Presentation: The Free-Lunch Guide to Idea Circularity
Holly Cummins · presentation · 2026-07-31 source ↗ Reader 🖥️
AI Labor DisruptionBurnout & RecoveryCognitive DebtIdea Circularity / Tradeoff Re-TiltingJevons ParadoxTechnical DebtVC Funding Incentive for Hype
Ideas in technology that look genuinely new are almost always old tradeoffs returning because the surrounding environment has re-tilted the balance, not novelties or evidence of forgetting -- and the same 'debt is a tradeoff with the future' logic that explains financial debt also explains technical, epistemic, ecological, and sleep debt, so recognizing which tradeoff underlies a current design is the transferable skill for spotting when a discarded old idea is about to become valuable again.
theoretical · mixed — Primarily theoretical/expository: the thesis is descriptive (why old tech ideas recur), built from a chain of historical narratives (sewage, grid computing, COBOL, ZIRP, Jevons Paradox) rather than deductive argument. It closes with a practical coda (be curious, watch for shifted assumptions, look for discarded ideas worth reviving) but that coda is secondary to the diagnostic claim.
Structure
| Part | Role |
|---|---|
| The Great Stink (1858 Thames sewage crisis) opening narrative | context/founding example |
| Grid-to-cloud and the general catalogue of returning tech ideas ('history rhymes') | evidence establishing the recurrence pattern |
| Carbon, data centers, and the tradeoff of green energy's own environmental costs; LightSwitchOps and efficient software (Quarkus vs. Spring Boot) as reduction levers | case study applying the tradeoff/re-tilting framework to sustainability |
| The dynamic-vs-static, centralized-vs-distributed architectural pendulum (Java reflection, microservices vs. monolith, Git vs. GitHub, Apple's on-device AI vs. hyperscaler data centers) | extended case study showing the same tradeoff re-tilting across four decades of infrastructure choices |
| Hype cycle mechanics: VC funding-round incentives (Meredith Whittaker quote) as the causal explanation for why hype recurs regardless of merit | argument for a causal mechanism behind the recurrence pattern |
| 'World without developers' predictions (COBOL, low-code, AI) debunked via Jevons Paradox and the knocker-up counter-example | argument applying the framework to predict developer jobs will not disappear |
| Job-market data reframed as interest-rate-driven (post-ZIRP) rather than AI-driven, plus 'AI washing' of layoffs | evidence/argument reattributing a widely-assumed AI causal story to financial debt dynamics |
| 'World without rest' instead of the promised 'world without work': sleep debt, burnout, default mode network, AI intensifying work hours | extends the debt framework to the human/physiological domain |
| Return of engineering discipline (TDD, specs) as a response to 'epistemic debt' (Wardley) | case study of a discarded idea (rigor/discipline) becoming valuable again once the tradeoff re-tilted |
| 'Connecting the dots' closing section: explicit statement that debt = tradeoff with the future, plus practical advice (ask what problem/tradeoff, be curious, look for shifted assumptions) | conclusion / explicit statement of unity |
Key terms
- Debt (general)
Any deferral of a cost into the future in exchange for a present benefit -- the single pattern Cummins claims unifies financial, technical, epistemic, ecological, and sleep debt. - Epistemic debt
The collapse of competence that results from shipping code (or systems) nobody understands; attributed by Cummins to Simon Wardley. - ZIRP (zero interest-rate period/phenomenon)
The roughly 15-year (2010-2023) stretch of persistently low (not literally zero, but under 1%) interest rates, treated as a historical anomaly rather than the norm, whose ending reshaped investor and hiring behavior. - AI washing
Attributing layoffs or business news to AI success/strategy as cover, when the real driver is overhiring, a weak core business, or investor FOMO. - Idea circularity / 'nothing new under the sun'
The phenomenon of old technology ideas returning to prominence -- explicitly reframed by Cummins as tradeoffs re-tilting due to a changed environment, not as history literally repeating or people forgetting. - Free lunch
An architectural or technology choice marketed as a pure win (more scalability, more decentralization, more autonomy) that in fact defers a real cost -- the title metaphor for every case study in the talk.
