they will capture motion natively).
  • Contradictions surfaced: none. The asymmetry “ethics as macro-concern / not product surface” is independently confirmed across two different geographic samples (German-led 2021 / Turkish 2022). Weber-implicit reading “AI capability = commercial success” is directionally contradicted by the LP1 vs nine-B2B pattern — motion may dominate capability.

[2026-05-25] edit | Weber-taxonomy-2026-gaps diagnostic page + source-hunting plan

  • Trigger: author goal — assess whether Weber et al.’s 4-category / 11-dimension / 39-characteristic AI-startup BM taxonomy is complete in 2026, identify blanks, and lay the ground for a parallel DISTRIBUTION / GTM morphological box. No new ingest in this pass; built entirely from existing wiki material.
  • New concept page: wiki/04_concepts/weber-taxonomy-2026-gaps.md — status working-draft (diagnostic; awaiting source-side fill from a 2024–2026 taxonomy update). Six missing dimensions documented (model sourcing · autonomy level · orchestration depth · distribution architecture · data flywheel · ethics/safety/compliance-as-product), each grounded in existing wiki evidence; plus the fifth archetype sketch (agentic-workflow providers — Sierra / Cognition / LangGraph / vertical orchestrators). Closes with a two-strand “Next move” plan: Strand A — ingest one post-2023 academic taxonomy update (BISE / ECIS / ICIS / HICSS 2023–2026, Nickerson-method follow-ups, or Bessemer / Menlo / Sapphire practitioner taxonomies); Strand B — ingest three GTM sources (Tomasz Tunguz outcome-pricing essays, Christoph Janz / Point Nine “5 ways” updated for AI, one of Bessemer 2025 / a16z service-as-software / Sapphire Service-as-Software / Foundation Capital outcome-pricing) and produce wiki/04_concepts/gtm-business-archetypes.md + outputs/gtm-taxonomy.html sibling to the existing Weber HTML.
  • Reverse-link patches (forward → reverse rule):
    • wiki/04_concepts/ai-startup-business-archetypes-weber.md — added paragraph in “2021 → 2026 vintage shift” section + entry in “See also” pointing to the new gaps page.
    • wiki/05_sources/weber-ai-startup-business-models.md — added paragraph in “2021 → 2026 vintage caveats” + entry in “See also”.
    • wiki/04_concepts/distribution-moat.md — added entry in “Related” naming this concept as the project’s central reference for the GTM sibling taxonomy (Strand B).
    • wiki/02_hypothesis/H3_orthogonal-axes-under-priced.md — added paragraph after the Weber activation-evidence section noting the six additional candidate orthogonal axes in the gaps page.
  • Index: added one-line entry in the Concepts catalog; refreshed last-edit header with 2026-05-25 summary; bumped frontmatter last-updated.
  • What this does not do: the gaps page is diagnostic, not authoritative. Each dimension is candidate-status until a 2024–2026 source has been ingested. The GTM sibling morphological box has not been built; it is sketched only in the planning notes inside the gaps page’s “Next move” section. The author hunts and ingests Strand A and Strand B sources — agent will not pull them.

[2026-05-24] edit | Site welcome callout + graph excludes index/log

  • wiki/index.md — added > [!abstract] callout (research one-liner + author: student/operator, Veneto).
  • quartz-config/quartz.layout.ts — new; excludeSlugs: ["index", "log"] on local + global graph; breadcrumbs hidden on home; footer → repo.
  • quartz-config/components/scripts/graph.inline.ts — patched upstream graph script to honour excludeSlugs.
  • .github/workflows/deploy.yml — copies layout + graph script on build.

[2026-05-24] edit | Glossary — inference and taxonomy entries

[2026-05-24] edit | New hypothesis H4 + tech/ reference subfolder (RL-driven specialization)

