Insights · AI

Who gains the leverage, who compounds it, and the parts of the work a model cannot take from you.

Breno A. Beirigo

14 August 2026

Interpolation versus extrapolation

Interpolation repeats the familiar. Extrapolation creates identity.

Sci-fi classic Extrapolation

RoboCop (1987) 88%critics 84%audience Certified fresh

VS

Bland action thriller Interpolation

RoboCop (2014) 49%critics 49%audience Rotten

Interpolation tends toward the generic. Extrapolation creates identity.

The 2014 film is bland because it recombines familiar action-thriller conventions without establishing a distinctive new point of view. It sits near the genre average. The 1987 film makes the extrapolative leap: corporate satire, body horror, ultraviolence, and deadpan comedy.

This is an analogy for AI, not a claim that either film was generated by a model. Interpolation can produce polished novelty while remaining inside a familiar conceptual space. Extrapolation is the move that creates a new direction.

Present this as a heuristic about creative leaps, not a universal technical limit. Cronin explicitly allows that cross-domain training may complicate the distinction.

[Sources]

• Reception scores were transcribed from the supplied reference cards in assets/ai/robocop-1987.png and assets/ai/robocop-2014.png.
• The two scene illustrations use those character references and the Stoa Codex picture guide in prompts/visual-style-guide.md.
• The idea was encountered in a now-unavailable X post by Denis McDowell: https://x.com/mcdowell_is/status/1741840187716411417. A surviving reply identifies Lee Cronin’s claim that AI can interpolate but cannot extrapolate. The primary passage is Cronin’s discussion of “novelty and interpolation” in Lex Fridman Podcast #404; see also the episode transcript.

[/Sources]

AI is sea level

A shared baseline, not a summit.

↘ Over-reliance lowers the ceiling

Distinctive work moves toward average.

shared-model average

↗ AI raises the floor

Weak work moves toward average.

Use AI to reach the surface. Bring your own judgment above it.

Lulu Cheng Meservey, via @clipsofcrypto (n.d.)

A shared model gives everyone the same output for fractions of a cent. If you are below average that is a gain. If you believe you could be better than average, you are trading that possibility away (@clipsofcrypto, n.d.).

The Jevons paradox of AI

Figure 1: One aristocratic patron receives a concentrated AI light, while many smaller rays illuminate varied forms of human work.

Cheaper AI

Lower cost unlocks many more uses.

Jevons paradox

Each task costs less. Total AI use can still rise.

Expensive AI

Only a few uses justify the cost.

Jevons argued that more economical coal use could enlarge the field of profitable applications and increase aggregate consumption (Jevons 1865). The modern term is a rebound effect: lower operating cost induces more demand for the service (U.S. Department of Energy 2023).

For AI, treat the completed task as the service. Figure 1 contrasts a high-cost state, where only a few applications are economical, with a lower-cost state, where many more applications become viable. Total use can therefore rise even as each task becomes cheaper.

Present the slide in two phases. First, hold on the aristocratic patron while the right side remains covered. Advance once to cover the patron and reveal the many forms of work. The visual contrast introduces the Jevons mechanism before naming it.

Efficiency does not only save resources. It changes what becomes worth doing.

The analogy has limits. Compute prices, model capability, complementary human labor, integration capacity, and new use cases move together. A Jevons effect is therefore a possibility, not an automatic law.

[Sources]

• Jevons, The Coal Question, Chapter VII (Jevons 1865).
• U.S. Department of Energy, “Rebound Effect Analysis” (U.S. Department of Energy 2023).

[/Sources]

AI can create a Matthew effect

Experts use the same AI to frame, verify, and compound better decisions.

Same capability AI

Expert judgment Frames the real problem ⟶ Tests against the domain ⟶ Builds better next questions

Novice judgment Delegates the framing ⟶ Accepts plausible output ⟶ Builds shallow fluency

The gap is cumulative advantage

“To everyone who has, more will be given, and he will have an abundance. From the one who has not, even what he has will be taken away.” Matthew 25:29 · ESV, shortened

Same capability, two trajectories. The gap is not caused by access to the model — both have it.

