Where AI Value Lands

Wiki Index

28 min read

About this research

An open secondary-research wiki asking where economic value lands across the agentic AI stack — hypotheses, sources, and citable data; an inquiry, not a fixed thesis.

About the author: Gabriele — student with entrepreneurial experience in the Veneto (Italy), using this project to learn how to move in this market as an operator, not only as an observer.

Wiki Index

Previous: 2026-05-26 (fourth 2026-05-26 edit) — /ask + crystallize on “what stops a founder from fully delegating to AI even when all five layers of the loop are technically in place?” One new concept page: personal-liability-exposure — decomposes Layer 4 of ai-first-company-loop into two distinct sub-mechanisms (hooks = typed pre-/post-action regulation triggered by action class; per-action-class autonomy slider with an automatic ratchet that loosens as the action class earns trust) plus a verification surface (Karpathy’s Iron Man). Names personal liability — every delegated action is taken in the operator’s name, every error debits reputation — as the variable that makes Layer 4 the binding constraint for solopreneurs and operator-builders. Sharpens the prescriptive half of H5_ai-as-operational-not-product (the cost-base advantage in software-as-temporary-artefacts only materialises if Layer 4 ratchets; otherwise the surplus is paid back as supervision time) and gives H6_industrial-ai-rollup-captive-suppliers a candidate transferable operational asset (a calibrated gate library installed across acquired firms). Reverse-link patches on ai-first-company-loop (Related + Pending Referenced by), autonomy-slider (Related), and H5_ai-as-operational-not-product (Related). Long-tail reverse-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 deferred to next /lint sweep.

Previous: 2026-05-26 (third 2026-05-26 edit) — /reflect on the 2026-05-26 daily thought (two threads: SME supply-chain roll-up + software as temporary artefacts / YC five-layer self-improving loop). Three new pages and four reverse-link patches. Created H6_industrial-ai-rollup-captive-suppliers as the first M&A-engine variant of H5 for the P2 paper: holding company acquires captive supplier SMEs in a large customer’s supply chain (40–60% single-customer concentration), replaces the management layer with the AI-native operating loop while preserving the manufacturing labor, amortises the loop-build across N acquired suppliers. Natural geography: Veneto / Lombardia / Emilia-Romagna industrial districts. Confidence low — speculative thesis; descriptive premises supported by wiki; rollup mechanic unvalidated. Created ai-first-company-loop — the YC Root Access five-layer firm-level operating model (sensor data → policy → tool → quality gates → learning), the firm-level cousin of ml-monitoring-quality-cycle, borrowed-framework citation (no full source page for the YC video). Authorial correction baked in: learning closes the loop into all four prior layers, including the quality gates. Created software-as-temporary-artefacts — the strategic-consequences page derived from Layer 3 of the loop: agents synthesise bespoke tooling on demand inside the loop; the middle layer of H2_u-curve-of-value is not just squeezed but in large part abolished, with the surviving middle re-classified as primitives for synthesis (Cursor’s IDE-and-model surface, LangGraph, RL APIs, eval tooling) rather than finished workflow products. Reverse-link patches: H5 (H6/loop/artefacts in Related); H2 (artefacts + loop sharpen middle-collapse from a third direction); middle-layer-defensibility (third-direction squeeze named); ai-development-facilitator (in-house-substitution mechanism added to the 2021 → 2026 vintage problem); ml-monitoring-quality-cycle (firm-level nesting cross-link); ai-productivity-firm-level (H6 names this page as its load-bearing empirical anchor). Open-questions page: 5 new H6-related P2 sub-questions + 2 loop/artefacts sub-questions. Paper-portfolio updated: P2 spine now H5+H6 with operator-side concept anchors ai-first-company-loop and software-as-temporary-artefacts. Long-tail reverse-links (autonomy-slider, continual-learning-paradigm, karpathy-software-3, sme-ai-adoption-gap, etc.) deferred to next /lint sweep.

Previous: 2026-05-26 (second 2026-05-26 edit) — added H5_ai-as-operational-not-product as the first project hypothesis written from the operator’s seat: two-track framing (Track A = VC-megafund race-in, AI-as-product, distribution-fortified — Cursor / Windsurf / Perplexity; Track B = operator-builder, AI-in-operations, non-AI deliverable). Created 09_paper-portfolio recording the decision that the wiki now feeds two papers: P1 (existing descriptive Where Value Lands) and P2 (new prescriptive operator-paper What to Start Now). H5 is P2’s founding hypothesis. Patches: forward-link integrity sweep across 13 concept / source / hypothesis pages (H1, H2, H3, H4, circular-ai-economy, middle-layer-defensibility, distribution-moat, ai-charged-product-service-provider, weber-taxonomy-2026-gaps, taker-shaper-maker, value-capture, foundation-model-layer, karpathy-software-3, scaling-gap, vertical-ai-orchestration, digital-core, fusion-skills) — each with a glossed reverse-link to H5; H3 received its first-ever ## Referenced by block (previously absent — H3 is a stub); H1’s ## Referenced by Hypotheses sub-block also gained the latent-missing H4 entry. Open-questions page received a new ”## Operator-side hypothesis and P2 paper-portfolio” section with 7 P2-tagged questions. Hypothesis touches: H1 neutral / H2 narrowed (top of U occupied not open) / H3 sharpened (operator-vs-seller named as a missing axis) / H4 complemented (seller-side sibling).