  • Trigger: author-proposed working hypothesis — pre-training has saturated, RL (especially rl-from-verifiable-rewards) is the still-scaling axis, and the entrepreneurial opportunity is domain-specific RL fine-tuning of base models sold as a proprietary operational asset to a single industrial customer or narrow vertical. Worked example: industrial-implant intervention-proposal model trained on a company’s own (state, proposal, outcome) history. Technical premise grounded; business case explicitly under-validated.
  • New hypothesis page: wiki/02_hypothesis/H4_rl-specialization-value-pocket.md — status working-hypothesis, confidence low-to-medium (technically grounded above; business case unvalidated). Sections cover claim, technical premise, business premise, the worked industrial-implant example, six bear arguments, four falsifiability conditions, six open research questions.
  • New subfolder wiki/tech/ introduced as a NEW page kind: short didactic technical-reference pages (200–500 words each) grounding the analytical layer in algorithmic detail. Seven content pages + a README:
    • tech/README.md — catalog + framing.
    • tech/rlhf.md — Reinforcement Learning from Human Feedback (InstructGPT recipe; SFT → RM → PPO/DPO).
    • tech/rlaif.md — Reinforcement Learning from AI Feedback (Constitutional AI; judge-model labelers).
    • tech/rl-from-verifiable-rewards.md — RLVR (o1/R1/AlphaProof recipe; the technical engine under H4).
    • tech/rl-data-preparation.md — trajectory collection, dedup, contamination filtering, reward shaping.
    • tech/rl-testing-validation.md — held-out evals, reward-hacking probes, real-world A/B; closes (or fails to close) eval-real-world-gap.
    • tech/rl-open-vs-closed-source.md — algorithms commoditised, recipes + data + value-function research stay closed; mirrors open-weight-asymmetry one floor down.
    • tech/rl-apis.md — frontier-lab tuning (OpenAI RFT, Anthropic custom, Vertex) vs open-weight tuning (Together, Fireworks, Predibase, Modal/Replicate/RunPod, Lamini, Anyscale); the most consequential vendor unit-economics choice in H4.
  • Patches to existing pages (forward + reverse link integrity):
    • wiki/index.md — H4 added to Hypotheses catalog; new ## Tech section listing the 7 pages.
    • wiki/01_paper-planning/05_open-questions.md — new “RL-driven specialization as value pocket” section with six open questions (break-even contract size, verifiable-reward feasibility across verticals, open-weight vs closed-frontier base choice, boutique vendor vs frontier-lab fine-tuning API, customer data exclusivity, position on H1 ladder).
    • wiki/02_hypothesis/H1_L0-L7-ladder.md — H4 added to Related (L3–L5 substrate, Maker posture at vertical scale).
    • wiki/02_hypothesis/H2_u-curve-of-value.md — H4 added to Related (vertical-scale top-of-U via RL specialization).
    • wiki/04_concepts/post-scaling-research-pivot.md — H4 added to Related (load-bearing premise).
    • wiki/04_concepts/value-functions-as-algorithmic-emotion.md — H4 + four tech pages added (value-function tractability is the gating constraint).
    • wiki/04_concepts/agentic-scaling-law.md — H4 + rl-from-verifiable-rewards added.
    • wiki/04_concepts/continual-learning-paradigm.md — H4 + three tech pages added (continual learning is the single-customer compounding moat).
    • wiki/04_concepts/scaling-wall.md — H4 added (entrepreneurial wager on the wall + bear pressure if wall generalises to RL).
  • Status note. H4 is the first hypothesis page with explicit technical-layer scaffolding underneath it. Confidence is asymmetric: higher that RL specialization will matter more in the next 18 months, much lower that boutique RL-fine-tuning vendors capture the surplus (vs frontier-lab tuning APIs or the customers themselves). The tech/ subfolder is intentionally light — sufficient depth for an operator to weigh the hypothesis above, not encyclopaedic. Future /lint sweep will populate the Referenced by blocks across the new pages.

[2026-05-23] edit | Quartz publish filter + baseUrl fix

  • Problem: Live site at https://supgrade.github.io/where-ai-value-lands/ showed only empty RSS XML — no HTML home page.
  • Cause: ExplicitPublish filtered all 111 wiki files (none had publish: true); deploy artifact had index.xml but no index.html.
  • Fix: Removed ExplicitPublish; use RemoveDrafts only. Corrected baseUrl to supgrade.github.io/where-ai-value-lands (matches GitHub repo slug). Local build now emits 303 files including index.html.
  • Files: quartz-config/quartz.config.ts, quartz-config/README.md, README.md

[2026-05-23] edit | Collaboration infrastructure (T0 + T1) shipped — repo go-public-ready

  • Decisions resolved (six of the open questions promoted from 06_collaboration):
    • License: MIT (code) + CC-BY-4.0 (content). Two separate files at repo root: LICENSE, LICENSE-content. Maximally compatible with T1–T4.
    • Repo: github.com/Supgrade/where-AI-value-lands — personal account, repo name includes “AI” to surface the subject at URL level.
    • Domain: GitHub Pages default (supgrade.github.io/where-AI-value-lands); custom domain deferred until T0 produces signal.
    • Static-site tool: Quartz v4, default theme. Cloned fresh by CI on every build (kept out of repo via .gitignore); our customizations live in quartz-config/.
    • /contribute skill scope at v1: all four contribution kinds (source, daily thought, concept note, open question). Implemented in .claude/skills/contribute/SKILL.md.
    • Contributor charter: terse + filtering posture; CONTRIBUTORS.md + per-page footer credit model. Lives at 08_contributor-charter.
  • Files created at repo root (public-facing entry surface):
    • README.md — rewritten for public consumption; project frame, audience, repo layout, skill table, license, contributing pointer
    • CONTRIBUTING.md — short pointer doc into the charter and the /contribute skill; out-of-scope list; license-acceptance clause
    • CONTRIBUTORS.md — seed roll with author listed; one-line entry format documented
    • LICENSE — MIT, copyright Gabriele Dalla Costa 2026, with scope clarification at the bottom
    • LICENSE-content — CC-BY-4.0 short-header form, with recommended attribution string
    • .gitignore — excludes quartz/, node_modules/, public/, .quartz-cache/, OS/IDE noise; preserves .obsidian/
  • Files created in .github/:
    • .github/workflows/deploy.yml — Node 22, clones Quartz v4, copies quartz-config/quartz.config.ts over the install, builds with -d ../wiki, uploads public/ artifact, deploys via actions/deploy-pages@v4; pinned modern Pages-via-Actions pattern (not deploy-from-branch)
  • Files created in quartz-config/:
    • quartz.config.ts — Quartz v4 config: pageTitle “Where AI Value Lands”, baseUrl supgrade.github.io/where-AI-value-lands, default plugin set, Source Serif 4 / Source Sans 3 / JetBrains Mono typography (matches /visualize register), analytics off, ignorePatterns: [private, templates, .obsidian, _staging]
    • README.md — explains the Quartz-as-build-tool pattern (clone-on-build, not committed), local-preview steps, required GitHub repo settings (Settings → Pages → Source = GitHub Actions), customisation notes
  • Files created in .claude/skills/contribute/:
    • SKILL.md — full v1 skill spec implementing 06_collaboration §T1: pre-flight check (verify fork, branch, gh auth), Step 1 scope check (reads 05_open-questions + index head), Step 2 contribution-kind menu (4 kinds), Step 3 per-kind workflows with file paths and frontmatter shapes, Step 4 branch + PR description template (dual Pages affected Obsidian wikilinks / Page links GitHub relative links), Step 5 forward → reverse link audit (no silent edits to existing pages; choice surfaced to contributor), Step 6 print-the-gh-command (skill does not push), explicit out-of-scope list, what-not-to-do
  • Files created in wiki/01_paper-planning/:
    • 08_contributor-charter.md — high-bar standard; terse + filtering; explicit out-of-scope (consumer AI, robotics, ML research, general AI safety, surface summaries); evidentiary standards (every claim sourced, quantitative claims have data pages, secondary syntheses flagged, bear and bull frames engaged); voice calibration; per-page contributor credit format; “friction is the feature” stance
  • Files updated:
  • Key takeaway: T0 (public repo + static site infrastructure) and T1 (/contribute skill) are now mechanically complete. The repo is ready to push to GitHub once the author runs git init + gh repo create. Quartz deploys on first push to main via Actions. The charter is status: draft and will mutate as real contributions accumulate — by design.
  • Recommended next:
    1. Author: git init, then gh repo create Supgrade/where-AI-value-lands --public --source=. --remote=origin, then push to main.
    2. Configure repo on GitHub: Settings → Pages → Source = GitHub Actions; verify the workflow run completes; check the live site at https://supgrade.github.io/where-AI-value-lands/.
    3. Optional T2 (auto-digest newsletter) and T3 (LinkedIn) remain deferred per 06_collaboration sequence.
    4. /lint to verify reverse-link integrity of the new 08_contributor-charter page against 06_collaboration and 05_open-questions.