Who gains the leverage

“AI won’t replace programmers, but rather make it easier for programmers to replace everyone else.”

Naval Ravikant

AI CAPABILITY × PROGRAMMING JUDGMENT ⟶ MORE WORK BECOMES PROGRAMMABLE

Naval’s line is a claim about where leverage lands, not about job losses. Capability alone does nothing until someone with programming judgment points it at a problem.

The opportunity cost of handwritten code

“Writing code by hand is no longer an economically viable skill for most programmers at most companies.”

David Heinemeier Hansson, Rails World 2026

The scarce work becomes deciding what deserves to ship.

The quotation states David Heinemeier Hansson’s argument at Rails World 2026. The Rails Foundation summarized 37signals’ position as “pencils down” on handwritten code (Heinemeier Hansson 2026; The Rails Foundation 2026).

This is a directional claim from one experienced practitioner and company, not a measured universal result. The economic comparison is between the extra value created by manual implementation and the opportunity cost of scarce human attention.

An anecdote often attributed to Milton Friedman makes the same economic point. At a canal project where workers used shovels instead of earthmoving equipment to preserve employment, Friedman supposedly replied that spoons would create still more work. The provenance is uncertain, and versions of the story predate the first located printed attribution to Friedman. Use it as an analogy, not as verified biography (Quote Investigator 2011). The narrow lesson is that manual effort is a cost, not proof of value.

Programming does not disappear. The work shifts toward specifying intent, choosing architecture, reviewing generated changes, testing behavior, and maintaining coherence. Handwritten code remains rational when agents fail, safety or accountability requires direct control, or writing the code is the learning objective.

[Sources]

• DHH, Rails World 2026 opening keynote (Heinemeier Hansson 2026).
• Rails Foundation summary of the keynote (The Rails Foundation 2026).
• Quote Investigator, “If You Want Jobs Then Give These Workers Spoons Instead of Shovels” (Quote Investigator 2011).

[/Sources]

Human slop preceded AI slop

Language is open source

Commodity content

Repeats what people already know.

Original content

Adds an idea that was not there before.

A human can write either.

This slide rejects the assumption that text becomes valuable merely because a human typed it. It does not claim that human and machine writing are identical. It separates the origin of the prose from the contribution embodied in it.

Cole describes language as shared infrastructure:

“Language is open source.”

The phrase is a metaphor. Letters, words, familiar structures, and most ordinary combinations of words are shared materials. It is not a claim about the legal status of a specific work. Copyright, trademark, and attribution remain separate questions.

Cole distinguishes commodity, personality, and original content. This slide compresses that three-part model into a two-part contrast so the audience can see the central distinction. Commodity writing repeats familiar ideas. Personality writing connects an idea to a specific person’s experience and point of view. Original writing changes how a reader understands the subject. AI did not create commodity writing. Humans were already producing it. AI made commodity production fast, abundant, and easy to recognize (Isenberg and Cole 2026, 05:40–08:20).

This changes the test for value. Manual effort does not prove understanding, originality, or usefulness. Human and machine writers can both produce generic text. Conversely, original does not automatically mean useful, and commodity does not automatically mean worthless. Cole explicitly argues that commodity content can be especially useful to beginners. The categories describe the contribution made by the text, not the moral worth of its author.

Cole’s operating rule is:

“I don’t want the robots writing for me. What I want is the robots repeating me.”

Writing first crystallizes the author’s choices. It creates what Cole calls a digital brain or personal language model: a library of approved language, stories, positions, and examples. AI can then repeat and recombine that material while the author spends more time developing the next idea (Isenberg and Cole 2026, 14:26–20:15).