Previous: 2026-05-26 — applied 2026-05-25 lint fixes. Created 2 new concept pages: foundation-model-layer (the frontier-model API tier — one of the three layers the U-curve actually names; covers commoditisation pressure from open weights and Jevons, partial defensibility via scaling-wall, and the trough-with-sub-peak shape inside the layer) and exclusive-industry-data-partnerships (stub; data-access variant of distribution-moat). Patched H2_u-curve-of-value with a new ## Confidence body block (four-band: Top high / Middle medium / Bottom under-pressure / Universal contested) + Gültekin and foundation-model-layer back-links. Patched H4_rl-specialization-value-pocket with its full 26-entry ## Referenced by block (previously empty) and bumped last-updated. Added gultekin-pinarbasi-commercialisation-ai-2025 reverse links to 7 concept pages. Added foundation-model-layer reverse links to 7 neighbours. Reordered log.md to consistent newest-first prepend convention (Westerveld + Protschky ingest entries moved from the bottom into the top group). Tier 2 page-splits (T2-B) and Strand B GTM taxonomy (T2-D) deferred.

Previous: 2026-05-25 — ingested Protschky et al. (2025) “What Gets Measured Gets Improved: Monitoring ML Applications in Their Production Environments” (IEEE Access, 13). Primarily a technical research source, anchored in the tech subfolder. Added 5 new pages: ml-monitoring (tech overview anchor), ml-production-environment-characteristics (the C1–C5 vocabulary), ml-monitoring-quality-cycle (the 17 practices on the define / measure / assess / act / control cycle), ml-drift-types (concept / data / virtual / adversarial drift), protschky-ml-monitoring-2025 (source). Patched eval-real-world-gap (C2 named as operational decomposition), continual-learning-paradigm (Practice 16 = operational instance), H4_rl-specialization-value-pocket (monitoring as operational complement to RL specialization — strongly supportive), rl-testing-validation (production-time counterpart cross-link), vertical-ai-orchestration (monitoring as operational stickiness substrate), ai-development-facilitator (monitoring tooling as candidate surviving sub-pattern in the squeezed middle — under-served per Protschky’s 15-tool review), scaling-gap (operational root cause as third reading), weber-taxonomy-2026-gaps (monitoring sophistication as candidate 8th gap dimension), tech/README (catalog). Key takeaway: monitoring is the operational layer that makes H4_rl-specialization-value-pocket and vertical-ai-orchestration possible — without it, the continual-learning-paradigm loop does not close and the deployed model rots. Empirical wedge under ai-development-facilitator: 15 monitoring tools reviewed, none implement the full 17-practice surface.

Previous: 2026-05-25 — ingested Westerveld et al. (2023) “The Business Model Portfolio as a Strategic Tool” (J. Strategy & Management) — recap document. Per author instruction, ingested as the prerequisite framing for weber-taxonomy-2026-gaps and the Weber classifier: the BM (not the firm) is the unit of strategic analysis. Added 2 new pages: business-model-portfolio (concept) and westerveld-business-model-portfolio-2023 (source). Patched weber-taxonomy-2026-gaps with a new “Prerequisite — the unit of analysis” section near the top + See-also entries; patched ai-startup-business-archetypes-weber with a “Unit of analysis” note + See-also; patched weber-ai-startup-business-models See-also. Reverse links added on H3 (BMP frame as structural reason for orthogonal-axes claim), karpathy-software-3 (Huang ↔ Complementing intent; Steinberger ↔ Cannibalization paradox), eval-real-world-gap (Sutskever’s “human reward hacking” ↔ Dogmatic paradox), and digital-core (seller-side mirror: functional architecture as connective tissue across BMs). Three strategic intents (Diversifying / Sensing / Complementing) and four paradoxes (Horizon / Cannibalization / Digital / Dogmatic) now in the wiki as diagnostic vocabulary.

Earlier: 2026-05-25 — partial ingest of Gültekin et al. (2025) “Commercialisation of AI” (BPMJ 31(2)). Per author instruction, ingested only for what bears on weber-taxonomy-2026-gaps. Added a new gap dimension — gap 7: Commercialisation motion / GTM posture — empirically anchored in the paper’s §5 (BAH vs Claessens vs fuzzy/emergent). This is the first gap-page dimension derived from an ingested source rather than from wiki-internal triangulation, and splits the original gap 4 into pure channel-geometry (now narrower) and process-motion (new gap 7). Also added 1 new source page (gultekin-pinarbasi-commercialisation-ai-2025) + “First independent evidence” addendum table on the gaps page (strong support for new gap 7; weak support for gaps 1, 6; silent on gaps 2, 3, 5). Patched weber-ai-startup-business-models with “Independent adoption” subsection and ai-startup-business-archetypes-weber See-also. Key empirical signal: in n=10 sample, the only firm with > 100k customers is the only one not on dev-led, demo-driven, contest-channelled motion — directionally aligned with distribution-moat.

