Open Questions

Decisions still to make, ordered by how soon they need to land.

Before research starts in earnest

• Byline / brand wrapper — Redacy, personal byline, or neutral “research collective” wrapper? Each implies a different distribution shape and credibility surface.
• Co-author or expert reviewer — invite a named external reviewer (academic, analyst, recognized operator) to lend credibility before publication, or after?
• Final title — current working title is provisional. Decide before first infographic is commissioned.

Before drafting scenarios

• Geographic frame — three blocs (US / China / Europe), or two (US-China duopoly + Europe as variant), or geography as a cross-cutting lens rather than its own section?
  • Sharpened by geopolitics-global-ai-divide / divergent-value-stack-optima: the three-bloc frame is empirically the right structure, but its role in the paper now matters more than its presence. If each bloc engineers a structurally different value-stack optimum (top-heavy US, bottom-heavy China, regulatory-centric EU), then Section 5 cannot be a standalone “geography” sidebar — it must function as a cross-cut of Section 3 (Where the margin lands). Realistic choices: (a) keep Section 5 as a standalone but reframe Section 3 around the bloc-specific shapes from the start; (b) dissolve Section 5 into Section 3, with each shape getting its own treatment; (c) keep both but make explicit that Sections 3 and 5 are two views of the same object.
• Bear-case anchor — pick one recognized skeptic voice (Marcus, Zitron, Acemoglu, etc.) to engage with seriously in-text. Decide who.
  • Sharpened by bear-case-synthesis: the four candidate skeptics attack different pillars — Marcus (cognitive ceiling), Acemoglu (labor/macro), GS/Covello (capex/infrastructure ROI), Zitron (product-layer circularity). They are complementary, not redundant. Realistic choices: (a) pick Acemoglu as the academic anchor and braid the other three around him as supporting evidence; (b) pick Zitron as the most direct threat to the U-curve and cite the other three as boundary conditions. Avoid forcing one “anchor” if the four pillars together are the actual frame.

Before publishing

• GitHub repo license — MIT vs. CC-BY vs. CC-BY-NC. Affects forking and commercial reuse.
• Embargo / private circulation list — finalize the ~20–30 recipients for pre-launch.
• Italian-language manifesto — translate for Italian press, or English-only?

Possibly out of scope (decide and document)

• Embodied AI / robotics — full chapter, paragraph, or excluded?
• Consumer AI / chat assistants — the paper is operator-facing, but consumer dynamics drive frontier-lab economics. How much to include?

Labor-side incidence (added 2026-05-23 via massenkoff-mccrory-labor-market-impacts-2026)

• Acemoglu vs Anthropic on labor incidence. Acemoglu (2024) predicted exposed workers would be low-education clerical women; Anthropic (2026) finds they are more educated, higher-paid, more white/Asian. The reconciliation depends on whether the theoretical-vs-observed-capability-gap (capability outruns deployment by ~3×) is structural or transitional. Either pick a primary frame for the paper or treat the disagreement explicitly as an open research front.
• Is “unemployment” the wrong indicator? unemployment-did-exposed-workers shows no detectable AI displacement in unemployment through mid-2025, but ai-young-worker-hiring-slowdown shows a ~14% drop in 22–25-year-old hiring rates in exposed occupations. If labor-side AI displacement appears as suppressed hiring rather than rising unemployment, the paper’s labor-side framing needs to use the right dashboard. Decide whether to highlight entry-level hiring as the leading indicator.
• Labor-side U vs value-stack U. Both U-shapes exist in the wiki but on different axes. The paper needs to either present them together as a coherent two-curve story (recommended) or separate them cleanly. Decide before Section 3 is drafted.
• Anthropic publishing labor-market research about its own deployment. The conflict-of-interest dimension is unstated in massenkoff-mccrory-labor-market-impacts-2026 but is itself an open question. Does the paper engage this directly, treat the Anthropic measure as one of several (alongside Brynjolfsson, Hampole), or ignore it?

