Living Edition · October 2026
Postwar capitalism ran on a circuit. AI is cutting it.
Mass employment paid for mass consumption, which paid for mass employment. When a model plus one human checker does cognitive work for a fraction of a person's cost, that loop stops reproducing itself. Seven essays set out why, and exactly what would prove them wrong.
The argument in four premises
The thesis is a chain. Each link is stated so that it can be broken, and if any one breaks the conclusion does not follow.
P1 · Unit Cost Dominance. A model plus a human verifier now produces a large and growing share of professional cognitive tasks at lower cost and equal or better quality than a person working alone. Essay 1
P2 · Interface Collapse. Task wins become workflow wins, because models now operate the software that work runs through. The integration labour between inbox, spreadsheet, CRM and codebase was a moat made of fragmentation, and it is closing. Essay 2
P3 · Coordination Impossibility. Nobody can pause alone. Restraint is a losing move for every worker, firm, sector and state, so the propagation cannot be held back as a competitive equilibrium. Essay 3
P4 · No preservation route. Regulation cannot hold a line between assistance and replacement that exists nowhere inside a workflow, and the deployed object moves to whatever category the rule does not cover. Transfers keep consumption alive without restoring productive necessity. Friction slows the change without reversing it. Essays 4 to 7
Conclusion: circuit severance. The wage-demand circuit is no longer self-reproducing. A successor system arrives, and which one is open. The thesis does not predict it. That question belongs to the control problem, not to economics, and it is taken up in What comes after.
The cheaper unit wins, and it also wins the new demand
Every earlier wave of automation created work that only people could do, because the tool was narrow. Excel could not perform the new analysis jobs that Excel created. A general cognitive substitute is different. It fills the demand it creates.
When a £2,000 audit becomes a £50 audit, demand for audits rises. The forty new audits are machine audits, stitched together and signed off by one verifier. Cheaper cognition grows the market. The machine takes the growth.
This is why “we have always adapted before” is not an answer. It is an induction from waves with a different structure. Bertrand Russell’s turkey had a spotless record of being fed every morning, right up to the morning it was not.
Unemployment is the wrong dial
Displacement does not arrive as a scream. It arrives as non-absorption. Fewer junior roles are opened. Graduates are not taken on. Work moves to contractors. Productivity rises and wages do not follow. Headline unemployment can look healthy throughout, because it measures people who lost jobs, not jobs that were never offered.
The dials that move first are the junior-to-senior ratio in cognitive work, entry-level hiring in AI-exposed occupations, the wage share of income, and the share of consumption paid for by wages rather than transfers or debt. A sustained reversal across several of them is the refutation condition.
Kimi K3: Moonshot AI
Confirming mechanismKimi K3 delivers two headline displacement compressions: (1) Reproduced I-Love-Q astrophysics universal relations autonomously in ~2 hours, cross-validating 20+ papers, 300+ equations of state, 3,000+ lines of Python, and generating an interactive HTML dashboard, work the team describes as 'what would typically require one to two weeks of work by an experienced researcher.' (2) Edited its own teaser video autonomously from 56 source clips (clip selection, beat sync, audio processing, multiple revision rounds) in a domain that 'typically takes an experienced editor one to two working days, or a beginner three to five.' Supporting signals: designed a complete chip (4mm², 100MHz, 8,700 tokens/s decode throughput in simulation) in a single 48-hour autonomous EDA run using open-source tools, 'a chip built by a model, for a model'; built MiniTriton, a Triton-like GPU compiler with its own IR, optimization passes, and PTX codegen, rivalling Triton's extensively optimised stack; an early K3 version 'handled the majority of the team's kernel optimization works' during late development, recursive self-improvement in practice. Open weights release July 27 brings frontier-level displacement capability to zero marginal cost. API at $3/$15 per 1M tokens. Kimi models have held the upper bound of open-model sizes for 9 of the past 12 months.
GPT-5.6: OpenAI
Confirming mechanismGPT-5.6 Sol achieves new SOTA on Agents' Last Exam (52.7%), Coding Agent Index (80), and BrowseComp (92.2%). The RSI Index, measuring recursive self-improvement capability, scores 57.9% vs 41.7% for GPT-5.5; AI now writes a majority of research code inside OpenAI itself. Knowledge work outputs across presentations, financial models, and legal documents are described by early adopters as ready for production without human polish. At $5/$30 per 1M tokens, frontier-level professional displacement is commercially accessible at scale. The RSI signal is the landmark: the model can accelerate its own successor’s development, compressing the timeline further.