Previous: 2026-05-25 — added weber-taxonomy-2026-gaps concept page: diagnostic on the six dimensions Weber’s 2021 taxonomy is missing for a 2026 sample (model sourcing, autonomy level, orchestration depth, distribution architecture, data flywheel, ethics-as-product) + the agentic-workflow fifth archetype + the two-strand source-hunting plan (Strand A: post-2023 academic update; Strand B: GTM sibling morphological box). Patched ai-startup-business-archetypes-weber, weber-ai-startup-business-models, distribution-moat, H3_orthogonal-axes-under-priced with reverse links. No new ingest — built entirely from existing wiki material; awaiting Strand A / Strand B ingests for source-side validation.

Previous: 2026-05-24 — ingested Weber et al. (2021) AI Startup Business Models. Added 6 new concept pages (ai-startup-business-archetypes-weber + 4 archetype pages + wiener-traditional-it-archetypes) + 1 source + 1 data page (Crunchbase 27.9k→97k). Patched H1/H2/H3/H4 with seller-side mirror; patched ecosystem-business-archetypes / middle-layer-defensibility / distribution-moat / ai-skill-shortage-as-diffusion-bottleneck / taker-shaper-maker / vertical-ai-orchestration / rl-apis with cross-links. Project now has the operational classifier for tagging future AI-startup ingests.

Previous (earlier 2026-05-24): added H4 hypothesis (RL-driven specialization as value pocket) + new tech/ reference subfolder with 7 RL explainer pages (rlhf, rlaif, rl-from-verifiable-rewards, rl-data-preparation, rl-testing-validation, rl-open-vs-closed-source, rl-apis). Patched H1, H2, post-scaling-research-pivot, value-functions-as-algorithmic-emotion, agentic-scaling-law, continual-learning-paradigm, scaling-wall with H4 forward references.

Previous: 2026-05-24 — ingested Sutskever “Age of Research” interview synthesis. Added 4 new concept pages (post-scaling-research-pivot, continual-learning-paradigm, eval-real-world-gap, value-functions-as-algorithmic-emotion) + 1 source (sutskever-age-of-research). Patched scaling-wall, agentic-scaling-law, karpathy-software-3, world-models-jepa, autonomy-slider with cross-source synthesis. ⇒

Catalog of every page in the wiki, one-line summary each. Read first when answering any question. Updated on every ingest.

Paper planning — 01_paper-planning/

  • 00_initial-brief — one-line frame: a secondary-research white paper on where AI value lands, for capital-rich operators.
  • 01_audience — who reads it, what they should be able to do after.
  • 02_purpose-and-justification — why this paper, why secondary research is the right vehicle.
  • 03_structure — proposed ~20-page structure, theoretical + practical halves.
  • 04_distribution — channels, derivatives, timing.
  • 05_open-questions — decisions still to make before research, drafting, publishing.
  • 06_collaboration — how others contribute to the wiki while it’s in flight; tiered surfaces from PRs to a chat-with-the-wiki app. Updated 2026-05-22: repo structure, Quartz static site spec, full /contribute skill specification. T0 + T1 shipped 2026-05-23.
  • 07_analytical-vocabulary — canonical glossary of all analytical axes (stack layer, autonomy level, business archetype, curve position, adopter maturity, strategic posture); vocabulary quick-reference table; sample coordinates.
  • 08_contributor-charter — high-bar standard for what counts as a good contribution; terse + filtering posture; out-of-scope list; evidentiary standards; per-page credit format. Operationalises 06_collaboration §T1.
  • 09_paper-portfolio — 2026-05-26 decision: wiki now feeds two papers. P1 = existing descriptive Where Value Lands. P2 = new prescriptive operator-paper What to Start Now. Shared wiki, shared hypothesis surface, distinct audiences (capital-rich vs capital-light) and distinct questions (where does value land vs what should an operator-builder start now). H5 is P2’s founding hypothesis.

