Handout—Editorial working document
Source Register
Every source the course depends on, in one table, with its tier, the lessons that rest on it, and when it must be re-checked. This is the document you work from during a refresh — not the lessons.
How to use this
Work down the tiers from the bottom. The trade tier moves fastest and carries the
least independent verification, so it is checked first and most often; primary research
rarely moves and is checked annually. Where a source has been superseded, update the lesson
and the verification date in that lesson's data-verified attribute — the
header stamp is generated from it.
Tier 5 — Trade Check first, every 2 quarters
Vendor blogs, consultancies, and industry aggregators. Directionally useful for describing a landscape; every specific figure must be re-verified before it appears on screen. Lessons 7 and 10 depend on this tier more than the rest because no independently audited alternative exists for current model rankings and product features.
| Source | Used in | What rests on it | Risk if stale |
|---|---|---|---|
| Swfte / Local AI Master / Kili — benchmark leaderboards, Aug 2026 | L7 | GPQA, SWE-bench, MMLU saturation figures | High — monthly churn |
| GMI Cloud — “Open-Weight Models Catch the Frontier,” Aug 2026 | L7, L8 | Top-of-market compression; open models needing datacenter hardware | High |
| Mettevo — “LLM Rankings 2026,” Jul 2026 | L7 | MMLU uninformative; tiered model policy | High |
| TeamAI — “The 2026 AI Frontier Model War” | L7 | Routing as production norm | Medium |
| Digital Applied / TechRT — MCP adoption statistics 2026 | L5 | SDK download and server counts | High — counts diverge by source |
| Kellton; Svitla — agentic AI trends 2026 | L5 | USB-C analogy; autonomy levels in production | Medium |
| Cogitx — “AI Agents: Complete Overview (2026)” | L5 | Agent component enumeration | Low — definitional |
| PDP Spectra; AI Business; Capacity; Presenc AI — sovereign AI 2026 | L9 | National programme inventory; announced investment figures | High — announced ≠ committed |
| Marvin-42 Insights — Hugging Face concentration summary, Mar 2026 | L8 | <200-download tail; Fortune 500 accounts | Medium |
| Stern Strategy Group — entry-level job evidence, Jun 2026 | L11 | Leadership-pipeline framing | Low — commentary |
| Swiss Cyber Institute; Open Digital AI — AI history timelines | L1, L3 | Second-winter dating; pattern-prediction characterisation | Low |
| AEI — “How AI Is like That Other GPT, Electricity” | L13 | Barcode lag; Solow paradox framing | Low — summarises primary work |
| KPMG — “AI: The second great shift” | L13 | Engels' Pause applied to the present | Low |
Tier 4 — Journalism Every 2–3 quarters
Reliable for events and dates. Any figure they report should be traced to its underlying source before it goes on screen.
| Source | Used in | What rests on it |
|---|---|---|
| CNN Business — Character.AI/Google settlement, 7 Jan 2026 | L4 | Settlement scope, parties, confidentiality, ongoing suits |
| K-12 Dive — mediation of wrongful-death suits | L4 | Corroboration of the above |
| MIT Technology Review — 10 Breakthrough Technologies 2026 | L4 | Companion AI category status |
| ScienceAlert — teen companion coverage, Jul 2025 | L4 | Companion vs. task-assistant definition |
| Forbes — 12 Aug 2026 | L11 | Aggregate ADP growth ≈6% vs ≈4% most-exposed quintile |
| Fortune — 27 Jun 2026 | L11 | Robustness checks; Acemoglu as principal sceptic |
| TechCrunch — AAIF formation, 9 Dec 2025 | L5 | Foundation membership |
Tier 3 — Documentation Every 2 quarters
First-party product and technical documentation. Authoritative about what a product is; not evidence about whether it works. Feature sets change fastest here.