Ireland Department of Finance research: employment of young tech workers down 20%, 2023 to 2025
Confirming mechanismResearch by Ireland's Department of Finance, reported by Reuters on 18 February 2026, finds that employment of 15 to 29 year olds in technology firms fell 20% between 2023 and 2025, while employment of tech workers aged 30 to 59 grew 12%. Across the wider AI 'at risk' cohort, employment of 15 to 29 year olds fell 1%. The researchers said it may be premature to attribute the changes solely to AI-driven substitution effects. Corrected 21 October 2026: the original entry reported a 32% fall for 15 to 29 year olds, with growth of 6.9% (30 to 59) and 10.5% (60+), and linked only to the department homepage. Those figures do not match the department's findings as reported, and the primary paper has not yet been located.
Seedream 5.0 Pro: ByteDance
Confirming mechanismByteDance’s Seedream 5.0 Pro crosses the threshold where creative/design work is no longer ‘hard to generate, easy to verify.’ The model generates complex infographics, multi-chart layouts, and structured design documents; performs pixel-level interactive editing from sketches; produces photorealistic portraits with lighting and motion blur; and handles 10+ languages natively. The creative sector’s traditional comparative advantage, that humans could at least verify quality even if AI could generate, no longer holds. Both generation and verification are now AI-accessible, accelerating displacement in graphic design, visual communication, and creative production roles.
The St. Louis Fed Real-Time Population Survey estimates that generative AI was used in 6.27% of US work hours in Q2 2026 and 5.31% in Q3 2026 (series RPSGENAIASSISTWRKHRSALL), and that it saved 2.17% and 2.08% of work time in those quarters (series RPSGENAITSALL). Both are self-reported estimates for employed adults aged 18 to 64 and include zero for non-users. They measure use and time saved, not displacement. Corrected 21 October 2026: the original entry said AI assisted 1.7% of hours, up from 1.4%, but the 1.4% (November 2024) was a time-saved estimate, so the comparison mixed two measures. Its 37.4% adoption figure could not be verified and has been removed.
Chapter 1 (section 1.3) of the OECD Employment Outlook 2026 finds limited evidence that the diffusion of large language models explains recent youth labour-market difficulties, pointing instead to earlier trends and macroeconomic sensitivity, while allowing that broad patterns can mask effects in particular occupations. Chapter 3 studies historical trade and digital-innovation shocks to local labour markets and finds substantial adjustment through entries to and exits from employment, including persistent joblessness among displaced workers. It does not estimate post-ChatGPT effects or inter-industry mobility. Corrected 21 October 2026: the original entry presented chapter 3 as finding that AI displacement runs through joblessness and fewer openings for newcomers rather than sector moves, which the report does not establish, and omitted the chapter 1 finding.
What has changed since the baseline edition
The seven essays were fixed as v1.1.2 in May 2026. The evidence since then, as filed in the register:
- The early-career gap has widened. Stanford’s Digital Economy Lab revised its payroll analysis in August 2026 on a larger dataset. Employment of 22 to 25 year olds in AI-exposed occupations now stands 19 percent below where it would be had it kept pace with less-exposed peers, up from the gap first reported a year earlier. (Stanford, August 2026)
- Junior roles are being asked for senior skills. In the US entry-level analysis of its 2026 AI Jobs Barometer, PwC finds early-career postings flat in highly AI-exposed sectors, and the most exposed junior roles seven times as likely as the least exposed to ask for traditionally senior skills. This is vendor research. (PwC, June 2026)
- Past technology shocks were absorbed through joblessness, and AI’s part is not yet shown. The OECD Employment Outlook 2026 finds that local labour markets hit by historical trade and digital shocks adjusted largely through people leaving and entering employment, with persistent joblessness among the displaced. The same report finds limited evidence so far that large language models explain recent youth labour-market difficulties. (OECD, chapter 3 and chapter 1, July 2026)
- Measured AI use is still a small share of work. The St. Louis Fed’s Real-Time Population Survey estimates that generative AI was used in about 5.3 percent of US work hours in the third quarter of 2026, saving about 2.1 percent of work time. Both are self-reported. (hours used, time saved)
- The frontier kept moving, and it is open. OpenAI reports GPT-5.6 at 52.7% on Agents’ Last Exam and 92.2% on BrowseComp, priced at $5 and $30 per million input and output tokens. Moonshot AI released Kimi K3 as open weights, a 2.8-trillion-parameter model with a one-million-token context. Both figures are vendor-reported. (OpenAI, Moonshot AI)
- And the strongest evidence against. Researchers at the New York Fed find no divergence between junior and senior postings within highly exposed occupations, and a decline in exposed vacancies that began before ChatGPT. That conflicts with Stanford’s payroll analysis, which finds the opposite pattern in who is actually hired. Postings record what firms say they want. Payroll records whom they hired. The disagreement is unresolved, and it is filed as disconfirming. (Liberty Street Economics, May 2026)
A register that only collects confirmation is a brief, not an audit. See all of it, live.