Hypotheses — 02_hypothesis/

  • H1_L0-L7-ladder — agentic AI stack as a ladder of substrate ownership and autonomy.
  • H2_u-curve-of-value — value concentrates at the top and bottom of the stack, evaporates in the middle. Top now empirically validated (Cursor, Windsurf). Bottom uncertain. U-shape may be US-stack artifact.
  • H3_orthogonal-axes-under-priced — deferred stub. Claims axes that cut across L0–L7 (autonomy, deployment topology, buyer maturity) are more predictive of where value lands than ladder position alone. Premise authored in kickoff; evidence not yet developed.
  • H4_rl-specialization-value-pocket — new 2026-05-24. RL-driven domain-specific fine-tuning of base models — sold to a single industrial customer or a narrow vertical as a proprietary operational asset — may be a real value pocket as pre-training saturates. Technical premise grounded in post-scaling-research-pivot + agentic-scaling-law; business case unvalidated. The thing sold is closer to AlphaFold-class domain specialization than to generic image/text-to-3D fine-tuning.
  • H5_ai-as-operational-not-product — new 2026-05-26. Two tracks: Track A (VC-megafund race-in, AI-as-product, distribution-fortified — viable but capital-locked) vs Track B (operator-builder, AI-in-operations, non-AI deliverable — accessible to capital-light builders). For builders outside the race-in track, AI captures durable value as operational input, not as delivered product. First project hypothesis written from the operator’s seat. P2’s founding hypothesis. Confidence medium on the descriptive half (composes existing wiki), low-to-medium on the prescriptive operator-side half (unvalidated). Bear case: “Cursor for local SEO” — a sub-Series-B small-vertical orchestration archetype that, if it exists, forces a third track.
  • H6_industrial-ai-rollup-captive-suppliers — new 2026-05-26. Rollup variant of H5 at the holding-company scale. A firm acquires supplier-tier SMEs that are captive in a single large customer’s supply chain (40–60%+ single-customer revenue concentration), keeps the manufacturing labor and physical production intact, and replaces the acquired firm’s management layer with the ai-first-company-loop while amortising one centralised AI-operations team across N acquired suppliers. Natural geography: Veneto / Lombardia / Emilia-Romagna industrial districts. Confidence low — speculative thesis; descriptive premises supported by wiki (especially ai-productivity-firm-level on the complementary-capital-as-gate finding); rollup mechanic unvalidated. Bear case rich: management layer may be doing tacit and relational work AI cannot absorb; dominant customer may claw back the savings; loop-build cost may not amortise as predicted; Italian labour-relations regime may make the management-layer reorganisation economically prohibitive at scale.