Talk track: human slop preceded AI slop. Language is shared material, so a human can reproduce commodity content as easily as a model can. Human authorship therefore guarantees neither originality nor quality. The useful question is what contribution the text makes.

[Sources]

• Greg Isenberg with Nicolas Cole, “$30M Writer: Never write AI slop again,” especially 05:40–08:20, 14:26–20:15, and 44:25–46:05 (Isenberg and Cole 2026).

[/Sources]

The student as a lossy proxy

Teacher feedback can improve the next answer without improving the student.

Learning begins when the student can defend the choice.

Gruhn uses meat proxy for a person who forwards AI output without reading, understanding, validating, or adding judgment. The intermediary adds another communication hop while transferring the cognitive work to the recipient (Gruhn 2026).

The education loop is the user’s extension of that idea. A student can submit generated work, forward the teacher’s comments to a model, and return the next generated version. The artifact may improve at every round while the student’s understanding remains unchanged. In that case, the teacher is effectively debugging an AI through a lossy human proxy.

Henle distinguishes empty forwarding from useful mediation. A valuable intermediary supplies missing context, checks the result, selects what should change, and accepts responsibility. Applied to learning, the student must interpret the feedback and be able to explain and defend the revised choice (Henle 2026).

Talk track: the problem is not that the student used AI. The problem appears when the student becomes only a transmission layer. The evidence of learning is the student’s ability to explain why the answer changed and why the final choice should remain.

[Sources]

• Niklas Gruhn, “Don’t be a meat proxy” (Gruhn 2026).
• E. Adrian Henle, “Let Me Ask Claude for You” (Henle 2026).

[/Sources]

Who is using whom?

The same tools can create opposite relationships.

Reverse centaur AI → you

The system chooses the route, pace, and task.

ride-hailing · delivery · picking

WHO LEADS?

Centaur You → AI

You choose the goal, constraints, and judgment.

research · analysis · creation

The goal: keep the human in the head.

The metaphor is about agency, not anatomy. In a reverse-centaur system the machine decomposes and directs the work while the person supplies perception, movement, and accountability (Doctorow 2021). Kasparov pioneered Advanced Chess in 1998 and players later adopted the centaur label, but he should not be credited with coining the word (ChessBase, n.d.; Wordorigins.org, n.d.).

The panopticon is already here

AI joins a surveillance system that already records daily life.

Bentham’s Panopticon plan, 1791 (published 1843)

The infrastructure of observation was already in place.

AI makes it easier to infer identity, intent, and risk from what it records.

Trust depends on who can act on the inference and whether people can challenge it.

The claim is not that AI surveillance is harmless. Its novelty is overstated: cameras, connected databases, and digital platforms already support pervasive observation. AI expands what institutions can infer and how quickly they can act. The governance question is therefore institutional trust: lawful and transparent use, proportionate action, and a meaningful route to challenge a decision (Information Commissioner’s Office, n.d.).

[Sources]

• The figure reproduces Jeremy Bentham’s Panopticon plan, originally designed in 1791 and published in 1843. The Wikimedia Commons file is marked public domain (Bentham and Reveley 1843).
• The UK Information Commissioner’s Office describes the growth of AI-based surveillance and identifies public trust, fairness, transparency, and proportionality as essential safeguards (Information Commissioner’s Office, n.d.).

The Spartan rule of AI

Don’t get caught.

Scarcity

Resourcefulness is the test.

Same rule

Abundance

Discernment is the test.

AI multiplies possibilities. You decide what deserves to remain.

Choose. Verify. Refine.

Xenophon 2.5–2.8 and Plutarch, Lycurgus 17.3–4: boys were kept hungry, permitted to steal, and beaten when caught — punished for stealing badly, not for stealing (Xenophon 1925; Plutarch 1914). These are literary accounts, hence “ancient accounts of Sparta” rather than verified social history; modern scholarship treats Spartan education through the “Spartan mirage” (Ducat 2006; Kennell 2010). The modern analogy follows Meservey (2026), with “steal” changed to “take” to avoid implying literal IP theft.