Propositions
- explicitTechnology ideas that appear genuinely new are typically old ideas returning because a tradeoff has re-tilted due to a changed environment, not because engineers forgot history or ideas are inherently cyclical.conf: high
- explicitDebt, as a general category, is a tradeoff with the future: a benefit taken now in exchange for a cost that comes due later.conf: high
- explicitThe current elevated cost of financial debt is the return to historical normal interest rates after a roughly 15-year zero-interest-rate period (ZIRP) that was itself the anomaly, not the baseline.conf: high
- explicitPost-ZIRP higher interest rates make safe government bonds relatively more attractive to investors than risky software-company equity, which reduces investment in software and, in turn, reduces software hiring -- a mechanism Cummins presents as a bigger driver of the recent tech job downturn than AI.conf: highextendsThe loudest warning about AI and jobs yet · p7 — Cummins supplies an explicit causal mechanism (bond-yield vs. equity-risk tradeoff for investors) for why interest-rate hikes -- one of several overlapping causes Newton's piece flags for job losses -- reduce software hiring specifically.
- explicitDeveloper job vacancies are recovering in the latest data even while overall hiring stays flat, which Cummins reads as evidence that the earlier vacancy collapse tracked interest-rate timing rather than AI-driven displacement (ChatGPT's launch produced no visible blip in the job figures).conf: medium
- explicitcorroboratedHype is a structural requirement of venture-funded business models, not an accident of any one technology: investor excitement, and confidence that others share it, is what attracts the follow-on funding rounds a startup needs to reach an exit, so technologies get inflated independent of their actual merit.conf: highsupports ←Everyone I recommend on AI is optimistic. Except Cory Doctorow. · p1 — Doctorow's claim that most AI discourse is 'a sales pitch aimed at investors' is the same structural mechanism Cummins states directly: hype is a structural requirement of venture-funded business models, independent of the technology's actual merit.
- explicitPredictions of 'a world without developers' (COBOL, low-code, and now AI) have consistently failed, and by Jevons Paradox each new layer of abstraction has historically increased the number of developers rather than decreased it, because cheaper software production increases total demand for software.conf: highsupportsSkill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI · p16 — Independent statement of the same Jevons-Paradox-for-developer-jobs claim, via different historical evidence (COBOL/low-code vs. contemporary agentic coding) -- though this corpus already flags the underlying disanalogy (task-level vs. occupation-level automation) as unaddressed in both.supportsYou Don't Get to Create Anything (Still Burning: Beck + Randy Shoup on Jevons paradox and distributed systems under AI) · p1 — Same Jevons Paradox claim applied to cognitive/software labor -- cheaper cognition/code production increases total demand rather than reducing headcount.
- explicitSome recent tech layoffs attributed to AI are better explained by 'AI washing' -- disguising ordinary causes (ZIRP-era overhiring, a weak core business, investor FOMO) as an AI success story, since companies with an AI product need to show it 'works' by cutting their own staff.conf: medium
- explicitEfficient software design (avoiding unnecessary runtime dynamism/reflection, as in Quarkus versus Spring Boot) measurably correlates with a better carbon footprint at a given throughput, making software engineering choices a direct lever on climate impact.conf: medium
- explicitEven zero-carbon/'green' energy sources carry significant non-carbon environmental costs (e.g. dam-driven ecological and even geophysical disruption), so decarbonizing the energy supply cannot be the sole climate strategy -- reducing consumption (e.g. 'LightSwitchOps,' turning idle systems off) must also be pursued.conf: high
- explicitdevelopedSwings between dynamic/centralized and static/distributed architectures (Java reflection vs. ahead-of-time compilation, monolith vs. microservices, decentralized Git vs. centralized GitHub, on-prem vs. cloud, hyperscaler AI data centers vs. Apple's on-device inference) are not one-directional technical progress but a recurring tradeoff that re-tilts whenever the deployment environment changes.conf: highextends ←Azure Virtual Desktop Hybrid Reaches GA with Licensing Details Unpublished · p2 — Cummins's claim frames on-prem-vs-cloud as a recurring, non-directional architectural pendulum that re-tilts with deployment constraints, not one-way progress toward the cloud. Azure Virtual Desktop Hybrid is a concrete, dated (2026) instance of exactly that re-tilt — a hyperscaler shipping a product whose whole premise is that some organizations' constraints currently favor pulling compute back toward customer-owned hardware.
