Simon Wardley's Post — Industrialisation of ML: nine notes on what to expect
Simon Wardley · article · 2025-12 Reader 🖥️
AI Labor DisruptionCognitive DebtCompute & Energy InfrastructureIndustrialization of AIJevons ParadoxMeasuring AI ImpactTechnology Diffusion & Transformation Curves
What is happening with AI is the industrialisation of machine learning — a one-way evolutionary shift to commodity/utility — and the right posture toward it is neither resistance nor hype but disciplined adaptation through the predictable 5-8 year co-evolution of practices, with explicit attention to where bottlenecks, supply-chain power, and inequality will actually land.
practical · mixed — A nine-point operating directive ('NB.') addressed to readers about how to position themselves for an ongoing industrialisation event. Mostly expository-with-prescription: each numbered point states a claim about the climatic landscape and then issues an instruction ('expect X', 'don't expect Y', 'focus on Z'). Built on Wardley's mapping framework even though no map is drawn.
Structure
| Part | Role |
|---|---|
| Frame (point 1): 'industrialisation of ML' as the underlying phenomenon | Sets the climatic premise from which everything else is derived |
| Maturity timeline (point 2): 5-8 years of co-evolution, not stable before 2029 | Predictive scaffolding that calibrates expectations about errors, failures and practice change |
| Counter-narrative on jobs (point 3): rejects 'jobs apocalypse', introduces Jevons paradox via adjacent possibles and reframes the real problem as redistribution | Argumentative — pushes against prevailing AI-doom and AI-utopia narratives in one move |
| Adaptation imperative (point 4): tasks will change, individuals must evolve | Personal-level prescription bridging climatic claim and doctrine |
| Software engineering critique (point 5): much of 'engineering' is craft; much of 'architecture' is statements of belief; metrics like override rate and P&L impact are missing or unmonitored | Domain-specific evidence that current practice is immature relative to incoming scale (30M LoC → 1B+ LoC) |
| Barriers and squeeze (point 6): lower barriers to entry; un-industrialised SaaS margins under pressure | Industry-structure prediction derived from the evolution-axis claim |
| Bottleneck migration (point 7): comprehension and chain-of-understanding become the new constraint; automate ancillary services first | Specific prescription about where to spend effort now |
| Geopolitics (point 8): China to dominate over time because the state understands the AI supply chain | Long-horizon prediction grounded in Wardley's recurring supply-chain/doctrine thesis about Western strategic illiteracy |
| Market correction (point 9): correction is normal and not evidence of low underlying value | Insulates the climatic claim from being falsified by a financial downturn |
| Coda ('Buckle up Dorothy. Hold firm.') plus the reshared NineTwoThree '2025 AI fails' list | Tonal anchor and evidential prop — the reshared failure list is the concrete instantiation of point 2's 'lack of maturity in practice' |
Key terms
- industrialisation (of ML) Industrialization of AI
The evolutionary movement of a capability along Wardley's evolution axis from custom-built/product toward commodity/utility — making it cheap, ubiquitous, and substrate for new industries. Here applied specifically to machine learning rather than to 'AI' as a hype category. - co-evolution of practice Technology Diffusion & Transformation Curves
When a technology industrialises, the surrounding doctrine, tooling, metrics, and professional practices lag and then converge over a typical 5-8 year window. This is a Wardley climatic pattern, not a guess. - Jevons paradox (Wardley's reading) Jevons Paradox
Explicitly NOT just price elasticity or unmet demand. The real engine is 'adjacent possibles becoming real and expanding' — once a capability is cheap and ubiquitous, entirely new industries become buildable on top of it (componentisation effects). - componentisation Industrialization of AI
Once a capability is a commodity component, higher-order products and industries can compose it — this is the mechanism by which new industries emerge from an industrialised substrate. - engineering vs craft (in software) Industrialization of AI
Most current software development is craft — apprenticeship, taste, reading-code-to-understand — not engineering. TDD is one of the few genuinely engineering-shaped practices. Craft will not scale to billion-LoC codebases. - architecture (as currently practiced)
Much of what passes for architecture is 'statements of belief and wishes' rather than disciplined reasoning grounded in measurable consequences. - metrics that matter (override rate, impact on P&L) Measuring AI Impact
The right way to know whether AI-assisted development is actually working is to measure how often humans override AI outputs and how those outputs ladder up to P&L — most companies don't measure either. - chain of understanding Cognitive Debt
The end-to-end ability of humans (and human-AI teams) to comprehend what a system is doing and why. As code production accelerates, this comprehension chain becomes the new bottleneck. - supply chain (in the AI sense) Compute & Energy Infrastructure
The full stack required to produce industrialised ML — compute, chips, energy, materials, talent, training data, deployment platforms. State actors who understand this stack as a whole will outcompete those who do not. - doctrine (implicit)
Repeatable principles for how to act given a map — Wardley uses 'doctrine' in his framework to mean universal practices like 'focus on user needs', 'use appropriate methods', 'know your supply chain'. The China point relies on the claim that Chinese state strategy embodies more of this doctrine than Western governments.
