# Changelog

## 2026-10-03
- Guide: rebuilt the opening around all ten ideas, grouped money, work, India, and coordination; added illustrated talk previews with user-controlled phone navigation. Removed the separate six-part Guide, introductory history, open questions, About section, and PDF promotion from the homepage.
- Library and audit: moved all 93 published readings to a separate searchable page, kept three homepage starting links, preserved source commentary and old reading bookmarks, and moved the changelog to a plain dated list with a text record.
- Talk: added The Economics of Innovation for Takshashila (4 October 2026), after Ashish approved the content and design. Twenty slides, academic citations and PDF readings, generated conceptual imagery, and a downloadable PDF. The two existing AI talks and reading-list entries remain as previously published in this release.

## 2026-10-01
- Library: added Nathan Goldschlag's "AI and Entrepreneurship: What is Really Happening?" (Economic Innovation Group) to The work — the federal-data dissent to the Stripe and NYT solopreneur entries: applications from businesses not planning to hire are up about 38% since 2024 on the higher of his two measures, but solo work as a main job is flat, so the boom so far looks like side gigs.
- Library: added a16z's "State of Markets II" (David George) to The money and Reading the news — the investor's case that the capex is working, read alongside the free-cash-flow squeeze and new borrowing on its own pages. 91 entries to 93.
- Talk (SocioEconomicAI): "The age of the solopreneur" now carries Goldschlag's side-gig finding and source. "The agentic shift" label now reads "June 2026 evidence" instead of "as we speak, June 2026". To fit the new appendix row, the two 23 September appendix rows were merged into one, keeping their content. Logged in the deck's changelog appendix. Slide count unchanged at 43.
- Talk (EconofAI): from a16z's State of Markets II. "The size of the bet" replaces the $3tn four-year estimate with $4.6tn of consensus hyperscaler capex for 2027–2030, and adds the financing: free cash flow near zero until about 2028, and over $300bn of new bonds and off-balance-sheet debt by hyperscalers and Nvidia by end-August, with a16z's answer that free cash flow recovers by 2030; the revenue-hurdle card now notes it was sized before capex passed $1tn a year. "Why the market wants to concentrate" adds OpenRouter as a second gateway: Anthropic, Google and OpenAI take about 70% of spend on about 35% of tokens (a16z co-led OpenRouter's 2025 funding; disclosed on the slide). "Depreciation" adds a16z's rental data to the bulls' side (rent-implied resale values for H100s and H200s rising through 2026) and the rebuttal: rents track today's demand, so a slowdown would hit every chip generation at once while the debt is still owed. Logged in the deck's changelog appendix. Slide count unchanged at 34.
- Talks: both decks and the guide now say the talks were first given in June 2026 and are updated weekly, replacing "preserved exactly as delivered" and "The Talks as Delivered", which had stopped being true. The about page now says ✓ sources were checked when added, not only in June. Logged in both decks' changelog appendices.

## 2026-09-30
- Exhibit: added a lab filter to both Pareto views, including Anthropic and OpenAI. Points, frontier, previous frontier and summaries follow the selected lab across dates; All labs restores the full comparison.
- Exhibit: refreshed the professional-work Pareto chart with the 30 September GDPval-AA capture, including Opus 5.5, Sonnet 5.5 and GPT-6.1 Sol. Preserved the 20 September capture; added snapshot navigation, confidence intervals, full-field viewing and benchmark-cost provenance. The comparison now covers 125 active priced configurations.
- Exhibit: restored a six-month Arena comparison from June 2024, with date controls and playback. Earlier scenes are explicitly labelled reconstructions from later ratings and prices. The latest scene selectively refreshes 24 rows and adds eight from the 25 September Text English leaderboard, checked on 30 September. Captured data and deterministic generators are retained.

## 2026-09-27
- Talk (SocioEconomicAI): "The canary" updated to the 12 August 2026 revision of Brynjolfsson, Chandar and Chen's "Canaries in the Coal Mine". The headline figure moves from 16% to 19%, and its label now says what it is: the raw employment gap for 22–25-year-olds in the most AI-exposed jobs versus less-exposed peers, November 2022 to June 2026, with no controls. The old 16% was the authors' earlier firm-controlled headline (September 2025 data), which the revision no longer leads with. On the authors' current filters the same method now gives about 7% for the most exposed fifth of jobs and 11% for the next; on the original filters the most-exposed-fifth estimate is about 5% and no longer statistically significant, while the next fifth's stays at about 12%. The bottom line now gives the 7% and 11%, adds that an education control roughly halves the uncontrolled estimates, and notes that national survey data to 2024 shows a much smaller gap, except in professional and finance jobs. Library entry updated to match. Logged in the deck's changelog appendix. Slide count unchanged at 43.
- Library: added four essays by labour economist Guy Berger (Macro Mostly) — the Census AI-adoption read ("tiny" labour impact so far), measuring graduate unemployment properly, the Great Stay's toll on young entrants, and "Data Centers and Butter" — plus the Census working paper behind the adoption figures (Bonney et al., CES 26-25). 86 entries to 91.
- Talk (SocioEconomicAI): "The aggregate job apocalypse has not shown up" now carries Census's figure that only about 1% of US firms say AI changed their headcount, split between up and down. "The canary" keeps the Brynjolfsson finding but adds Berger's dissent: AI is playing a role but not a dominant one, hiring is weak (US hiring rate 3.2% in July 2026, comparable to 2010–11), and graduation timing can inflate graduate-only effects, though he thinks the Canaries result survives it. To make room, the canary's closing line about high-intensity adopters expanding entry-level headcount was dropped; that evidence stays on "Labor evidence" (Ramp's 10.2%) and "Firm-level evidence" (+12.0% entry-level). "Policy one - measure" cites Berger's weekly round-up as the habit India's dashboard should copy. Logged in the deck's changelog appendix. Slide count unchanged at 43.
- Talk (EconofAI): "The size of the bet" gains one sentence on the macro cost of the buildout (Berger: butter, AI data centers, low inflation — pick two). Logged in the changelog appendix. Slide count unchanged at 34.