Search — 03_search/

Concepts — 04_concepts/

  • scaling-wall — the architectural ceiling on next-token-prediction reasoning.
  • task-based-framework — Hulten’s theorem applied to AI; the <0.71% TFP ceiling.
  • circular-ai-economy — the financial ouroboros from VC → apps → models → hyperscalers.
  • ai-capex-cycle — boom-then-bust pattern in hyperscaler training capex → inference demand → potential overcapacity; bear-case mechanism.
  • capital-labor-divergence — how AI directs surplus toward capital and away from labor.
  • displacement — the substitution side of AI’s labor effect, distinct from augmentation; measurement debate (theoretical vs observed exposure) and the unemployment-invisibility / hiring-suppression evidence.
  • value-capture — the share of generated surplus that a firm or stack position actually retains; the value-creation-vs-capture wedge.
  • application-layer — L0–L2 territory where end-user AI products live; the layer the top-of-U thesis depends on and the circular-economy thesis threatens.
  • software-3-paradigm — Karpathy’s 1.0/2.0/3.0 framing of computing.
  • autonomy-slider — continuous human↔agent control spectrum; “Iron Man suit” partial-autonomy default.
  • llm-as-operating-system — LLM as CPU, context window as RAM; 1960s-time-sharing analogy.
  • agentic-scaling-law — Huang’s test-time-compute axis; the strongest counter to scaling-wall.
  • democratization-of-programming — natural language as programming language; “vibe coding”.
  • divergent-value-stack-optima — the U-curve is not universal; each geopolitical bloc engineers a different optimum (top-heavy / bottom-heavy / regulatory-centric).
  • digital-empires-tripartite — Bradford’s three governance models (US market / China state / EU rights), each exporting a different form of power.
  • diffusion-vs-innovation — Ding’s GPT-marathon thesis: economic power transitions are decided by diffusion capacity, not invention.
  • open-weight-asymmetry — China’s strategic commoditization of the foundation-model layer; the “Android of the AI era” play.
  • foundation-model-layer — the frontier-model API tier (GPT-4o / Claude / Gemini as APIs); one of the three layers the U-curve actually names; mostly trough, with a sub-peak at the frontier-lab club.
  • brussels-effect-and-mirage — EU regulatory imperialism and its structural limits in high-risk state-controlled domains.
  • sovereign-ai — the European hardware play; AI factories, sovereignty traps.
  • middle-layer-defensibility — the empirical case that distribution-embedded platforms (Cursor, Windsurf, LangGraph) survive the “middle dies” thesis.
  • distribution-moat — owning the direct user relationship as the decisive AI competitive variable (Srinivas, Truell, Mohan).
  • exclusive-industry-data-partnerships — stub. Data-access variant of the distribution moat: contractually-exclusive industry corpora as a structural moat for AI-charged providers in regulated verticals.
  • context-control — Chase’s insight: LLMs fail due to poor context supply, not poor intelligence; orchestration = context management infrastructure.
  • world-models-jepa — LeCun’s JEPA architecture; proposed successor to autoregressive LLMs; next paradigm moat.
  • jevons-paradox-ai — efficiency gains in AI inference expand total compute demand rather than contracting it.
  • scaling-gap — the 74% / 16% empirical artifact (BCG / Accenture): most firms cannot scale AI past pilots; interpretation-flexible evidence for managerial / bear / middle-dies readings.
  • enterprise-adoption-ladder — WEF / Accenture five-phase organizational maturity ladder; buyer-side parallel to L0–L7.
  • foundational-enablers — WEF / Accenture’s six-factor framework (ecosystem, trust, self-governance, talent, cybersecurity, digital core) claimed to close the scaling gap.
  • agentic-revolution — analytical→agentic transition; action-orchestration replaces prediction as the value-capture unit.
  • vertical-ai-orchestration — multi-agent industrial-domain orchestrators as the durable AI moat in asset-heavy verticals (non-software instantiation of middle-layer-defensibility).
  • synthetic-data-generation — generative methodology for dark-data verticals; partially relocates the operational-data moat but inherits a version of scaling-wall.
  • ai-factory-huang — Huang’s framing of the AI-era datacenter as a tokens-per-second-per-watt manufacturing unit; bottom-of-U mechanism, federated across central / regional / edge.
  • digital-core — buyer-perspective 3-layer enterprise stack (apps / data backbone / infra); architectural precondition the bull-managerial frame treats as load-bearing.
  • taker-shaper-maker — WEF strategic-adoption typology (rent / customize / train the model); third orthogonal axis to substrate ladder and maturity ladder.
  • fusion-skills — workforce-capability frame; humans + AI as unit of analysis; labor-side counterpart of the autonomy slider.
  • sme-ai-adoption-gap — the structural firm-size AI adoption gap (40% large / 11.9% small OECD-wide 2024); holds across G7 / applications / sectors / controls.
  • oecd-sme-adopter-taxonomy — OECD four-quadrant SME adopter taxonomy (Novices / Optimisers / Explorers / Champions); two axes (complexity × scope) + digital-maturity diagonal.
  • sectoral-ai-diffusion-pattern — ICT-led, traditional-sector lagging diffusion pattern; ICT ~45% vs Construction 7.2% OECD-wide 2024.
  • ai-productivity-firm-level — firm-level AI productivity premium evidence and methodological caveats; premium shrinks once complementary digital capital is controlled.
  • oecd-sme-enabler-quartet — OECD four-enabler frame (connectivity, AI-enabling inputs, skills, finance); SME-specific cousin of foundational-enablers.
  • ai-skill-shortage-as-diffusion-bottleneck — skills as the most consistently cited SME AI-adoption barrier; 50%+ of SMEs in four G7 countries.
  • sme-ai-finance-gap — credit-condition mismatch for multi-year AI integration; short-term lending shift 2022–2024.
  • sme-connectivity-divide — persistent ~27 pp small-vs-large broadband gap; widening urban-rural mobile gap (5 → 45 Mbps, 2019–2024).
  • sme-policy-pathway-novice-to-champion — stage-conditional policy matrix mapping OECD adopter stages to enabler interventions.
  • g7-sme-ai-policy-pluralism — four operationally distinct G7 SME-AI policy optima (market-led / state-coordinated industrial / state-coordinated soft-law / federal-regional); convergence on instruments alongside divergent framing.
  • observed-exposure-measure — Massenkoff–McCrory usage-weighted occupation-level AI displacement-risk measure; theoretical capability + realized Claude work-context usage.