[2026-05-22] edit | Create wiki/data/ — citable-evidence layer

[2026-05-22] edit | Add 06_collaboration to paper-planning

  • New page: 06_collaboration — brainstorms how others contribute to the wiki while it is in flight. Distinct from 04_distribution (which is about the finished paper).
  • Source for the page: today’s daily thought (raw/daily tougths/2026-05-22.md), which articulated the PR-via-/contribute-skill contribution model, the operator-only friction tolerance, and the long-term “tool for founders to share by prompting” framing.
  • Structure: five additive tiers (T0 read-only repo → T1 PRs + /contribute skill → T2 auto-digest newsletter → T3 community channel (LinkedIn / X / Discord, pick one) → T4 chat-with-the-wiki web app), plus a recommended ship sequence, a risks section, and a how-it-fits-with-existing-skills table.
  • Stance taken (user-directed): contributions are technical-by-default (CLI + PR), author is the single merge gatekeeper, voice coherence beats contribution volume — “reject more than you accept, especially early.”
  • Reverse links added: 04_distribution (Related section), 05_open-questions (new “Collaboration model” block listing four decisions promoted up: /contribute skill scope, newsletter platform, contributor charter, community channel choice), 00_initial-brief (added to the plan map).
  • Index updated: new entry under “Paper planning”.
  • Suggested next: /edit to author the contributor charter referenced in 05_open-questions; spec out the /contribute skill before publishing the repo (currently a placeholder, not a real skill); decide repo license — T1/T2 work in MIT or CC-BY, T4 is complicated by CC-BY-NC.

[2026-05-22] reflect | Business archetypes, analytical vocabulary, personal research frame

  • Daily note: 2026-05-22
  • Annotated: 2026-05-22 annotated
  • Memory marker: archetypes-vocabulary-personal-frame
  • Wiki changes:
    • 07_analytical-vocabulary (created) — six analytical axes, vocabulary quick-reference table, sample coordinates for Cursor and a North Italian SME; canonical source for “layer” vs “level” vs “archetype” vs “tier”
    • 06_collaboration (updated) — T0 expanded with repo structure (full folder tree) and static site spec (Quartz recommended, GitHub Action deployment); T1 expanded with full /contribute skill specification (scope check, 4 contribution kinds, branch + PR description template, forward→reverse link check)
    • 05_open-questions (updated) — three new sections: collaboration model additions (static site domain, Quartz vs. alternatives), analytical vocabulary decisions (5 open Qs on terminology and axis scope), business archetype taxonomy decisions (completeness, SME/Italian archetype gap, archetype-to-U-curve mapping)
    • index.md (updated) — 07_analytical-vocabulary entry added; thoughts section initialized
  • Kept as conversation: personal life-design framing (what to start, where, what kind of life) — live material for Section 6 drafting; U-curve mapping of archetypes (open question, not yet a wiki page)

[2026-05-22] discover | SME & mid-cap ER&I AI adoption

  • Mode: both (wiki-driven + news-driven)
  • Candidates surfaced: 8
  • Gaps targeted: enterprise-adoption-ladder (no SME phase anchors), scaling-gap (SME shape unknown), H2_u-curve-of-value (tested against large firms only), divergent-value-stack-optima (EU firm-level data absent), foundational-enablers (universalist claim untested for SMEs)
  • Top picks: IEA Energy and AI 2025 (energy-sector primary source), OECD AI Adoption by SMEs Dec 2025 (76% are “AI novices”), Eurostat 2025 sector×size breakdown (EU energy AI = ICT security, not operations), PIIE 2026 industrial sectors (manufacturing AI outside production), Fed monitoring note April 2026 (14:1 investment gap), St. Louis Fed EU/US gap Mar 2026 (employer encouragement drives 95% of gap), TwinLadder Mittelstand (94% not ready), McKinsey State of AI Nov 2025 (6% high performers)