The analogy does not encourage hiding AI use. The work itself exposes careless tool use: invented claims, unexamined cliché, brittle code, or decisions you cannot explain. The ancient sources support resourcefulness directly. Judgment, verification, and ownership are the modern extension of the analogy.

Iron sharpens iron

Proverbs 27:17

AI alone

Feedback without stakes

Single perspective

Culture at a distance

✦

AI + people

Real feedback

Challenges your framing

Exposes you to culture

AI can help you practice. People sharpen your thinking.

The Hebrew describes iron sharpening iron and a person sharpening the “face” of a friend; the NET notes read it as constructive criticism sharpening character (NET Bible, n.d.). A later rabbinic reading applies it to two scholars sharpening one another through joint study (Sefaria, n.d.).

Bastani et al. found unrestricted GPT access improved assisted practice but reduced later unassisted performance, and that tutor-style guardrails largely removed the harm (Bastani et al. 2025). Dai et al. warn that apparently cooperative AI interactions can entrench simplified cultural narratives (Dai et al. 2025).

Talk track: a model cannot induct you into a living culture by itself.

The human edge

Insights · AI

AI can imitate patterns.

Your perspective comes from what you have lived.

What you notice and what you reject depend on who you are.

Experience ⟶ Taste ⟶ Judgment

A model can imitate any pattern in its training data; what it cannot do is have been somewhere. End on the three-word sequence.

Don’t build your work on rented land

Digital sovereignty means the system survives when a supplier changes.

Digital feudalism One company owns the stack

interface

agents + automations

memory + knowledge

model

They change the rules price ↑ · API shifts · product sunset

Your workflow breaks.

OR

Digital sovereignty You own the system

open files + data

your knowledge + workflows

portable interfaces

model A → model B

A supplier changes swap the model, preserve the system

Your workflow continues.

Own the system. Rent the intelligence.

The issue is not using a large platform; it is letting the cost of exit grow larger than the value of choice. The CMA found technical and commercial barriers lock cloud customers in (Competition and Markets Authority 2025). The EU Data Act bans switching charges from 12 January 2027 and requires machine-readable exports (European Union 2023; European Commission, n.d.). NIST names portability and standardised formats as the safeguards (National Institute of Standards and Technology, n.d.). “Datafeudalism” is a debated political-economy concept, narrowed here to a practical design warning (Galindo 2024).

Talk track: convenience is not sovereignty. The model should be a socket, not the house.

References

Bastani, Hamsa, Osbert Bastani, Alp Sungu, Haosen Ge, Ozge Kabakcı, and Rei Mariman. 2025. “Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics.” Proceedings of the National Academy of Sciences 122 (26). https://doi.org/10.1073/pnas.2422633122.

Bentham, Jeremy, and Willey Reveley. 1843. Plan of the Panopticon. Wikimedia Commons, https://commons.wikimedia.org/wiki/File:Penetentiary_Panopticon_Plan.jpg.

ChessBase. n.d. How Technology Is Changing Our Minds. Https://en.chessbase.com/post/how-technology-is-changing-our-minds/25.

@clipsofcrypto. n.d. Post on AI Writing, Templates, and “Sea Level”. X, https://x.com/clipsofcrypto/status/2088827366491762895.

Competition and Markets Authority. 2025. CMA Independent Inquiry Group Publishes Provisional Findings in Cloud Services Market Investigation. Https://www.gov.uk/government/news/cma-independent-inquiry-group-publishes-provisional-findings-in-cloud-services-market-investigation.

Dai, David Wei, Hua Zhu, and Guanliang Chen. 2025. “How Does Interaction with LLM-Powered Chatbots Shape Human Understanding of Culture? The Need for Critical Interactional Competence (CritIC).” Annual Review of Applied Linguistics 45: 28–49. https://doi.org/10.1017/S0267190525000054.