## 2026-09-23
- Library: added "Quiet Quitting Comes to India. AI Will Make It Worse" (Andy Mukherjee, Bloomberg Opinion) to The work and India — Gallup's engagement numbers for India, TeamLease's IT fresher-hiring series, and Autor's patent-lawyer trial on who keeps the skill AI lends.
- Library: added Autor et al., "Does AI Assistance Enhance or Erode Expertise?" (NBER w35720) to The work — the randomised trial behind the apprenticeship worry.
- Talk: SocioEconomic deck, "The fragile part is the ladder into good work" — added Autor's trial to the entry-level card and Gallup (23% engaged) and TeamLease (IT fresher-hiring intent 81% → 76%) as early Indian signals; logged in the deck's changelog appendix. Slide count unchanged at 43.
- Talk: SocioEconomic deck, "India is a large Claude market" — refreshed from Anthropic's November 2025 India brief to our calculation from Anthropic's April–May 2026 Economic Index data: share 7.1% (#2), per capita about #102 of 121 (tied), four states 51.2%, about 10× time ratio. Software fell from the world's highest share (45.2%) to 26.2% against 23.8% globally, on a measure not strictly comparable. Headline changed from "a narrow one" to "a thin, concentrated one".
- Library: added Anthropic's June 2026 Economic Index report and data; rewrote the India brief's take as the November baseline.

## 2026-09-20
- Exhibit: refreshed the falling-price-of-intelligence Pareto frontier against Arena's September 13 Text English overall leaderboard; added eleven priced August–September entrants including GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash, GLM 5.3, Qwen 3.8 Max, and Muse Spark 1.2, refreshed the current ceiling/frontier contenders, and documented the exclusion of unpriced Muse Spark 1.3.
- Records: archived the exact pre-update public exhibit HTML and data in the personal Obsidian presentation archive before editing.

## 2026-09-19
- Talk (EconofAI): added Bryan and Gans's human-AI system model to the organization argument, OpenAI's internal research-workflow evidence to the firm-reorganization slide, and METR's independent incident evidence to the frontier-pause verification design. Added a dated depth appendix; the deck is now 34 slides, with the 27-slide core path unchanged.
- Talk (SocioEconomicAI): added three core slides on work's non-income functions, Anthropic's economic scenarios, and the India Country Brief; strengthened the task lens with Bryan and Gans and the frontier-access slide with METR and Farrell. Added a dated depth appendix; the deck is now 43 slides, with 41 on the core path.
- Library: added seven source records for the new deck evidence, bringing the library from 76 to 83 entries.
- Records: archived the exact pre-update public HTML for both talks in the personal Obsidian vault before editing, with source URLs, slide counts, byte counts, and SHA-256 hashes; added `scripts/archive_live_talks.py` and made this a required pre-publish step.

## 2026-08-07
- Talk (SocioEconomicAI): merged the three “What changed?” appendix slides into one compact dated changelog; the deck is now 39 slides, with the 38-slide core path unchanged.
- Talk (SocioEconomicAI): tightened ten text-heavy core slides and two dense changelog slides for live presentation, especially the services-paradox slide; preserved the claims, figures, links, and argument while removing essay-like repetition from the canvas.