[2026-05-22] visualize | enterprise-adoption-ladder (ER&I reading)

  • Output: outputs/enterprise-adoption-ladder.html
  • Concept resolved: enterprise-adoption-ladder, read through the ER&I lens (user asked for “timeline of AI adoption + value creation for an ER&I company, plus small/medium vs big differences”).
  • User scope choice: visualize the existing concept ER&I-flavored, not invent a new page. The small-vs-big-firm dimension is flagged as a wiki gap in the output’s closing tension — every current ER&I case study (BMW, Aker BP, SBB, Merck, Siemens, Chevron) is a large firm; SME industrial AI adoption is not represented in the wiki yet.
  • Nodes: 25 (1 center + 5 phase markers + 8 concepts + 2 hypotheses + 6 ER&I use cases + 3 sources).
  • Edges: 42 (33 wikilinks + 9 citations).
  • Anchors: P1 generic-utility-pilot (no concrete case in wiki); P2 disconnected-pilots (most ER&I firms here); P3 BMW supply-chain digital twin (30–40% productivity claim); P4 Aker BP Yggdrasil (periodically unmanned offshore); P5 no public exemplar.
  • Design: Tufte-style two-column layout, Source Serif 4 body, JetBrains Mono for marginalia and chrome, ink-blue accent (#1f3a5f) on off-white (#f4f1ea), d3-force graph clustered along the phase axis (left→right), scroll-triggered reveals, no emoji / no cards / no gradients.
  • Size: ~40KB, self-contained except for Google Fonts + d3 CDN.
  • Suggested next: /discover on SME and mid-cap ER&I AI adoption surveys (EU SME digital-maturity, IEA small-utility AI) before re-running /visualize with a firm-size dimension folded in.

[2026-05-22] edit | Materialize ER&I use cases into 06_ideas/ folder

[2026-05-22] ingest | “Digital Core” and the Structural Requirements for Non-Tech Enterprises (synthesis note)

  • Pages created: non-tech-digital-core-synthesis (source), digital-core, taker-shaper-maker, fusion-skills
  • Pages updated: foundational-enablers (digital-core now a wikilink + Related expanded), wef-ai-in-action-2025 (Frameworks section: digital-core / taker-shaper-maker / fusion-skills now wikilinks + Related expanded), enterprise-adoption-ladder (Related expanded), autonomy-slider (Related expanded: fusion-skills as labor-side counterpart), democratization-of-programming (Related expanded), karpathy-software-3 (Related expanded), H1_L0-L7-ladder (Tensions: third orthogonal axis taker-shaper-maker added; Related expanded), index (4 new entries)
  • Key takeaway: this source is a secondary LLM-generated synthesis of material already in the wiki (wef-ai-in-action-2025 + karpathy-software-3 + Huang). Its real contribution is naming: it elevates three handles that were latent bullet items in the primary sources — digital-core (3-layer enterprise stack), taker-shaper-maker (strategic-adoption typology), and fusion-skills (workforce capability) — into addressable concepts. With these, the wiki now has the full three-axis coordinate system for non-tech firms: (substrate L0–L7, maturity Phase 1–5, positioning Taker/Shaper/Maker).
  • Contradictions surfaced:
    • Methodologically thin. The source claims “PhD-level analysis” but provides no anchored citations. The “GANs for synthetic data” claim is two cycles behind the 2026 frontier (LLM-distilled / diffusion-generated synthetic data). Flagged on the source page.
    • No bear-case engagement. The source repeats the bull-managerial frame (scaling-gap is a readiness problem) without acknowledging zitron-circular-economics, task-based-framework, or marcus-world-models-failure exist.
    • The “digital core” frame silently locks the enterprise into the L0–L2 / renter position. By treating L3 infrastructure as “procurement,” it collapses substrate-ownership into a non-question — exactly the move H1_L0-L7-ladder is meant to make visible. Documented on digital-core.
    • Fusion-skills is a managerial answer to capital-labor divergence, not a refutation. The framing implies the AI-augmented worker captures surplus; the structural critique (capital-labor-divergence) argues the vendor of the AI tool does. Flagged on fusion-skills.
  • Suggested next: this source is meta-evidence of how the bull-managerial frame propagates into non-tech firms via LLM-generated executive digests — worth treating as a discourse artifact in the paper, not just a content source. Possible follow-ups: /lint to verify reverse-link integrity across the four new pages; /ask to materialize the three-axis (substrate / maturity / positioning) coordinate map as a draft section; /discover to find the primary Huang GTC keynote material the source quotes but does not cite.

[2026-05-22] ingest | Deloitte AI Dossier — AI Use Cases in Energy, Resources & Industrials