Propositions
- explicitAI is best understood specifically as the industrialisation of machine learning, not as a generic technology wave.conf: high
- explicitcorroboratedPractices will not stabilise before 2029, because co-evolution of practice with industrialising technology typically takes 5-8 years.conf: highsupports ←Ordinary Engineers, Not Heroic Inventors · p8 — Same claim in different vocabulary: diffusion / co-evolution of practice with an industrializing technology takes years, not quarters.
- explicitContinued mistakes, errors, and failures should be expected and read as immaturity-of-practice signals, not as evidence the technology is wrong.conf: high
- explicitdevelopedThe 'AI jobs apocalypse' framing is wrong; industrialisation will create new industries via adjacent possibles (componentisation), not collapse the labour market.conf: highrefines ←You Don't Get to Create Anything (Still Burning: Beck + Randy Shoup on Jevons paradox and distributed systems under AI) · p2 — Accepts Wardley's no-jobs-apocalypse frame but adds the qualification that the transition still inflicts major displacement on current cognitive workers (agriculture analogy).
- explicitdevelopedThe real problem with technology industrialisation in the West is inequality and absence of redistributive mechanisms — not the technology itself.conf: highextends ←The Tidy House · p4 — Where Wardley diagnoses missing redistributive mechanisms as the real problem, O'Reilly extends this with concrete lab-level mechanism designs (subsidized tokens, ContentID-style splits).
- explicitReading code to understand a system will not survive the scale shift from ~30M to 1B+ lines of code.conf: high
- explicitMost of what is called software 'engineering' is actually craft, and most of what is called 'architecture' is statements of belief.conf: high
- explicitcontestedOverride rate of AI outputs and impact on P&L are the right metrics for AI-assisted development; almost no one is monitoring them.conf: highcontradicts ←Itchy Brain (Still Burning: Michael Grinich & Kent Beck on AI adoption, enterprise developer tools) · p8 — Wardley says P&L impact is the right metric for AI-assisted development; Grinich argues Red Queen dynamics make profit unusable as a measure because every competitor's baseline shifts too.
- explicitcorroboratedBarriers to entry will fall in many creative and SaaS-like fields; small teams will build blockbuster movies, games, etc., and SaaS players that haven't industrialised their offerings (for margin reasons) will get squeezed.conf: mediumsupports ←AI and the Looming Competition for Margin · p1 — Wardley's prediction that SaaS players who haven't industrialised their offerings will get margin-squeezed is an instance of exactly the 'same with less'/'more with same' squeeze this article frames as a general law of AI-driven productivity growth.
- explicitestablishedAs development accelerates, comprehension and the chain of understanding become the new bottleneck.conf: highsupports ←Presentation: The Free-Lunch Guide to Idea Circularity · p12 — Cummins's 'epistemic debt' (collapse of competence from shipping unread-code) matches Wardley's own claim that comprehension and the chain of understanding become the bottleneck as development accelerates -- same source, consistent framing.supports ←Five studies that are changing how I think about AI in software engineering · p2 — 'Generating code faster than the systems needed to understand it' is the same claim as Wardley's 'comprehension and the chain of understanding become the new bottleneck.'supports ←Presentation: The Future of Engineering: Mindsets That Matter When Code Isn't Enough · p4 — Corroborates Wardley's claim that comprehension becomes the binding bottleneck as development accelerates, from a hands-on practitioner's anecdotes rather than systems theory.
- explicitRight now, the highest-leverage automation work is ancillary services — legal, sales, etc. — not core engineering.conf: high
- explicitcorroboratedChina will become the dominant AI player over time because the Chinese state understands the AI supply chain in a way Western states largely do not.conf: mediumsupports ←Howard Yu's Post — American AI models went from 70% of usage to 30% in 12 months · p1 — Reaches Wardley's China-ascendancy conclusion by a different mechanism (cost/distribution here vs. state supply-chain mastery there).