  • theoretical-vs-observed-capability-gap — Claude work-context usage covers ~33% of Computer & Math tasks vs ~94% theoretical; “the red area will grow to cover the blue” diffusion lag at occupation granularity.
  • exposed-worker-demographics — AI-exposed workers are older, female, more white/Asian, more educated (4× graduate-degree share), 47% higher-paid; partial contradiction of the Acemoglu wage-incidence prediction.
  • ai-young-worker-hiring-slowdown — the only detectable AI labor-side signal: ~14% drop in 22–25-year-old job-start rate in exposed occupations post-ChatGPT; “canary in the coal mine.”
  • eloundou-beta-exposure — theoretical LLM task-exposure scoring (β=0/0.5/1) used as input to most modern exposure measures; 97% of Claude work-context usage falls on β≥0.5 tasks.
  • ecosystem-business-archetypes — Choudary’s four-archetype taxonomy (Aggregator / Integrator / Infrastructure / Capability) for horizontal business models; the cleanest off-the-shelf sub-vocabulary for Axis 3.
  • vertical-to-horizontal-unbundling — structural thesis: falling transaction costs unbundle vertical chains into modular horizontal ecosystems where value concentrates in archetype positions.
  • cross-archetype-confusion — strategic failure mode of benchmarking the wrong ecosystem archetype (B2B firms aping BigTech aggregators).
  • post-scaling-research-pivot — Sutskever’s “Age of Research” reframe; pre-training data exhausted, algorithmic research replaces brute-force scaling as the binding leap.
  • continual-learning-paradigm — sample-efficient on-the-job learning as the proposed successor to static pre-trained models; the algorithmic answer to anterograde amnesia.
  • eval-real-world-gap — superhuman benchmarks coexisting with brittle real-world performance; “human reward hacking” by researchers optimising for evals.
  • value-functions-as-algorithmic-emotion — Sutskever’s RL proposal: biological emotions as the proof-of-concept for cheap robust intermediate trajectory signals during long-horizon learning.
  • ai-startup-business-archetypes-weber — Weber et al.’s 2021 four-pattern AI-startup taxonomy + the operational classifier the project uses to tag future AI-startup ingests. Seller-side mirror of H1_L0-L7-ladder.
  • ai-charged-product-service-provider — Weber Pattern 1: ready-trained AI model embedded into a standardized product; top of H2_u-curve-of-value; Cursor / Windsurf / Perplexity are 2026 instances.
  • ai-development-facilitator — Weber Pattern 2: tooling to help other firms build AI (SDKs, APIs, no-code workbenches); the squeezed middle of the U; survives only in agent frameworks and rl-apis in 2026.
  • ai-data-analytics-provider — Weber Pattern 3 (largest cluster, 30/100): integrate + analyze customer data for decision support; classical analytics with ML inside; shoulder of U.
  • ai-deep-tech-researcher — Weber Pattern 4: frontier R&D as the deliverable (robotics, autonomous driving, drug discovery, neurotech); bottom of U; structural lineage to H4_rl-specialization-value-pocket.
  • wiener-traditional-it-archetypes — the pre-AI baseline IT-BM triad (data users / suppliers / facilitators) Weber positions the AI-startup taxonomy against.
  • weber-taxonomy-2026-gaps — diagnostic on what Weber’s 2021 11-dim / 39-char box fails to encode in 2026: seven missing dimensions (model sourcing, autonomy level, orchestration depth, distribution architecture (channel geometry), data flywheel, ethics-as-product, commercialisation motion / GTM posture — empirically anchored in gultekin-pinarbasi-commercialisation-ai-2025) + the agentic-workflow fifth archetype + the BMP unit-of-analysis prerequisite (business-model-portfolio); working-draft input to the GTM sibling taxonomy.
  • business-model-portfolio — Westerveld et al. (2023). Operating multiple business models has shifted from defensive contingency to proactive strategic tool. Three intents (Diversifying / Sensing / Complementing), four paradoxes (Horizon / Cannibalization / Digital / Dogmatic), and — operationally — the prerequisite scoping rule that BM (not firm) is the unit of analysis when applying ai-startup-business-archetypes-weber to multi-BM AI startups.
  • ml-production-environment-characteristics — Protschky et al.’s C1–C5 vocabulary (data representation / metrics proxy / assumptions / changes / entanglement); operational decomposition of why ML applications fail in production in ways software does not.
  • ml-monitoring-quality-cycle — the 17 ML monitoring practices on a 5-step define / measure / assess / act / control cycle + 3 cross-sectional; the operational bill of materials for keeping an ML application alive in production.
  • ml-drift-types — concept / data / virtual / adversarial drift; the operational form of the C4 (Changes) characteristic; load-bearing for continual-learning-paradigm and H4_rl-specialization-value-pocket.
  • ai-first-company-loop — new 2026-05-26. 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”). Firm-level cousin of ml-monitoring-quality-cycle: Protschky scaffolds production discipline for one ML application; this loop scaffolds the entire company. Borrowed-framework citation; no full source page for the YC video. Authorial correction: learning closes back into all four prior layers, including the quality gates themselves. Anchors the operating-model claims in H5_ai-as-operational-not-product and H6_industrial-ai-rollup-captive-suppliers.
  • software-as-temporary-artefacts — new 2026-05-26. Strategic-consequences page derived from Layer 3 (tool) of the ai-first-company-loop. Agents in the loop synthesise bespoke tooling on demand at near-zero marginal cost. The middle of H2_u-curve-of-value is 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) rather than finished workflow products. Sharpens middle-layer-defensibility and the ai-development-facilitator 2021 → 2026 vintage problem from a third direction (buyer-side in-house substitution).
  • personal-liability-exposure — new 2026-05-26 (fourth edit). Decomposes Layer 4 of ai-first-company-loop into hooks (typed pre-/post-action regulation triggered by action class) + per-action-class autonomy slider with auto-ratchet (from “ask every time” to “routine” as the action class earns trust) + verification surface (Karpathy’s Iron Man). Names personal liability — every delegated action is taken in the operator’s name — as the variable that makes Layer 4 the binding constraint for solopreneurs and operator-builders. Architectural condition for the prescriptive half of H5_ai-as-operational-not-product and a candidate transferable operational asset for H6_industrial-ai-rollup-captive-suppliers (calibrated gate library installed across acquired firms).