Doctorow, Cory. 2021. Reverse Centaurs and the Failure of AI. Pluralistic, https://pluralistic.net/2021/02/17/reverse-centaur/.

Ducat, Jean. 2006. Spartan Education: Youth and Society in the Classical Period. Classical Press of Wales.

European Commission. n.d. Data Act Explained. Https://digital-strategy.ec.europa.eu/en/factpages/data-act-explained.

European Union. 2023. Regulation (EU) 2023/2854 on Harmonised Rules on Fair Access to and Use of Data (Data Act). Https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32023R2854.

Galindo, Luis. 2024. “Datafeudalism: The Domination of Modern Societies by Big Tech Companies.” Philosophy & Technology 37: 121. https://doi.org/10.1007/s13347-024-00777-1.

Gruhn, Niklas. 2026. Don’t Be a Meat Proxy. Personal website, https://gruhn.me/blog/2026-08-03/.

Heinemeier Hansson, David. 2026. Rails World 2026 Opening Keynote. Ruby on Rails, YouTube, https://www.youtube.com/watch?v=V9SxpJpHuus.

Henle, E. Adrian. 2026. Let Me Ask Claude for You. Personal website, https://henletech.net/2026/08/06/let-me-ask-claude-for-you.html.

Information Commissioner’s Office. n.d. Video Surveillance, Including Guidance for Organisations Using CCTV. Https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/cctv-and-video-surveillance/guidance-on-video-surveillance-including-cctv/.

Isenberg, Greg, and Nicolas Cole. 2026. $30M Writer: Never Write AI Slop Again. YouTube, https://www.youtube.com/watch?v=YuOSyRj3sXg.

Jevons, William Stanley. 1865. The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal-Mines. Macmillan; Co. https://www.econlib.org/library/YPDBooks/Jevons/jvnCQ.html.

Kennell, Nigel M. 2010. Spartans: A New History. Wiley-Blackwell.

Meservey, Lulu Cheng. 2026. Post Comparing Spartan Training with AI-Assisted Work. X, https://x.com/lulumeservey/status/2092369479934038461.

National Institute of Standards and Technology. n.d. Cloud Computing Synopsis and Recommendations. Special Publication 800-146. National Institute of Standards; Technology. https://nvlpubs.nist.gov/nistpubs/legacy/sp/nistspecialpublication800-146.pdf.

NET Bible. n.d. Proverbs 27:17. Bible.org, https://classic.net.bible.org/passage.php?passage=Pro+27:17.

Plutarch. 1914. “The Life of Lycurgus.” In Plutarch’s Lives, Volume i. Loeb Classical Library 46. Harvard University Press. https://penelope.uchicago.edu/Thayer/E/Roman/Texts/Plutarch/Lives/Lycurgus*.html.

Quote Investigator. 2011. Quote Origin: If You Want Jobs Then Give These Workers Spoons Instead of Shovels. Quote Investigator, https://quoteinvestigator.com/2011/10/10/spoons-shovels/.

Sefaria. n.d. Taanit 7a:12. Https://www.sefaria.org/Taanit.7a.12.

The Rails Foundation. 2026. Rails World 2026 Opening Keynote – DHH. LinkedIn, https://www.linkedin.com/posts/ruby-on-rails-org_rails-world-2026-opening-keynote-dhh-activity-7508700625996775424-KPoz.

U.S. Department of Energy. 2023. Rebound Effect Analysis. U.S. Department of Energy. https://www.energy.gov/sites/default/files/2023-03/PH_FR_TSD_2023-03-31.pdf.

Wordorigins.org. n.d. Centaur, Reverse Centaur. Https://wordorigins-org.ghost.io/centaur-reverse-centaur-free-copy/.

Xenophon. 1925. “Constitution of the Lacedaemonians.” In Xenophon: Scripta Minora. Loeb Classical Library 183. Harvard University Press. https://topostext.org/work/798.