  • Pages created: deloitte-ai-dossier-eri, agentic-revolution, vertical-ai-orchestration, synthetic-data-generation, ai-factory-huang
  • Pages updated: H2_u-curve-of-value (vertical-domain corroboration block + 4 new Related links), H1_L0-L7-ladder (Related: ER&I use cases sit at L3–L5 in industrial verticals), autonomy-slider (vertical-industrial restatement of the Iron Man Suit imperative; liability allocation as the constraint), middle-layer-defensibility (Non-software instantiation block; 3 new Related links), scaling-wall (Vertical-application manifestation block: OOD generalization failure in ER&I; 3 new Related), karpathy-software-3 (4 new Related links), wef-ai-in-action-2025 (4 new Related links), index (5 new entries)
  • Key takeaway: this is the first ingested source with a deep vertical-domain decomposition. Deloitte’s central strategic claim — “the competitive moat in ER&I will no longer be determined solely by physical assets, but by the maturity of an organization’s digital core, its data ecosystems, and its ability to securely manage hybrid human-AI workforces” — places the moat at orchestration + data + workforce coordination, not at the AI model. This is a clean non-software instantiation of middle-layer-defensibility and supplies the first vertical-domain corroboration of H2_u-curve-of-value in industrials: top = multi-agent operational orchestrator, bottom = proprietary operational data + physical assets, squeezed middle = generic foundation models + generic agent frameworks. The source also introduces agentic-revolution (analytical → agentic transition) as the load-bearing paradigm shift the paper must address.
  • Contradictions surfaced: (1) OOD failure is self-aware bear evidence inside a bull source — Deloitte’s candid admission that models fail on “novel ores” / unfamiliar geology lands closer to scaling-wall than its own conclusion does. (2) Magnitude is unquantified — unlike WEF ($7.6–17.9T) and task-based-framework (<0.71% TFP cap), Deloitte does not commit to a sector-wide number. Harder to test, harder to falsify. (3) Agentic full-stack vs. static task-set — the agentic-revolution narrative implicitly argues the automatable task-set expands as orchestration improves, directly rebutting Acemoglu’s static-task framing; the source does not draw the contradiction out. (4) Bear cluster absent — Deloitte does not engage with circular-ai-economy or task-based-framework at all. (5) Federated AI-factory topology opens room for vertical-incumbent or sovereign substrate strategies that bypass hyperscaler rental — consistent with sovereign-ai but pulls a different lever than divergent-value-stack-optima.
  • Suggested next: ingest the primary Deloitte AI Dossier PDF directly to verify the synthesis (this file is LLM-authored); ingest a Huang GTC keynote for the AI-factory framing as a primary source; /lint to check reverse-link integrity across the 5 new pages and the heavy update fan-out; /ask to draft a per-layer value-capture summary that integrates the vertical-industrial frame with the horizontal-SaaS frame (where-value-lands-2026) — the wiki now has both and they should be reconciled.

[2026-05-22] ingest | WEF — AI in Action: Beyond Experimentation to Transform Industry (2025)

[2026-05-22] ingest | Where Value Lands: AI Redistribution of Economic Surplus (2026)

  • Pages created: where-value-lands-2026, middle-layer-defensibility, distribution-moat, context-control, world-models-jepa, jevons-paradox-ai
  • Pages updated: H2_u-curve-of-value (status lifted to low-to-medium; empirical validation section added for the top of the U; Jevons Paradox nuance added for the bottom; Related section expanded), index (6 new entries)
  • Key takeaway: the strongest empirical validation yet for H2’s top — Cursor (2B ARR) and Windsurf ($2.4B acquisition battle) prove distribution moats are real and measurable. But the taxonomy matters: what the source calls “middle stack” is what H2 calls “the top” (distribution + workflow ownership). The generic thin wrapper still dies; the sophisticated workflow-embedded platform survives. LeCun’s JEPA adds a long-horizon paradigm-reset risk for current LLM-based moats. Jevons Paradox partially defends the bottom of the U against the circular-economy critique.
  • Contradictions surfaced: (1) Jevons Paradox vs. circular-ai-economy — efficiency unlocks new demand, but is that demand real or VC-subsidized? (2) LeCun’s JEPA threat is largely absent from the bull-frame practitioners’ worldview; neither Truell nor Srinivas addresses the architectural ceiling risk their moats face in a post-LLM paradigm. (3) EU and China stacks absent from this source — bull frame is implicitly US-centric.
  • Suggested next: /ask to draft a consolidated per-layer value-capture summary that integrates bull and bear frames across this source, bear-case-synthesis, and karpathy-software-3; or /discover to find primary sources for Cursor/Windsurf (Lenny’s Podcast transcripts) and LeCun (Davos talk / JEPA papers) to verify synthesis claims.

[2026-05-21] ingest | The Geopolitics and Governance of the Global AI Divide

[2026-05-21] ingest | AI Redistribution of Economic Surplus — The Skeptics and Bear Case

[2026-05-21] ingest | Karpathy — Software Is Changing (Again)

  • Pages created: karpathy-software-3, software-3-paradigm, autonomy-slider, llm-as-operating-system, agentic-scaling-law, democratization-of-programming
  • Pages updated: H1_L0-L7-ladder (competing-taxonomy + agentic-scaling counter to scaling-wall added; confidence note refined), H2_u-curve-of-value (bull-frame pressure points added — install-base reinforces top, AI-factory reinforces bottom, agent-eliminates-apps threatens GUI top), scaling-wall (added direct counter via agentic-scaling-law), index (5 new concept pages + 1 new source).
  • Key takeaway: this is the first strong bull-frame paradigm document in the wiki. It introduces a competing single-axis taxonomy (autonomy-slider) to H1_L0-L7-ladder, partially restores the bottom of H2_u-curve-of-value via inference-as-economy, and offers the strongest direct counter to scaling-wall via agentic-scaling-law (test-time compute as second scaling axis).
  • Contradictions surfaced:
    • Scaling wall vs. agentic scaling law — base-model reasoning plateau (Marcus) vs. test-time compute as second axis (Huang). Both can be partially right; the synthesis is task-distributional.
    • GUI verification vs. GUI elimination — Karpathy’s “Iron Man suit” assumes verification GUIs are the durable design pattern; Steinberger predicts agents eliminate 80% of apps. Unresolved within the source.
    • Bull paradigm vs. circular economy — Karpathy treats LLMs as a utility (“intelligence on tap”) and is silent on whether the substrate paying for it is solvent (cf. circular-ai-economy).
    • OS metaphor vs. security reality — Steinberger’s OpenClaw experience suggests prompt injection is the new buffer overflow but the LLM-OS lacks the syscall ABI / permissions model that production OSes spent 40 years building.
  • Suggested next: ingest the original Karpathy talk and a Huang GTC keynote directly (this file is a secondary synthesis); ingest primary Steinberger / OpenClaw material to verify security claims; run /lint to verify reverse-link integrity across the five new concept pages; revisit 05_open-questions re: bear-case anchor in light of a now-present bull anchor.