Sources — 05_sources/

  • bear-case-synthesis — internal synthesis of the four-pillar bear case.
  • acemoglu-simple-macroeconomics — NBER WP 32487 (stub via synthesis).
  • goldman-sachs-too-much-spend — Top of Mind #129, June 2024 (stub via synthesis).
  • marcus-world-models-failure — Marcus on AI, June 2025 (stub via synthesis).
  • zitron-circular-economics — Better Offline / wheresyoured.at (stub via synthesis).
  • karpathy-software-3 — Karpathy’s Software Is Changing (Again) lecture (with Huang + Steinberger synthesis); bull-frame paradigm document.
  • geopolitics-global-ai-divide — internal synthesis of the tripartite AI governance architecture (US / China / EU) and the divergent value-stack optima each engineers.
  • bradford-digital-empires — Digital Empires (2023) and the tripartite governance frame (stub via synthesis).
  • ding-diffusion-marathon — Ding’s RAND + FPRI work on GPT diffusion and China’s diffusion deficit (stub via synthesis).
  • lee-01ai-pivot — Kai-Fu Lee on 01.AI’s B2B pivot and open-source-for-Global-South thesis (stub via synthesis).
  • where-value-lands-2026 — 2026 synthesis of the technical frontier; empirical validation of distribution moats (Cursor, Windsurf, Perplexity, LangGraph); LeCun JEPA; DeepSeek commoditization; Jevons Paradox.
  • wef-ai-in-action-2025 — WEF / Accenture flagship white paper (Jan 2025); canonical bull-multilateral synthesis of enterprise AI adoption, foundational enablers, and the $7.6–17.9T projection; surfaces the 74% scaling-gap.
  • deloitte-ai-dossier-eri — vertical-industrial deep-dive on Energy, Resources & Industrials; 12 ER&I AI use cases; introduces the agentic-revolution, vertical-ai-orchestration, synthetic-data-generation, and ai-factory-huang concepts (stub-via-synthesis of the Deloitte AI Dossier integrated with WEF and Karpathy/Huang material).
  • non-tech-digital-core-synthesis — secondary LLM-generated executive digest condensing WEF + Karpathy + Huang into a non-tech-enterprise prescription; elevates digital-core, taker-shaper-maker, fusion-skills to dedicated concept pages.
  • oecd-sme-ai-adoption-2025 — OECD discussion paper for the G7 (2025); canonical official-multilateral document on SME-specific AI diffusion; four-enabler quartet, four-stage adopter taxonomy, G7 country policy profiles; load-bearing for diffusion-vs-innovation and the bottom-right of H2_u-curve-of-value.
  • massenkoff-mccrory-labor-market-impacts-2026 — Anthropic internal study (March 2026); introduces observed-exposure-measure; finds no detectable AI unemployment effect through 2025 but tentative entry-level hiring slowdown for 22–25-year-olds; partial contradiction of Acemoglu’s incidence prediction; bull-frame internal datapoint.
  • choudary-ecosystem-teardown — Sangeet Paul Choudary 2022 newsletter teardown; pre-AI canonical statement of the four-archetype ecosystem-business-model taxonomy; structural precursor to H2_u-curve-of-value and to distribution-moat.
  • sutskever-age-of-research — Sutskever (2026) declares the Age of Scaling over and the Age of Research begun; primary-architect concession of the empirical scaling wall; proposes continual learning + value functions + a “straight-shot” deployment posture.
  • weber-ai-startup-business-models — Weber et al. (BISE 2021) — empirical four-pattern AI-startup BM taxonomy (AI-charged P/S Provider / AI Development Facilitator / Data Analytics Provider / Deep Tech Researcher) over 100 Crunchbase startups; 11-dimension × 39-characteristic dimension catalogue; load-bearing for the seller-side reading of H1_L0-L7-ladder / H2_u-curve-of-value / H3_orthogonal-axes-under-priced.
  • gultekin-pinarbasi-commercialisation-ai-2025 — partial ingest. BPMJ 31(2). 10-interview qualitative study of Turkish AI startups (Istanbul / Antalya, Nov 2022). First independent peer-reviewed adoption of Weber’s classifier in the wiki; anchors the new gap 7 (Commercialisation motion / GTM posture) on weber-taxonomy-2026-gaps via the BAH-vs-Claessens-vs-fuzzy contrast in its §5; weakly supports gaps 1, 6; silent on gaps 2, 3, 5.
  • westerveld-business-model-portfolio-2023 — Westerveld et al. (2023), Journal of Strategy and Management. Developmental literature review over 38 empirical cases (2010–2021) on the Business Model Portfolio as a strategic tool. Anchors business-model-portfolio: three strategic intents (Diversifying / Sensing / Complementing) and four permanent paradoxes (Horizon / Cannibalization / Digital / Dogmatic). Ingested in this wiki as the prerequisite framing for the Weber classifier: BM (not firm) is the unit of analysis. Ingested from a structured recap document, not the full paper.
  • protschky-ml-monitoring-2025 — Protschky, Lämmermann, Hofmann, Urbach (IEEE Access 13, 2025). Multivocal literature review (25 academic + 56 practitioner articles) + 10 expert interviews + 15 ML monitoring tools reviewed. Derives 5 production-environment characteristics (C1–C5) + 17 monitoring practices on a 5-step quality-management cycle. Primary source for the technical floor under H4_rl-specialization-value-pocket and the operational layer under vertical-ai-orchestration; surfaces “monitoring sophistication” as a candidate 8th gap dimension on weber-taxonomy-2026-gaps.