- explicit'Epistemic debt' -- the collapse of competence from shipping code nobody understands, a term Cummins attributes to Simon Wardley -- is driving a partial return of pre-AI engineering discipline (TDD, incremental commits, written specs).conf: highsupportsSimon Wardley's Post — Industrialisation of ML: nine notes on what to expect · p10 — 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.
- explicitCurrent AI-driven productivity gains have produced a 'world without rest' rather than the historically promised 'world without work' (Keynes's 15-hour week): professions using AI report working roughly 3 hours more per week than before, an instance of Jevons Paradox applied to time and cognition.conf: highsupportsAI productivity is burning out your best engineers · p7 — Both describe AI's apparent productivity gains as partly illusory -- borrowed against increased working hours / uncredited human validation labor rather than representing a clean net gain.
- explicitUnstructured rest has independent cognitive value (the brain's default mode network becomes more active and helps solve problems during idle time), evidenced historically by Henry Ford's move to a 5-day week without a drop in productivity -- supporting the idea that 'sleep debt' is a real, not merely rhetorical, cost.conf: highextendsDaniel Pink's Post · p3 — Adds a mechanism (default mode network) and a historical proof point (Henry Ford's 5-day week preserving output) to the claim that burnout comes from the brain never getting to stop, not from stress per se.
- explicitCode is a liability rather than an asset -- even good, well-tested code adds to technical debt merely by existing -- and cheap AI-assisted code generation ('vibe now, pay later') is accelerating that liability's accumulation into a materializing cost, evidenced by Amazon reportedly rolling back agent-authored code and requiring senior sign-off after outages.conf: medium
- implicitBecause all five discussed debts (ecological, financial, technical, epistemic, sleep) are unified only at the level of a shared verbal definition ('a tradeoff with the future'), the talk implicitly claims the same corrective posture -- watch for the tradeoff re-tilting -- transfers equally across all of them, even though the mechanisms generating each debt (compound interest, code entropy, understanding loss, sleep physiology) operate on entirely different causal logics, actors, and timescales.conf: medium
- implicitThe talk implicitly treats 'this is a repeat, not something new' as itself a form of insight superior to naive novelty claims, without testing whether the current AI moment could still contain a genuine discontinuity that would break the historical pattern the talk relies on.conf: medium
Arguments
Software developer jobs are not being eliminated by AI (nor were they by low-code or COBOL before it). strength: medium
Explicit premises:
- Historical predictions of developer extinction via a new abstraction layer (COBOL, low-code) have repeatedly failed to materialize.
- Jevons Paradox predicts that cheaper software production increases total demand for software and hence total developer demand.
- Recent job-vacancy weakness correlates in time with rising interest rates (ZIRP ending) rather than with AI's introduction: ChatGPT's launch produced no visible change in the vacancy curve, and vacancies are recovering in the latest data even as general hiring stays flat.
Implicit premises:
- Demand for software is effectively insatiable in a way that demand for a single-purpose service (e.g., a human alarm clock) is not, so productivity gains can never saturate demand the way they did for eliminated professions.
- A timing correlation between interest-rate changes and vacancy changes is sufficient to assign the dominant causal weight to interest rates rather than AI, without ruling out AI as a lagged or confounded contributing factor.
- AI is not qualitatively different from COBOL or low-code as an abstraction layer -- i.e., historical base rates for 'abstraction layers don't eliminate developers' reliably extend to this new case.
Inferences:
- If total software demand really is insatiable, even large per-unit productivity gains from AI could never reduce aggregate developer headcount, only reshape its composition -- a stronger and more specific claim than the vacancy-timing evidence actually establishes.
Financial, technical, epistemic, ecological, and sleep debt are fundamentally the same phenomenon, and the discipline for anticipating when an old, discarded idea will matter again is identical across all of them: identify the underlying tradeoff and ask whether the environment has shifted. strength: low
Explicit premises:
- All five domains involve deferring some cost into the future in exchange for a present benefit.