Collaboration model (added 2026-05-22, partially resolved 2026-05-23)

Decisions promoted from 06_collaboration:

• /contribute skill scope — resolved 2026-05-23: full at v1 (all four kinds — source, daily thought, concept note, open question). Implemented in .claude/skills/contribute/SKILL.md.
• Newsletter platform — Buttondown vs. Substack vs. self-hosted Mailgun. Decision driven by list vs. publication framing.
• Contributor charter — resolved 2026-05-23: see 08_contributor-charter. Terse + filtering posture, CONTRIBUTORS.md + per-page footer credit model.
• Community channel choice (T3) — LinkedIn page, X account, or Discord — pick one, defer the others. See 06_collaboration for tradeoffs.
• Static site domain — resolved 2026-05-23: GitHub Pages default (supgrade.github.io/where-AI-value-lands). Custom domain deferred until T0 produces signal.
• Quartz vs. alternatives — resolved 2026-05-23: Quartz. Recommended generator for Obsidian-flavored wikis; native wikilink support; deployed via GitHub Action to GitHub Pages.
• Repo license — resolved 2026-05-23: MIT (code) + CC-BY-4.0 (content). See LICENSE and LICENSE-content.
• Repo owner + name — resolved 2026-05-23: github.com/Supgrade/where-AI-value-lands (personal account, name includes “AI” to surface the subject at URL level).

Analytical vocabulary (added 2026-05-22)

Decisions promoted from 07_analytical-vocabulary:

• “Archetype” as canonical term — the project uses “business archetype” to name the 15 organizational types (agency, vertical SaaS, micro-SaaS, etc.). Is this the right term, or should the paper use “value-capture pattern,” “firm type,” or “go-to-market form”? Needs to be locked before the paper’s glossary (Section 8) is drafted.
• Interface layer vs. application layer — the top two stack layers blur in practice (Cursor is both application and interface). Does the paper need a clean distinction, or is the overlap acceptable? If we collapse them, the stack has four layers, not five.
• Multi-archetype organizations — OpenAI is simultaneously a foundation model lab, an AI developer tool (API), and a consumer app. The paper needs a rule: do we assign a primary archetype, or treat them as multi-archetype? Answer affects how we use the archetype list in Section 6.
• Axis 6 (Taker/Shaper/Maker) scope — this posture typology was designed for enterprise AI buyers (WEF/Accenture). Does it apply cleanly to AI-native suppliers (e.g., Cursor is a Shaper building for developers who are Makers)? Or should Axis 6 be buyer-side only?
• Which axes appear in the paper explicitly — all six axes currently live in 07_analytical-vocabulary. The paper probably cannot introduce all six in Section 2 without losing the reader. Candidate collapses: Axis 5 (adopter maturity) appears only in Section 5 and 6; Axis 6 (posture) is footnoted in Section 6. Decide which axes are load-bearing for the main argument vs. which are supporting evidence.

RL-driven specialization as value pocket (added 2026-05-24 via H4_rl-specialization-value-pocket)