Ideas — 06_ideas/

Concrete use-case patterns: how AI is actually being used on the ground. Pattern-level pages (not theoretical frameworks) — what the AI does, who’s deploying it, what’s been measured, what concepts it connects to.

ER&I (Energy, Resources & Industrials) — via deloitte-ai-dossier-eri

SME adoption patterns — via oecd-sme-ai-adoption-2025

Business ideas — 07_business-ideas/

Speculative business / company pages the author is considering building or testing. Each page describes one concrete business — what it sells, to whom, why it would work, what conditions it depends on, what would kill it. Distinct from 06_ideas/ (observed use-case patterns) and from 02_hypothesis/ (project-paper research claims). See README for the folder contract.

  • specialist-subagent-for-orchestrators — new 2026-05-27. Vertical-specialist subagent exposed as an MCP tool, sold into other people’s agent orchestrators (Opus, GPT, in-house). Buyer-side composition framing of H4_rl-specialization-value-pocket: agent builders pay for a drop-in specialist on the one node in their graph where the orchestrator’s accuracy is not good enough. Three required conditions (proprietary data / vertical narrow enough to evade frontier-lab horizontal absorption / low-friction MCP-class integration surface) and three GTM options sequenced A → B → C (single industrial customer → metered API → data-aggregator consortium). Bear cases: frontier-lab horizontal absorption, single-customer surplus clawback, MCP commodification, operations cost dominance, first-party specialist marketplace from the orchestrator vendor. Confidence low.

Tech — tech/

Reference layer: short technical explainer pages grounding the analytical concepts and hypotheses in real algorithmic detail. New 2026-05-24 to support H4_rl-specialization-value-pocket. See README for the catalog.

  • rlhf — Reinforcement Learning from Human Feedback. SFT → reward model → PPO/DPO. The production recipe behind aligned chat models.
  • rlaif — Reinforcement Learning from AI Feedback. Replace the human labeler with a strong model labeler (Constitutional AI).
  • rl-from-verifiable-rewards — RLVR. Deterministic checker as reward (unit tests, theorem provers, simulators). The mechanism behind reasoning-tuned models (o1, R1).
  • rl-data-preparation — trajectory collection, preference labeling, deduplication, contamination filtering, reward shaping.
  • rl-testing-validation — held-out evals, reward-hacking probes, distribution-shift checks, real-world A/B; closes (or fails to close) eval-real-world-gap.
  • rl-open-vs-closed-source — open algorithms (PPO, DPO, GRPO) and frameworks (TRL, OpenRLHF, verl) vs closed recipes inside frontier labs; strategic implications for moat location.
  • rl-apis — the commercial fine-tuning surface: OpenAI RFT, Anthropic custom models, Google Vertex tuning, Together, Fireworks, Predibase, Modal-hosted TRL.
  • ml-monitoring — the production-time monitoring discipline; operational counterpart to rl-testing-validation. Anchored in protschky-ml-monitoring-2025: 5 ML production-environment characteristics (C1–C5) + 17 monitoring practices arranged on a quality-management cycle + 4 drift types.

Data — data/

Citable units of evidence: statistics, tables, figures, images. Each numeric claim in a concept page should resolve to one of these. See README for conventions.

Infrastructure evolution

  • evolution-of-it-infrastructure — McKinsey-framed H2M→M2M shift from CLI/bare-metal (2000) through IaC/NLP (2020s) to intent-based autonomous infrastructure (2030 vision).

Bear-case ceilings

Capex and unit economics

Adoption and bull projections

Distribution-moat empirical anchors

Vertical-industrial deltas

Labor-side AI exposure (Massenkoff–McCrory, 2026)

  • most-exposed-occupations — Top-10 list: computer programmers (74.5%), customer service reps (70.1%), data entry keyers (67.1%), medical record specialists, market research analysts.
  • ai-exposure-vs-bls-growth — Slope −6.07 (SE 1.32), R² 0.027 — 10pp more coverage → 0.6pp less BLS projected growth; theoretical β alone does not correlate.
  • high-vs-low-exposure-worker-characteristics — CPS Aug–Oct 2022: top-quartile-exposed vs zero-exposure demographic/wage/education table.
  • unemployment-did-exposed-workers — DiD pooled post-ChatGPT +0.0020 (SE 0.0019) — no detectable unemployment effect.
  • young-worker-hiring-did — DiD pooled post-ChatGPT −14.3% (SE 7.2) — tentative entry-level hiring suppression for 22–25-year-olds.

SME diffusion (OECD G7, 2025)

Glossary — Glossary/

  • inference — the computational act of running a trained model to produce outputs; why inference cost, Jevons dynamics, and test-time compute are central to the project’s stack economics.
  • taxonomy — a MECE discrete classification (vs. ladder, axis, or curve); lists all six taxonomies used in this project and the capitalization rule for named types.

Thoughts — wiki/thoughts/

Graph View

Backlinks