- explicitA market correction in AI is normal and predicted, and should not be read as evidence that the underlying components lack value.conf: high
- implicitWestern governments are characterised by strategic illiteracy about their own technological and economic landscape — and this is a rarity-of-understanding problem, not a resource problem.conf: high
- implicitEvolution along the genesis→commodity axis is one-way; resistance is wasted energy.conf: high
- implicitThe framing battle ('industrialisation of ML' vs 'AI') is itself strategic — choosing the right description determines which playbook you reach for.conf: high
Arguments
Expect continued errors and failures, and read them as immaturity signals rather than as evidence against the underlying shift. strength: medium
Explicit premises:
- Practices co-evolve with industrialising technology
- Co-evolution typically takes 5-8 years
- We are only a few years in
Implicit premises:
- Industrialisation events follow recognisable climatic patterns (Wardley's framework)
- The 2025 AI-fails list (reshared) reflects practice-immaturity, not technology-failure
Inferences:
- Any 'AI has failed' narrative grounded only in 2025-2026 evidence is premature
- Investments in maturity-of-practice (metrics, doctrine, governance) have an unusually high return right now
The jobs apocalypse framing is wrong; the real problem is redistribution. strength: medium
Explicit premises:
- Industrialisation creates raft of new industries through componentisation
- Jevons paradox here operates via adjacent possibles, not just demand elasticity
Implicit premises:
- Historical industrialisation cycles are valid analogues for ML industrialisation
- Adjacent-possible expansion happens faster than labour displacement at the aggregate
Inferences:
- Policy attention should shift from 'protecting jobs from AI' to 'redistribution mechanisms'
- Frameworks built around individual obsolescence under-predict the new-industry creation rate
China will dominate AI over time. strength: medium
Explicit premises:
- China Gov understands its technological and economic landscape
- China Gov understands the AI supply chain
- China has positioned itself for the long term
Implicit premises:
- Understanding the supply chain is necessary AND sufficient (or near-sufficient) for long-run dominance
- Western governments will not close this understanding gap in time
- Doctrine and gameplay are the decisive variable, not raw compute or talent flow
Inferences:
- Western strategy debates that treat AI as fundamentally a US-vs-EU-regulation problem are mis-scoped
- Sovereignty-of-stack initiatives (energy, chips, models, data) matter more than the model layer in isolation
Engineering practice must overhaul both 'engineering' and 'architecture' or break under scale. strength: high
Explicit premises:
- Most development is craft, not engineering
- Much architecture is belief and wishes
- Codebases will grow from 30M LoC to 1B+
- Reading code to understand a system will not survive that scale
- Override-rate and P&L metrics exist but are rarely monitored
Implicit premises:
- Scale of code volume is the binding constraint (rather than e.g. modularity or compositional bounds)
- Engineering-shaped practices (TDD-like, measurement-driven) generalise to AI-generated code
Inferences:
- Investments in chain-of-understanding tooling are higher-leverage than further code-generation speed
- Override-rate dashboards and P&L attribution become new core engineering disciplines
Correction in the AI market is expected and orthogonal to underlying value. strength: medium
Explicit premises:
- What is being built are essential underlying components of future industries
- Markets routinely overshoot
- Correction and recovery is normal
Implicit premises:
- The components currently being built are the right components (not the wrong abstraction)
- Recovery to current-or-higher levels is the base rate after corrections of this kind
Inferences:
- Holders of the components survive; pure narrative-driven valuations do not
- Position-sizing during a correction is a doctrine question, not a market-timing question
Narratives
- The reshared 'Biggest AI Fails of 2025' list from NineTwoThree (VW Cariad $7.5B loss; Taco Bell drive-thru; Google AI Overviews hallucinations; $25.6M Arup deepfake heist; Replit agent wiping production DB; McDonald's/Paradoxai 64M records exposed; UnitedHealth/Humana algorithmic care denial; Earnest $2.5M bias settlement; Workday age-discrimination class action).
→ Operationalises point 2: each fail is an instance of immature practice (governance, evaluation, deployment, oversight) rather than a refutation of ML industrialisation. The reshare functions as evidence the post depends on rhetorically — 'this is what immature practice during industrialisation looks like'. - Implicit analogy to past industrialisation events (electricity, computing) where 5-8 year co-evolution of practices was observed.
→ Supports the 2029-stability prediction and the climatic-pattern claim without Wardley having to defend it explicitly in this post. - 'The small team that builds the next blockbuster movie, game, etc.'
→ Illustrates the lowering-barriers prediction in point 6. Functions as the canonical adjacent-possibles example. - 'Buckle up Dorothy. Hold firm.'
→ Tonal coda — frames the whole post as steadying advice for a turbulent passage, signalling doctrine over panic.
Questions addressed
- ✅ What is the right name for 'AI', and why does the name matter?
'Industrialisation of ML' — because this name imports a Wardley-mapping decision framework with known climatic patterns and timing. - ✅ When will practice stabilise?
Not before 2029, based on 5-8 year co-evolution norms. - ✅ Is the jobs apocalypse real?
No — the real problem is redistribution, not aggregate job destruction. - ✅ Where should organisations spend effort right now?
Automate ancillary services (legal, sales); invest in comprehension/chain-of-understanding; build override-rate and P&L metrics. - ✅ Who wins the long-run AI competition?