[2026-05-22] ingest | OECD — AI Adoption by SMEs (G7 Discussion Paper 2025)

[2026-05-23] ingest | Choudary — Ecosystem Business Models: A Teardown

  • Pages created (4): choudary-ecosystem-teardown (source), ecosystem-business-archetypes (master concept), vertical-to-horizontal-unbundling (structural thesis), cross-archetype-confusion (strategic failure mode).
  • Pages updated (4): 07_analytical-vocabulary (Axis 3 sub-vocabulary table added — Choudary’s four archetypes as canonical decomposition of ecosystem position; Related expanded), H2_u-curve-of-value (Structural precursor section added — pre-AI horizontal-value thesis mapped onto top/middle/bottom of the U; Related expanded), middle-layer-defensibility (Disambiguating-the-middle section added — Integrator-as-switchboard vs failed-Capability; Related expanded), distribution-moat (Precursor section added — Aggregator archetype as canonical antecedent of distribution-moat reasoning; Related expanded).
  • Key takeaway: Choudary’s 2022 framework supplies the cleanest off-the-shelf sub-vocabulary for Axis 3 (Business Archetype) and a pre-AI structural precursor to H2_u-curve-of-value. The four archetypes (Aggregator / Integrator / Infrastructure / Capability) disambiguate what “middle” actually contains — the distinction between defensible Integrators (Stripe, Plaid) and failed Capabilities is the lens H2’s “squeezed middle” was missing. The distribution-moat AI firms (Cursor, Perplexity, Windsurf) are best read as Aggregators in B2B clothing, not as novel AI-era inventions. Choudary’s vertical-to-horizontal-unbundling mechanism locates H2 inside a recurring transaction-cost pattern across industries.
  • Contradictions / open questions surfaced:
    • Integrator-class value capture vs U-curve. Plaid (Integrator) and Stripe (Capability with Aggregator-like economics) exhibit top-of-U margins from middle-stack positions. Counterexamples to H2 or refinements?
    • Fifth archetype for the agentic era? vertical-ai-orchestration looks like a candidate Orchestrator archetype not covered by Choudary’s 2022 four-way taxonomy.
    • US-stack bias. Choudary’s framework does not cleanly accommodate open-weight-asymmetry or brussels-effect-and-mirage; needs cross-checking against divergent-value-stack-optima.
    • Should the paper adopt Choudary’s four as primary Axis-3 sub-vocabulary with the fifteen-archetype roster as elaboration? Promoted to 05_open-questions candidate.
  • Suggested next: /lint to verify reverse-link integrity across the 4 new pages and 4 updates; /edit to fold “Should Choudary’s four be the primary Axis-3 sub-vocabulary?” into 05_open-questions; /discover for a primary Stripe / Plaid case-study source to stress-test the bimodal-Integrator hypothesis; /ask to draft the per-archetype value-capture map (Aggregator → top of U; Infrastructure → bottom of U; Integrator → bimodal middle; Capability → IP-defended or failed).

[2026-05-23] ingest | Labor Market Impacts of AI — A New Measure and Early Evidence (Massenkoff & McCrory, Anthropic 2026)

[2026-05-23] lint | Full wiki health check (124 content pages)

  • T1 findings: 30 broken wikilinks (9× to never-created [[H3_orthogonal-axes-under-priced]]; case bug on [[OECD-sme-enabler-quartet]]; stub demo links in [[06_collaboration]] and wiki/data/README; skill-name links in [[08_contributor-charter]]; missing [[ai-capex-cycle]]); 459 missing reverse links (forward→reverse invariant decayed broadly — top hubs [[H2_u-curve-of-value]] (77 in), [[H1_L0-L7-ladder]] (37), [[wef-ai-in-action-2025]] (40) all under-linked back); 3 pages missing from wiki/index.md ([[inference]], [[suggested-sources]], [[The evolution of IT infrastructure]]); 2 genuine slug violations (README, The evolution of IT infrastructure in wiki/data/).
  • T2 findings: [[H1_L0-L7-ladder]] lacks explicit body Confidence block (only in frontmatter); 4 widely-used terms lack pages (value capture ×14, application layer ×23, displacement ×14, model layer ×6); 7 pages over 1500 words ripe for splitting, led by [[H2_u-curve-of-value]] at 2852 words.
  • T3 findings: 4 orphans (mostly legitimate planning/discover output); 31 hubs ≥15 incoming (healthy distribution); all pages carry last-updated 2026-05-21 to 2026-05-23 (no staleness); log chronological order intact.
  • Suggested next: /edit to resolve H3 ghost references and add reverse-links to top hubs; split [[H2_u-curve-of-value]]; create concept pages for value-capture, application-layer, displacement, ai-capex-cycle; ingest queued sources in raw/Clippings/ and the Massenkoff/McCrory-companion Anthropic PDF.