• Break-even contract size. At what customer contract value (€/year, multi-year) does dedicated RL fine-tuning + harness build break even against an equivalently good base-model + retrieval + tools play? Unit economics of the boutique vendor archetype are unproven.
• Verifiable-reward feasibility across verticals. Which industrial domains actually admit a constructible verifiable reward in the rl-from-verifiable-rewards sense? Software is easy (tests). Energy retrofitting / industrial implant intervention proposals — the author’s worked example — have month-long noisy feedback loops and contestable attribution. Open question whether value-functions-as-algorithmic-emotion-class intermediate signals close the gap.
• Open-weight vs closed-frontier base model choice. The single most consequential vendor decision in this hypothesis (rl-open-vs-closed-source, rl-apis): owning the weights enables continual-learning-paradigm-style compounding moat inside one customer; renting them via OpenAI RFT / Anthropic custom models gives better base capability but disposable IP. Both paths exist; the economics under each are not yet legible.
• Boutique vendor vs frontier-lab tuning API. Does a startup that does domain RL fine-tuning as a service survive the moment frontier labs ship verticalised fine-tuning at scale? Closest precedent: vertical AI orchestrators in industrials (vertical-ai-orchestration, deloitte-ai-dossier-eri) — but the model layer was rarely the moat there.
• Customer data exclusivity. Will industrial / SME customers — especially the Italian family-business archetype — agree to let proprietary process know-how flow into a vendor’s training data? If not, the cross-customer amortisation that the vendor’s unit economics need is contractually forbidden.
• Where on the ladder does H4 sit. Tentatively L3–L5 substrate (agent + harness + RL-tuned model on rented or owned compute), Phase 3+ buyer maturity (enterprise-adoption-ladder), taker-shaper-maker Maker posture at vertical scale, oecd-sme-adopter-taxonomy Champion-adjacent customer. Confirm or revise as evidence accumulates.

Operator-side hypothesis and P2 paper-portfolio (added 2026-05-26 via H5_ai-as-operational-not-product and 09_paper-portfolio)

The operator-side reframe of 2026-05-26 has spawned a sibling paper (P2 — What to Start Now) and a new working hypothesis (H5). See 09_paper-portfolio for the portfolio decision. Open questions:

• [P2] What is the operator-builder’s capital threshold for Track B? H5 claims two tracks (VC-megafund race-in vs operator). The boundary needs a load-bearing variable: capital threshold ($XSeriesA),revenuethreshold($Y ARR), or distribution-prior threshold (audience of size Z). The three correlate but are not identical.
• [P2] Does the “Cursor for local SEO” archetype exist in 2026? Bear case for H5: small-vertical AI-orchestration products at sub-Series-B scale could constitute a viable third track between A and B. Empirical search across 3–4 verticals (local SEO, dental marketing, restaurant ops, real-estate listings) needed before H5 hardens.
• [P2] What is the right Weber-classifier home for the operator-builder? None of the four 2021 Weber patterns fits — operator-builders are not AI startups in Weber’s seller-side sense. This becomes the next gap dimension on weber-taxonomy-2026-gaps.
• [P2] How does H5 interact with H4_rl-specialization-value-pocket? H4 (seller-side) and H5 (operator-side) — are H4 vendors suppliers to H5 operator-builders, or are they competitors for the same surplus? Decide which framing the paper uses.
• [P2] Italian-market or G7? sme-ai-adoption-gap / g7-sme-ai-policy-pluralism suggest the operator-track is structurally harder in Italy than in the US or UK. Is H5 a global hypothesis or specifically guidance for the author’s home market?
• [P2] Where does P2 live commercially? P1 is free PDF + GitHub-Pages wiki. P2 might want a different distribution surface (Substack, cohort-based course, Italian-language landing page). Decide before P1 ships so the P2 surface can be staged.
• [P2] Italian-language P2 vs English-only? The operator-builder reader the author has most empathy for is Italian. Is the natural primary language for P2 Italian (with English secondary), or the same English-primary stance as P1?

Industrial-AI rollup of captive suppliers (added 2026-05-26 via H6_industrial-ai-rollup-captive-suppliers)

H6 is the rollup variant of H5 — a holding company acquiring captive-supplier SMEs in a single industrial customer’s supply chain and installing the ai-first-company-loop in each, replacing the management layer while preserving production. The thesis is speculative; the four sub-questions below are the ones the author flagged in the 2026-05-26 daily-thought reflection.