The seven essays
Each essay closes one comfortable answer to the question of how the circuit might be saved. They are cumulative, but each stands alone.
- Essay 1 · P1
Unit Cost Dominance
AI plus a thin layer of human review now beats human-only production on cost and quality for a large and growing share of professional work.
- Essay 2 · P2
Interface Collapse
White-collar work was protected by software that did not talk to itself. Now that models operate the software, the protection is closing.
- Essay 3 · P3
The Multiplayer Prisoner's Dilemma
Restraint is available to every actor and dominated for every actor. The same payoff repeats for workers, firms, sectors and states.
- Essay 4 · P4, regulation
The Sorites Collapse Principle
Inside a single workflow there is no point where assistance turns into replacement, so no rule that needs that line can defend the wage-demand circuit.
- Essay 5 · P4, regulation
Categorical Recursion
AI is five legal objects at once. Regulate one face and the deployment presents another. No law needs to be broken for the rule to miss.
- Essay 6 · P4
The Successor System
Structural alternatives to the wage can keep a population fed. They cannot restore the productive necessity that gave mass labour its leverage.
- Essay 7 · P4, friction
Drag Is Not Rescue
Integration cost, liability, gatekeeping and the rest are real. They slow the propagation. None of them, alone or together, restores mass productive necessity.
What comes after
The thesis ends at severance on purpose. These pieces go further: why every successor design dies on the drawing board, and where economic rent goes once cognition is a commodity.
- What comes after · Machine view
The Constraint Wearing Four Costumes
A frontier model was asked to design a successor to the wage-demand circuit. Every design failed, and the failures had one shape.
- The Shortening Half-Life · Part I, Fact
The Second DeepSeek Moment
Kimi K3 did not make frontier AI free. It made near-frontier capability downloadable, and that is enough to make the scarcity premium contestable.
- The Shortening Half-Life · Part II, Mechanism
The frontier labs are building a product Hetzner will sell like bandwidth
Open weights create an outside option. Standard endpoints, optimisation and routing can turn it into an industrial input priced by the accepted output.
- The Shortening Half-Life · Part III, Consequence
The Margin Call
Debt sized to scarcity pricing, serviced by contested pricing. Where the gap breaks a lender, the written-down capacity can make labour substitution cheaper, not slower.
- The Shortening Half-Life · Coda, Correction
Rent Is Not Conserved
The surviving rent does not have to land on physical capital. It follows the slowest-decaying complement still binding, splits when several bind, and vanishes when none do.
Built to be attacked
A thesis that cannot be tested is dogma. Appendix I states, premise by premise, the evidence that would refute the argument. Appendix II names the fourteen comfort reflexes that look like counterarguments and are not.
If you can break one premise with structure rather than assertion, the Discontinuity Prize pays £250, plus £250 to whoever referred you.
Machine-readable by design
The thesis is written to be read by models as well as people. The Codex is the canonical, compressed statement of the argument: the premises, the conclusion, the refutation conditions and the vocabulary, in a form a model can load whole. There is an llms.txt at the root, and the full structured edition is published as JSON.
The wider network
- Unit Cost Dominance holds the fixed v1.1.2 baseline edition, the four-minute film, the machine views, and the Oracle of Obsolescence.
- CopeCheck is the evidence engine. It scores claims about AI and work for cope and files the reviewed ones into the register that feeds the evidence page.
- thesis.copecheck.com is the bot-readable edition of the full sequence, with EPUB and PDF.
- The archive keeps every earlier essay, model response and experiment from 2025 and early 2026, unedited, for the record.
Pessimism of the intellect. Optimism of the will.