[2026-05-23] edit | Lint Tier 1 cleanup + hypothesis health + 4 concept pages + top-5 hub reverse-link sweep

  • Orchestrated three waves: parallel Sonnet pair → Opus → Sonnet sweep.
  • Wave 1A (Sonnet, planning/data/index): [[06_collaboration]] cleaned of demo [[index.md]] refs (→ [[index]] or backticks); [[08_contributor-charter]] skill wikilinks ([[ingest]], [[lint]], [[edit]], [[contribute]], [[reflect]], [[log]]) converted to slash-commands or backticked filenames; [[CLAUDE]]/[[CONTRIBUTORS]] → backticked filenames; wiki/data/README.md template placeholders wrapped in backticks; wiki/data/The evolution of IT infrastructure.md renamed via git mv to wiki/data/evolution-of-it-infrastructure.md (substantive McKinsey-derived analysis, kept); wiki/index.md gained Glossary section + [[inference]], [[suggested-sources]] under Search, [[evolution-of-it-infrastructure]] under Data.
  • Wave 1B (Sonnet, hypothesis files): CREATED [[H3_orthogonal-axes-under-priced]] stub (~270 words; status deferred-stub; reconstructs premise from how 9 linking pages use it — that axes cutting across L0–L7, e.g. autonomy, deployment topology, buyer maturity, are more predictive of value-landing than ladder position alone; resolves all 9 ghost references). Added ## Confidence body section to [[H1_L0-L7-ladder]] (3 bands: high L0–L3, medium L4–L5 conditional on agentic-scaling-law vs scaling-wall, speculative L6–L7). Fixed [[OECD-sme-enabler-quartet]] casing in [[H2_u-curve-of-value]] (×2) and [[sme-policy-pathway-novice-to-champion]] (×1). Converted [[04_search-log]] in [[03_keyword-search-strings]] to backticked filename.
  • Wave 2 (Opus, concept pages): CREATED 4 concept pages with proper forward + non-hub reverse links:
    • [[value-capture]] — share of generated surplus a stack-position keeps vs. creates; bear-case wedge; reverse links from [[lee-01ai-pivot]], [[agentic-revolution]], [[digital-empires-tripartite]], [[sovereign-ai]], [[world-models-jepa]].
    • [[application-layer]] — L0–L2 territory; top-of-U as a claim about workflow-embedded app-layer firms; reverse links from [[circular-ai-economy]], [[middle-layer-defensibility]], [[distribution-moat]], [[zitron-circular-economics]], [[karpathy-software-3]].
    • [[displacement]] — substitution side of labor effect (vs augmentation); names theoretical-vs-observed measurement debate; reverse links from [[capital-labor-divergence]], [[task-based-framework]], [[observed-exposure-measure]], [[ai-young-worker-hiring-slowdown]], [[exposed-worker-demographics]].
    • [[ai-capex-cycle]] — hyperscaler boom-then-bust mechanism (training capex → uncertain inference demand → overcapacity → write-downs on short-half-life GPUs and stranded power); reverse links from [[circular-ai-economy]], [[hyperscaler-capex-trajectory]], [[ai-factory-huang]]; previously-broken [[ai-capex-cycle]] refs in [[bear-case-synthesis]] and [[goldman-sachs-too-much-spend]] now resolve.
    • wiki/index.md updated with all 4 entries under concepts.
  • Wave 3 (Sonnet, reverse-link sweep): appended ## Referenced by sections to the top 5 hubs listing all inbound pages grouped by directory. Counts: [[H2_u-curve-of-value]] 79, [[H1_L0-L7-ladder]] 39, [[wef-ai-in-action-2025]] 41, [[oecd-sme-ai-adoption-2025]] 33, [[diffusion-vs-innovation]] 32. New concept pages picked up automatically. Frontmatter last-updated refreshed to 2026-05-23.
  • Orchestrator follow-up: added [[H3_orthogonal-axes-under-priced]] to wiki/index.md under Hypotheses.
  • Net change vs pre-edit lint: content pages 124 → 129 (+5: H3, value-capture, application-layer, displacement, ai-capex-cycle). Broken wikilinks 30 → 7 (remaining 7 are false positives — wikilinks intentionally inside backticks or image-embed syntax for evolution-of-it-infrastructure.png). Index gaps 3 → 0. Missing reverse links 459 → 424 (full remediation still requires sweeping non-hub pages; deferred).
  • Suggested next: /lint re-run to confirm; consider splitting [[H2_u-curve-of-value]] (2852 words, now even larger with Referenced-by section); /ingest the queued sources in raw/Clippings/; future reverse-link sweep on the next tier of hubs ([[capital-labor-divergence]], [[autonomy-slider]], [[middle-layer-defensibility]], [[bear-case-synthesis]], [[task-based-framework]], [[circular-ai-economy]]) to recover most of the remaining 424.