| Source | Used in | What rests on it |
|---|---|---|
| Claude Code documentation — Overview | L10 | Agentic coding definition; CLAUDE.md; MCP integration; surfaces |
| Anthropic — “Introduction to agentic coding” | L10 | Assistant vs. agentic distinction; “compresses the build loop” |
| Anthropic — “Building Effective AI Agents” | L5, L15 | Workflow vs. agent; simplest-solution recommendation |
| Anthropic / MCP blog — AAIF donation, 9 Dec 2025 | L5 | MCP release date, governance transfer |
| Anthropic — “Labor market impacts of AI” | L11 | Observed exposure; postings ≠ harm caution (interested party — flagged in lesson) |
| OpenAI — “What are tokens and how to count them” | L3 | ≈4 chars / ≈¾ word per token |
| OpenAI — “Why language models hallucinate” | L6 | Research summary accompanying arXiv:2509.04664 |
| IBM Research — Deep Blue | L2 | 1997 match result; ≈200M positions/second |
Tier 2 — Institutional Every 2–3 quarters
| Source | Used in | What rests on it |
|---|---|---|
| Brynjolfsson, Chandar & Chen — Canaries in the Coal Mine?, Stanford Digital Economy Lab, rev. 12 Aug 2026 | L1, L2, L11, L12 | The six facts; ≈19% gap for ages 22–25. Under active revision — always cite the version. Prior versions reported ≈13% and ≈16%. |
| Common Sense Media — Talk, Trust, and Trade-Offs, 16 Jul 2025 | L4 | n = 1,060 teens 13–17; 72% ever-used; ≈50% regular; risk assessment |
| Hugging Face — State of Open Source, Spring 2026 (17 Mar 2026) | L8, L9 | 13M users; >2M models; 49.6% top-200 share; 41% Chinese-origin; producer shift; Korean models trending |
| IMF SDN/2026/001 — Bridging Skill Gaps for the Future, Jan 2026 | L11 | AI-skill diffusion vs. employment in high-exposure occupations. Frequently miscited by subtitle alone. |
| SIEPR — “What is really happening to jobs?”, Jul 2026 | L11 | Recent-graduate unemployment 5.6%, +1.6pp over three years |
| Gartner — task-specific agents forecast, 26 Aug 2025 | L5 | 40% of enterprise apps by end-2026. A forecast, not a measurement. |
| Linux Foundation — AAIF formation, 9 Dec 2025 | L5 | Founding projects and members |
| TechPolicy.Press — “Rethinking Sovereign AI as Strategy,” Mar 2026 | L9 | The layer framing; sovereignty as non-binary |
| Carnegie Endowment — “Early Lessons in the Pursuit of Sovereign AI,” Jun 2026 | L9 | Corroboration of the layer framing |
| Draft South Africa National AI Policy 2026 — cabinet approval 25 Mar 2026 | L9 | Policy-formation breadth beyond the G7 |
| Duke University Libraries — “Why Are LLMs Still Hallucinating?”, 5 Jan 2026 | L6 | Working definition; data-quality dependence |
| Lighthill — Artificial Intelligence: A General Survey, SRC 1973 | L1 | Combinatorial explosion; first winter |
| Rockefeller Foundation / Rockefeller Archive Center | L1 | $13,500 requested, $7,500 awarded |
| Dartmouth College — “AI Coined at Dartmouth” | L1 | Founding conjecture wording |
| METR — early-2025 developer productivity study | L10, L14 | −19% measured / +20% perceived. METR labels it historical. |
Tier 1 — Primary Annually
Peer-reviewed papers, preprints, and archival documents. Load-bearing; put these on screen. Preprints are flagged in their endnotes as not peer reviewed.
| Source | Used in | Status |
|---|---|---|
| McCarthy, Minsky, Rochester & Shannon — Dartmouth proposal (1955) | L1 | Archival |
| McCulloch & Pitts (1943); Turing (1950); Rosenblatt (1958) | L1 | Peer reviewed |
| Newell & Simon (1956); McCarthy, Lisp (1960); Weizenbaum, ELIZA (1966) | L1, L4 | Peer reviewed |
| Minsky & Papert, Perceptrons (1969) | L1 | Monograph |
| Rumelhart, Hinton & Williams, Nature 323 (1986) | L1 | Peer reviewed |
| Krizhevsky, Sutskever & Hinton, NeurIPS 25 (2012) | L1, L2 | Peer reviewed |
| Silver et al., Nature 529 (2016) and 550 (2017) | L2 | Peer reviewed |
| Vaswani et al., “Attention Is All You Need” (2017) | L2 | Peer reviewed |
| Brown et al., GPT-3 (2020) | L2, L3 | Peer reviewed |
| Sennrich, Haddow & Birch, subword units (2016) | L3 | Peer reviewed |
| Liu et al., “Lost in the Middle,” TACL 12 (2024) | L3 | Peer reviewed |
| Wei et al., chain-of-thought (2022) | L3 | Peer reviewed |
| Bender et al., “Stochastic Parrots,” FAccT (2021) | L3 | Peer reviewed |
| Kalai, Nachum, Vempala & Zhang, arXiv:2509.04664 (2025) | L6, L15 | Preprint |
| Lewis et al., RAG, NeurIPS 33 (2020) | L6, L15 | Peer reviewed |
| Becker et al. (METR), arXiv:2507.09089 (2025) | L10, L14 | Preprint |
| Hosseini & Lichtinger (2025), via arXiv:2507.16078 | L11 | Preprint |
| Makrehchi, arXiv:2510.12859 (2025) | L12 | Preprint |
| “Closer to Language than Steam,” arXiv:2506.10281 | L12, L13 | Preprint |
| Crafts, Oxford Review of Economic Policy 37(3) (2021) | L13 | Peer reviewed |
| David, American Economic Review 80(2) (1990) | L13 | Peer reviewed |
| Brynjolfsson, Rock & Syverson, AEJ: Macro 13(1) (2021) | L13 | Peer reviewed |
| Allen, Explorations in Economic History 46(4) (2009) | L13 | Peer reviewed |
| Ericsson, Krampe & Tesch-Römer, Psychological Review 100(3) (1993) | L14 | Peer reviewed; effect size contested |
| Roediger & Karpicke (2006), both papers | All lessons | Peer reviewed |
Three sources that need watching more than the rest
- Canaries in the Coal Mine? — the headline figure has moved 13% → 16% → 19% across revisions. Check for a newer version before every use, and never cite it without the version date.
- Benchmark leaderboards (L7). Vendor-reported, unaudited, and mutually inconsistent — during preparation they disagreed about which model versions existed.
- MCP ecosystem counts (L5). Registry counts, directory listings, and GitHub topic counts differ by an order of magnitude. Always state what was counted and when.
Refresh sign-off
Review date ______________ Reviewer ______________________
Sources found to have moved, and the lessons updated:
Sources found dead or superseded, and their replacements:
data-verified attributes updated in which
lesson files?