## 2026-08-04
- Library: added thirteen sources from Stanford's MS&E 435 "Economics of the AI Supercycle" (Apoorv Agrawal, Altimeter Capital) — both of Agrawal's gross-profit-by-layer essays, six of the nine course lectures, the Vercel AI Gateway index, Altimeter's Q1 2026 SEC 13F, the Stanford Review interview, the course's compute-demand calculator, and Korinek and Lockwood's Brookings tax piece. 63 entries to 76.
- Library: rewrote seven existing takes to carry the tensions the new evidence creates — SemiAnalysis on value migrating to model labs against Agrawal's 7% app-and-lab layer; the capex ledger and the $600B question now disclose that their authors are investors in what they analyse; Thompson's open-weight ceiling against gateway data showing open weights take 29% of tokens on under 4% of spend; and the three-way token-price conflict where only Du documents his basket.
- Talk (EconofAI): split the four-row capex ledger into five — model labs and applications were previously dissolved into "token buyers" — and added a new core slide carrying Agrawal's profit-pool split (semis ~79%, infra ~14%, apps ~7%) with its assumed-margin and double-counting caveats. Extended "So is anyone actually making money?" from three verdicts to five. Replaced the "unknown" disclosure tile with the estimate and its limits. Reconciled the two NVIDIA revenue figures on slides 2 and 6. Added concentration evidence to the market-structure slide and a counter-case to the depreciation slide. Retired the third-hand 5.7% margin-cliff figure in favour of a transcript-sourced one.
- Talk (SocioEconomicAI): rebuilt the services-paradox slide — revenue per employee now computed from TCS, Infosys and Wipro filings rather than resting on a Lightspeed piece about a different metric — and added Ghodsi's connector redesign as counter-evidence, with his commercial interest disclosed. Attributed the "taxing steel during the industrial revolution" line to Korinek and Lockwood directly.
- Correction: an earlier draft of the services-paradox slide put Ghodsi's redesign gain at "twenty-one times the throughput." That was wrong — his old process ran one engineer per connector, so the like-for-like gain is about threefold. The slide now states the raw progression and no multiplier.
- Records: `README.md` no longer claims the guide is unpublished; `tasks/todo.md` carries a superseded notice; `STATUS.md` created as the authoritative current-state record.

## 2026-08-03
- Guide: removed the embedded falling-price-of-intelligence Pareto frontier from The machine; the interactive exhibit remains available as a direct link.
- Talk: added two EconofAI Act III slides on the balance-sheet economics of a frontier-AI pause — why a unilateral pause is unstable, and why a coordinated pause must stand still contracts and credit support as well as training.

## 2026-08-01
- Exhibit: refreshed the falling-price-of-intelligence Pareto frontier against Arena's August 1 Text overall leaderboard; added Claude Opus 5 High and Inkling, updated the July entrants' current scores, and preserved July as a dated comparison scene.

## 2026-07-31
- Exhibit: refreshed the falling-price-of-intelligence Pareto frontier through July 2026 with Muse Spark 1.1, Hy3, Claude Opus 5, GPT-5.6 Sol, Kimi K3, Claude Sonnet 5, and Google’s latest Flash models; preserved the original default-overall Arena ruler, documented OpenAI’s July 30 Terra/Luna price cuts separately, and marked preliminary ratings.

## 2026-07-21
- Talks/library: strengthened the EconofAI open-weight market-structure slide and the SocioEconomicAI frontier-access slide with Ben Thompson’s mechanisms—cost per completed task, rent migration into learning and harnesses, and diffusion/distillation as part of the social search process.
- Library: added Ben Thompson’s “Who’s Afraid of Chinese Models?” to Money/Choices/Questions for its account of AI marginal costs, open-weight diffusion, and rent migration — while reframing its U.S.–China conclusion as a test of portability, forkability, and contestability.

## 2026-07-19
- Library/talks: added Sydney Ember’s New York Times report to The work and updated the firm-formation treatment in both Takshashila decks with the broader Census signal: 5.7 million applications in 2025, while keeping applications distinct from the roughly 500,000 businesses actually formed in 2023.

## 2026-07-09
- Talks: standardized presentation-mode navigation chrome across both June 2026 decks — left/right arrows advance/back as before, with a thicker blue progress bar at the top.

## 2026-07-08
- Library: added Urbanomics on the AI trade to Money/News/Questions as reinforcement for the capex-ledger reading: downstream adoption over the next three years has to do real work before upstream margins and financing pressure decide the bubble question.

## 2026-07-05
- Library: added Rohit Krishnan on Fable/GLM model-market equilibria to Money/News/Questions as live reading-list material; not promoted to the talks yet.
- Talks/library: added Carlo Cordasco on what AI cannot yet measure as conceptual support for Amara’s law, productivity-statistics lag, and managed experimentation.
- Talks/library: added Ishan Bakshi on India’s missing long-term-growth foundations as non-causal India baseline evidence for services fragility, gig-work absorption, and AI shock measurement.
- Talks: updated both June 2026 decks with Ramp/Revelio firm-level AI spending and employment evidence, plus per-deck changelog appendix slides for August-readiness provenance.
- Library: added "A New Look at AI’s Impact on Jobs" (Kharazian, Simon & Stevens) to The work; firm-level Ramp/Revelio data makes AI employment effects a reorganization threshold, not a one-sign jobs story.

## 2026-07-03
- econofai begins. The hub: a ten-idea TL;DR, the six-part spine with the falling-price-of-intelligence chart embedded live, both June 2026 Takshashila talks preserved as delivered, a 55-source reading list with an outbound link on every entry, and six questions for the room.