- In each domain, Cummins supplies a historical example where the 'debt' was taken on for good short-term reasons and later came due when circumstances changed (ZIRP ending, AI-generated code compounding, Wardley's epistemic debt, Industrial-Revolution overwork).
Implicit premises:
- A shared verbal definition ('a tradeoff with the future') is sufficient grounds to treat structurally and causally different phenomena as instances of one underlying pattern for practical, predictive purposes.
- The recommended remedy (watch for the tradeoff re-tilting) is equally actionable and sufficient across financial markets, codebases, individual/team understanding, and human sleep physiology, despite those systems having entirely different feedback loops, actors, and timescales.
Inferences:
- Taken literally rather than rhetorically, the unification would imply that epistemic/cognitive debt could in principle be 'serviced' the way financial debt is (paid down with interest, refinanced, defaulted on) -- but other extractions in this corpus (e.g. the 'triple debt' framework) argue cognitive/epistemic debt has no such repayment mechanism, which undercuts the strength of treating the five debts as truly parallel.
Narratives
- The Great Stink: London's 1858 sewage crisis, caused by a well-intentioned earlier law eliminating cesspits in favor of piping waste into the Thames, which created a throughput/scaling problem once the population grew and the summer heat compounded it.
→ Establishes the core pattern that a 'good idea' (removing cesspits) creates a tradeoff (centralized stink/disease vs. distributed stink/disease) whose costs only surface once external conditions (population, heat) shift -- the template for every later example in the talk. - Grid computing disappeared, then returned relabeled as 'cloud.'
→ A compressed, memorable instance of the idea-circularity thesis: the same underlying capability came back under a new name once conditions favored it again. - Amazon Prime Video publicly announced it saved 90% of its costs by moving from microservices back to a monolith.
→ Supports the 'free lunch' framing: it isn't evidence microservices are simply bad, but evidence that microservices' promised benefits came with a real, sometimes hidden, price -- 'the lunch was not free.' - Apple's strategy of licensing AI models cheaply and putting AI capability into on-device hardware rather than building its own $100 billion data centers.
→ Maps the current AI infrastructure debate onto the same centralize-vs-decentralize tradeoff already seen in Git-to-GitHub and grid-to-cloud, reinforcing the idea-circularity thesis with a live, current-cycle example. - The 'knocker-up' profession (a human alarm clock during the Industrial Revolution, which persisted into the 1970s in parts of England before disappearing).
→ Functions as a contrast case, not a supporting example: it implicitly argues why software work is unlike professions technology genuinely eliminated -- demand for a wake-up call is finite (once per person per day) while demand for software is not. - The Three Gorges Dam moved so much water that it measurably lengthened Earth's day (by 0.06 microseconds) and tilted its axis.
→ A vivid illustrative aside (not tightly load-bearing) supporting the claim that even large, ostensibly 'green' infrastructure choices carry real, sometimes startling, physical costs. - Greenspun's Tenth Rule is revealed to have been written by Philip Greenspun at age 29, not as an old-and-experienced observer.
→ Undercuts Cummins's own first-instinct answer to 'how do you get better at seeing repeating patterns' (her initial answer was 'be old'), redirecting the advice toward asking 'what problem, what tradeoff' rather than relying on accumulated age/experience.
Questions addressed
- ✅ Why do old technology ideas keep coming back?
Because the underlying tradeoff re-tilts when the surrounding environment (cost structure, scale, regulatory context) changes -- not because people forgot the idea or because history is inherently cyclical. - ✅ Is AI going to eliminate software developer jobs?
No, provisionally -- based on historical base rates (COBOL, low-code), Jevons Paradox, and a reinterpretation of recent job-vacancy data as interest-rate-driven rather than AI-driven. - ⭕ Whether AI could be a qualitatively different abstraction layer than COBOL or low-code, such that the historical 'developers don't go away' pattern might not hold this time.
- ✅ Why are so many company layoffs currently blamed on AI?