[2026-05-24] ingest | Sutskever — From the Age of Scaling to the Age of Research

  • Source: secondary analysis of a Sutskever interview/talk (Main thesis / Evidence / Critique / Methodology structure). LLM-generated structural digest, not verbatim transcript — claims are Sutskever’s; scaffolding is the analyst’s. Promoted to sutskever-age-of-research.
  • Routing: inline path (53 lines, no subsections — single orchestrator pass, no extraction agents).
  • New pages (5):
  • Patches (5):
    • scaling-wall — added Sutskever inside-operator concession of the empirical wall (not the architectural form).
    • agentic-scaling-law — added demand-side validation; Sutskever and Huang agree on the axis (test-time compute + RL), diverge on methodology (aesthetic top-down vs first-principles physics).
    • karpathy-software-3 — cross-linked Karpathy’s “anterograde amnesia” with Sutskever’s continual-learning gap (same diagnosis, different label).
    • world-models-jepa — added Sutskever as a third convergent diagnosis of the autoregressive-pretrain bottleneck (algorithmic successor, complementary to LeCun’s architectural answer).
    • autonomy-slider — added Iron Man vs straight-shot tension; Karpathy’s partial-autonomy default vs Sutskever’s silo-and-release posture; eval-real-world gap anchors the slider on the left.
  • Hypothesis touches: scaling-wall gains a high-credibility inside-operator endorsement. agentic-scaling-law gains cross-source agreement on direction. H2_u-curve-of-value is reshaped without being adjudicated — the bottom-of-U mechanism partially shifts from “owns the GPUs” to “owns the algorithmic insight.” H3_orthogonal-axes-under-priced modestly reinforced — Sutskever’s aesthetic-vs-physics methodological axis is exactly the kind of cross-cutting variable H3 names.
  • Contradictions surfaced:
    • Sutskever’s straight-shot deployment posture directly contradicts Karpathy’s Iron Man partial autonomy.
    • Value-functions claim is explicitly unfalsifiable in public (“safety + competitive reasons”) — scientifically unverifiable.
    • “Neuralink plus plus” alignment equilibrium falls outside the WEF/Deloitte regulatory frame; pushes alignment from policy to transhumanist biology.
  • Open questions surfaced: Does the research pivot favour incumbents or small algorithmic teams? Is the “Age of Research” a phase (months) or a paradigm reset (years/decades)? What signal drives next-gen training if natural data is exhausted?
  • Source moved to raw/ingested/.

[2026-05-24] ingest | Weber et al. — AI Startup Business Models (BISE 2021)

  • Source: Weber, Beutter, Weking, Böhm, Krcmar (2021), AI Startup Business Models: Key Characteristics and Directions for Entrepreneurship Research, Business & Information Systems Engineering 64(1), 91–109. doi:10.1007/s12599-021-00732-w. 19-page research paper; promoted to weber-ai-startup-business-models.
  • Routing: inline path (19-page PDF read directly; no marker needed; single orchestrator pass).
  • New pages (8):
  • Patches (8):
    • H1_L0-L7-ladder — added “Seller-side mirror” section pointing to Weber’s four patterns.
    • H2_u-curve-of-value — added “Seller-side instantiation” section mapping all four Weber patterns onto the U; documents the mild counter-evidence to the strong “middle dies” form (the 30-startup Data Analytics Provider cluster).
    • H3_orthogonal-axes-under-priced — added “Seller-side empirical reinforcement”; Weber’s 11 dimensions are the strongest activation evidence yet for H3.
    • H4_rl-specialization-value-pocket — added “Seller-side archetype lineage” paragraph; H4 sits inside the Deep Tech Researcher pattern with a 2026 substitution (does not train own base model).
    • ecosystem-business-archetypes — added “Companion taxonomy” section explaining how Choudary (horizontal) and Weber (BM) complement each other inside Axis 3.
    • middle-layer-defensibility — added “Seller-side population evidence” sharpening generic-middle-dies / workflow-embedded-middle-survives with Weber’s cluster sizes.
    • distribution-moat — appended Weber-archetype framing to the “what doesn’t qualify” section.
    • ai-skill-shortage-as-diffusion-bottleneck — added “Seller-side mirror” pointing at Deep Tech Researcher demand.
    • taker-shaper-maker · vertical-ai-orchestration · rl-apis — added one-line reverse references.
  • Hypothesis touches: H1 modestly reinforced (gains its seller-side sibling). H2 qualitatively reinforced + empirically complicated (large viable middle cluster). H3 modestly activated (strongest evidence so far that orthogonal axes outperform ladder position). H4 lineage clarified (Deep Tech Researcher with 2026 base-model substitution).
  • Contradictions surfaced:
    • The 30-startup ai-data-analytics-provider cluster is mild counter-evidence to the strong form of “middle dies.” Resolution: H2 needs to distinguish firm count from value-capture share.
    • Weber et al. expected ethics as a salient dimension; taxonomy contains none. Worth a future-paper note.
  • Open questions surfaced (from Weber Table 4 / discussion §5):
    • When is data not essential to value creation for AI startups?
    • How can AI startups create competitive advantage via AI model leadership?
    • What type of AI technology is easier to replicate than others?
    • At least the first three intersect with active project questions on value-capture and distribution-moat; should be folded into 05_open-questions on next pass.
  • 2021 → 2026 vintage caveat documented: Foundation-model commoditization has reshaped AI Development Facilitator pattern; a plausible fifth pattern (agentic-workflow providers) may need adding on next refresh.
  • Source moved to raw/ingested/.
  • Next: /lint to verify reverse-link integrity; consider promoting the four Weber open questions into 05_open-questions entries.

[2026-05-24] visualize | weber-ai-startup-taxonomy

  • Output: outputs/weber-ai-startup-taxonomy.html
  • Two synced views of the Weber et al. (2021) 4-category / 11-dimension / 39-characteristic taxonomy:
    • (01) Morphological box — chips, multi-select per dimension, sticky “your model” recipe bar, hover tooltips.
    • (02) Radial mind map — all 39 characteristics around the perimeter; selected leaves and the radial path back through their dimension/category/root highlight in accent.
  • 4 archetype quick-loads (ACPS / ADF / ADAP / ADTR) preload canonical patterns.

Graph View