• [P2] Why are manufacturers not AI-first — newness or structural barrier? The wiki’s read (ai-productivity-firm-level, ai-skill-shortage-as-diffusion-bottleneck, sme-connectivity-divide, sme-ai-finance-gap, oecd-sme-enabler-quartet) is structural, not newness — AI is a multiplier of pre-existing complementary capital, and most SMEs do not have the complementary capital to multiply. Confirm with primary Italian-district data (ISTAT, Cerved) and decide whether the paper presents this as a settled finding or as a live debate.
• [P2] Will it be a natural transformation with time, or will the gap widen? sme-ai-adoption-gap (40% large / 11.9% small OECD-wide 2024) is widening between 2020 and 2024, not closing. sme-broadband-firm-size-gap has been flat at ~25–28 pp for five years. McElheran’s J-curve says even adopters take years to capture the surplus. The empirical bet behind H6 is not a bet on time alone — it is a bet that targeted acquisition + deliberate rebuild beats organic diffusion.
• [P2] Can acquisition accelerate it, and at what unit economics? This is the question the wiki has almost no material on. The H6 thesis depends on (a) the management layer being a 8–20% addressable cost share, (b) the dominant-customer relationship surviving ownership change, (c) the loop-build cost amortising across 5+ acquired suppliers in similar positions, (d) Italian regulatory and labour-relations exposure not making the management-layer reorganisation prohibitive at scale. Each is an empirical sub-question.
• [P2] Are AI-first manufacturers actually more competitive? ai-productivity-firm-level reports labor-productivity premia of >4% (sometimes >15%) for AI users within cohort across G7 — but most of the premium is not AI-caused: it reflects pre-existing competitiveness. The AI-attributable share for SMEs specifically is smaller. The H6 bet is that a deliberate rebuild (not organic adoption) captures the surplus organic adoption cannot, by installing the complementary capital across all five layers of the ai-first-company-loop simultaneously.
• [P2] Is H6 structurally distinct from H5, or H5 with leverage? H5 is one operator-builder running an AI-native firm; H6 is N operator-builder firms acquired and rolled up. Mechanically H6 is a multi-firm H5; strategically the questions are different (M&A pipeline, post-acquisition integration, loop portability, relational risk at ownership change). Decide whether the paper treats them as separate hypotheses or as a single hypothesis with single-firm and rollup variants.

AI-first-company operating model and software-as-temporary-artefacts (added 2026-05-26 via ai-first-company-loop and software-as-temporary-artefacts)

• [P1+P2] Does “software as temporary artefacts” tighten H2 into an L? The mechanism predicts that the buyer-side market for finished middle-layer workflow products shrinks as AI-first firms route tool needs through in-house synthesis. The surviving middle is primitives for synthesis (LangGraph, RL APIs, eval tooling, frameworks). If empirically true at scale, H2_u-curve-of-value’s middle is sharper than the existing wiki frames it — not just squeezed, partly abolished. Decide whether the paper presents this as a sharpening of H2 or as a separate sub-hypothesis.
• [P2] Is gate design (Layer 4 of the ai-first-company-loop) a defensibility surface for operator-builder firms? The page argues yes — gate-calibration data is accumulated, transferable, and durably valuable; firms that learn to gate well make fewer reputation-damaging mistakes than firms that don’t. Empirical work needed: can we point to any firm in 2026 whose operational edge is plausibly gate-design quality rather than model or distribution quality?

Business archetype taxonomy (added 2026-05-22)

The taxonomy emerged from the daily thought of 2026-05-22. Decisions needed:

• Completeness — are there archetypes missing from the current 15? Candidates: AI research lab (non-commercial, e.g., DeepMind inside Google), government / public-sector AI deployer, AI-enabled hardware company (e.g., Humane, Rabbit), embodied AI / robotics startup.
• North Italian / SME archetype gap — the taxonomy is currently US-stack-centric. The author’s home market (North Italy, traditional SME industry) is an AI adopter, not an AI supplier. A 16th archetype may be needed: “traditional-industry AI integrator” — the type of firm the author is, and the type of firm reading the paper.
• Archetype-to-U-curve mapping — should the mapping of archetypes to H2 positions (top / squeezed middle / bottom) become a standalone wiki page or live inside H2_u-curve-of-value? It is currently a table in the annotated daily thought only.