Partly genuine AI capability, but significantly 'AI washing' -- disguising ordinary causes (ZIRP-era overhiring, a weak core business) as an AI success story, plus investor-FOMO-driven imitation. - ✅ How does an individual engineer get better at spotting these repeating tradeoffs?
Ask what problem is being solved and what tradeoff underlies the current design; watch for changed assumptions; be curious about both the past and the present -- explicitly not 'be old' (rejected via the Greenspun-was-29 story). - ⭕ Is the current rate of technical/epistemic debt accumulation from AI-generated code sustainable, and how does it get paid down?
Cummins presents evidence the debt is already 'coming due' (Amazon's reported agent moratorium after outages) and that discipline is 'coming back,' but does not specify a repayment mechanism or a criterion for when the debt is under control. - ⭕ Are financial, technical, epistemic, ecological, and sleep debt actually the same kind of phenomenon, or only verbally similar?
Gaps
- illogicalThe five-way 'debt' unification (financial, technical, epistemic, ecological, sleep) elides real differences between the domains. Financial debt has a mechanistic, quantifiable repayment schedule (compound interest) enforced by a counterparty; technical/cognitive/epistemic debt, per other extractions in this corpus, has no equivalent repayment mechanism and is only partially, informally repairable (e.g. via retrospectives or deliberate re-implementation, not a 'payment'); sleep debt is a biological phenomenon recoverable through actual rest, a third distinct dynamic again. Treating all five as instances of 'the same underlying pattern' does more rhetorical unification work than analytical work -- memorable, but it risks implying a false interchangeability (e.g. that principles for managing financial debt, like refinancing, transfer to epistemic debt, which they plausibly do not).
Read next: extract-triple-debt-storey and extract-ai-coding-two-kinds-debt, which analyze cognitive/technical/intent debt specifically and explicitly note debt-type asymmetries (e.g., cognitive debt 'has no repayment mechanism' unlike technical debt) -- read against Cummins's talk to test the limits of her unification. - incompleteCummins argues the current tech job downturn is mostly interest-rate-driven (ZIRP ending) rather than AI-driven, using a single aggregate vacancy-timing correlation as her main evidence. She does not address the disaggregated finding, present elsewhere in this corpus, that AI is already measurably eroding entry-level/early-career roles specifically even while aggregate figures look stable -- a granular pattern her aggregate graph would not detect.
Read next: extract-loudest-warning-ai-jobs, for the disaggregated (by seniority/AI-exposure) labor data that complicates the pure interest-rate explanation. - incompleteThe talk doesn't examine whether AI-driven computing efficiency could itself be subject to the same Jevons-Paradox rebound Cummins invokes to defend developer jobs -- i.e., if cheaper computation increases total computation demand, the same logic likely undercuts her own claim that efficient software (Quarkus) straightforwardly reduces carbon, since more efficient compute could simply induce more compute usage and offset the gains. This exact rebound-effect argument, applied to AI energy specifically, is developed elsewhere in this corpus and represents a direct, unaddressed tension with Cummins's efficient-software-is-climate-action argument.
Read next: extract-ai-environmental-costs-un, which argues efficiency improvements are unlikely to offset rising AI resource demand because of a rebound effect. - uninformedCummins frames 'epistemic debt' as a fresh insight heard from Simon Wardley, but doesn't connect it to the more developed 'triple debt' framework already circulating in the same discourse, which distinguishes cognitive debt (team-level loss of shared understanding) from intent debt (undocumented rationale/goals) as separate, causally-interacting debts rather than a single undifferentiated 'epistemic debt.' That finer-grained model would sharpen her own diagnosis of why 'shipping code you don't understand' is a problem and what specifically would fix it.
Read next: extract-triple-debt-storey and extract-ai-coding-two-kinds-debt for the more granular cognitive/intent-debt distinction.
Related documents
- Daniel Pink's Post
- Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI
- Fragments: February 13
- Presentation: The Future of Engineering: Mindsets That Matter When Code Isn't Enough
- Simon Wardley's Post — Industrialisation of ML: nine notes on what to expect
- The 8 software engineering metrics AI broke
- 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
- The "funhouse mirror": How AI reflects the hidden truths of your software pipeline
- AI productivity is burning out your best engineers