China, because of supply-chain literacy at the state level. - ✅ Is the AI market over-valued?
Yes, a correction is expected — but that is orthogonal to underlying value. - ⭕ By what concrete mechanisms should benefits be redistributed?
Named as 'the main problem' but left entirely unspecified. - ⭕ What are the actual replacement practices for 'reading code to understand a system'?
Wardley says current practice won't survive, but does not name what replaces it. - ⭕ What does 'understanding the supply chain' look like in operational terms for a Western state or enterprise?
Asserted as decisive but not operationalised.
Gaps
- incompleteNo mechanism is offered for the redistribution problem Wardley calls 'the main problem in this space'. The post diagnoses inequality precisely and then walks away from the prescription.
Read next: Read alongside political-economy work on technology and redistribution — Mariana Mazzucato on entrepreneurial state, Daron Acemoglu's recent work on AI and labour share, Carlota Perez on technological surges and golden-age policy regimes. Also Frederiksen's Coop Bank piece for the sovereignty/redistribution angle at the firm and national level.closed byThe Tidy House — Supplies the missing prescription: concrete value-distributing mechanisms (subsidized tokens for community work, ContentID-style revenue splits) that labs could build for the redistribution problem Wardley diagnoses but leaves without a mechanism. - incompleteThe 'chain of understanding' is named as the new bottleneck but not operationalised. What tools, practices, or organisational shapes produce or sustain comprehension at billion-LoC scale?
Read next: Read against the harness-engineering corpus you have this session (Anthropic harness paper, Cloudflare Mythos, Pocock, smallcode). Bernhard's value-stream framing is a candidate operationalisation: comprehension as a stream-level property, not an individual one. Also Donella Meadows on bounded rationality and information delays — the chain of understanding is a feedback-loop integrity problem.closed byPresentation: The Future of Engineering: Mindsets That Matter When Code Isn't Enough — Greene operationalizes 'protecting comprehension' with two concrete practices: a review-time rule that code must be rewritten until a reader's cognitive load is low ('after correctness, think always of the reader'), and a dedicated public Slack channel that normalizes admitting non-understanding once informal in-person modeling of curiosity was lost to remote work. - illogicalThe China-dominance argument compresses 'supply chain understanding' into a near-sufficient condition for long-run dominance. Other variables (capital allocation efficiency, demographics, talent migration, capability of domestic models, export-control regimes, energy supply, civil-military fusion costs) are not weighed.
Read next: Read against Dan Wang on Chinese tech policy, CSIS/RAND supply-chain analyses, and any sceptical takes on PRC industrial policy execution (e.g. semiconductor self-sufficiency progress vs targets). - incompleteThe Jevons-paradox reframing (adjacent possibles becoming real) is asserted but the conditions under which adjacent possibles actually become real industries — vs stalling — are not specified. Sometimes the adjacent possible is real and stays small.
Read next: Read against the slower-than-expected industrialisation histories — early electricity (the 40-year productivity paradox), enterprise software in the 1990s. Beck and Shoup on Jevons-in-coding is a sharper instantiation of this same claim worth juxtaposing. - uninformed (mild)The post under-weights regulatory ratchets — particularly EU AI Act, sector-specific liability regimes, GDPR-style data sovereignty rules — as climatic variables that shape what 'industrialised ML' actually looks like in different jurisdictions. Wardley's mapping treats regulation as doctrine; this post treats it as background.
Read next: Read alongside EU AI Act implementing acts, NSM/Datatilsynet guidance for the Norwegian context, and Frederiksen on Coop Bank sovereignty for how an enterprise actually reasons about regulatory ratchets as architectural constraint. - incompleteThe post is silent on the cultural and organisational variation in how 'practice converges' over the 5-8 year window. Nordic public-sector institutions, Japanese manufacturers, and Silicon Valley startups will not converge to the same practice — but the post talks as if there is one convergence trajectory.
Read next: Read against Cutler on organisational operating systems and the Heis stable's Norwegian-leader interviews — concrete evidence that practice-convergence is mediated by national culture and institutional shape, not just by elapsed years.
Related documents
- Presentation: The Free-Lunch Guide to Idea Circularity
- AI and the Looming Competition for Margin
- How Organizations Use AI: Evidence From ChatGPT
- Itchy Brain (Still Burning: Michael Grinich & Kent Beck on AI adoption, enterprise developer tools)
- 46 thoughts on the near future
- The 8 software engineering metrics AI broke
- Agentic Code Review
- AI coding creates two kinds of debt. You're only measuring one
- AI Enthusiasts Are in a Race Against Time, AI Skeptics Are in a Race Against Entropy
- AI's environmental costs threaten water, land and climate