The Empty Quadrant

The centaur, a human working with a machine as one entrant, holds a lease on the operator's usefulness. That lease has an expiry date.

A narrator in 1940s America walks up to a picket line of hand-tool-only carpenters and asks why one of them is holding a signboard instead of a tool. The demonstrator says power tools are a menace, then makes his case. Eric S. Raymond posted it on 11 September 2026, at 10:48 UTC.

How are you going to train apprentices if they can't feel the wood through their hands? What happens when we can chew through aged quality timber faster than we can grow new trees? Carpentry isn't just a set of skills and habits, it's a mindset. It won't be any true carpenters left if we let this go on.

The demonstrator, in Raymond's parable

The demonstrators plan to blacklist anyone who has ever used a power tool, and say that with enough people on their side they can bury all this. The parable resolves into a sorting, and the giveaway is that the post is about something else. It circulated widely, and the replies I read largely accepted the analogy.

Raymond never answers the objections. The narrator asks what the real problem is, listens, and walks away. Three of the four objections (apprenticeship, timber supply, and craft as a mindset) concern judgement, training supply, or input supply, and those are the serious ones. The fourth, the exclusion pledge, is the cartoon, and Raymond attributes it to the demonstrators rather than to anyone making the argument in 2026. The real dispute is about junior pipelines, liability, and who absorbs the cost of maintaining code nobody understands.

Put at its strongest, his case is this: the tools demonstrably help the people who use them, the objections are often made in bad faith, and refusing the tools is a counsel of despair rather than a plan. That case is arguable, and it is not the case the parable makes.

The post gives no answers, and the choice it offers is demographic. Which tool you would rather hold is a different question from whether the objections are correct, and a saw offers no judgement either way. That gap is the part the parable never reaches.

Two axes decide where a tool sits: is its output repeatable, and does the tool decide the output or does the operator? A tool augments when the operator still decides and the tool supplies force, speed, or reach. It performs when the tool decides the output and the operator only feeds or watches. A saw is repeatable and augments, so its operator stays. A power loom is repeatable and performs, so the hand-loom weavers it displaced lost their trade. A horse is not repeatable and augments, because the rider decides the route, so its handler lasted millennia.

The fourth cell, non-repeatable and performing, has been empty. That is the one the LLM occupies, and it explains why its arrival is different from either a loom or a horse. Replacement is coming, on a clock set by how expensive the output is to verify, and the delay is the lease.

Sampling makes a model non-repeatable. At a nonzero temperature the same prompt can take a different path, a changed context window turns the same prompt into a different prompt, and a fluent answer can be wrong without looking wrong. The failure mode is fluent error: the fabricated case citation arrives in the same register as the real one. That is why the 2023 Mata v. Avianca sanction, where a lawyer filed chatbot-invented authorities, is the pattern. A model run at temperature zero is repeatable and can still be wrong, which is why repeatability alone is not the axis; the pairing of variability with apparent judgement is what does the work.

Industrial replacement was sold as repeatability. A manufactured, fluent, cheap, apparently judging system aimed at the whole task has not been mass-deployed before, and where a close analogue exists it teaches a sharper lesson than learn to use it. Handling skill is real, and it has a shelf life.

I. What industrial replacement was sold as

The stocking frame, the power loom, the spinning mule, the pocket calculator, the automatic telephone exchange: same input, same output, by design. A loom that produced a different cloth each time you fed it the same wool would have been a defect. Capital promised that the future would look like the past, only faster.

That is a design ideal. Breeding, fermentation, sailing, and gunnery in weather all substituted for human work while staying stubbornly variable, and they were sold anyway. Where industrialisation institutionalised a substitute, it sold the substitution on repeatability.

The Luddites of 1811-1816 smashed specific frames amid wage cuts and the dilution of skilled work. They were fighting a labour settlement, and machinery was the terrain: machine-breaking was a bargaining tactic in a decade of trade depression, and the frames they destroyed were the ones undercutting their rates. Raymond's picketer is the cartoon of that fight.

Every labour substitution the industrial economy institutionalised was sold on the promise that the tool would do the same thing again. The LLM is the first widely deployed cognitive tool sold without that promise. You can still learn to use it. You cannot treat that skill the way you treat a trade built on a saw, because a saw makes a carpenter faster and a loom makes the weaver optional. The saw does not get better at being the carpenter.

II. The myths point elsewhere

Cultures tell two kinds of story about made servants, and the difference is speech.

The made being that cannot speak is a tool. Sanhedrin 65b has Rava send a man to Rabbi Zeira; Zeira speaks to him, gets no answer, and knows from the silence that the man was made. The Golem of Prague is the same type, clay animated by a written name, and the legends keep him silent. Goethe's broom fetches water forever because it was told to start and cannot be told to stop. Mute means tool, and the fear is runaway obedience.

Speech moves the story into another genre. Shelley's creature is manufactured, willed, and eloquent, and the horror is recognition. Čapek's robots in R.U.R. speak, organise, and inherit the earth, and their name comes from robota, servitude. In the Solomon cycle of Gittin 68a-b, Ashmedai outwits the binder: handed the ring, the chain comes off, and he throws Solomon four hundred parasangs and sits on the throne in the king's shape.

Homer supplies an exception: Hephaestus has golden handmaidens with understanding in their hearts, speech, strength, and knowledge of their work (Iliad 18.417-421), and self-moving tripods that reach the assembly of the gods (18.373-377). They support their lame master, and nothing goes wrong, because their maker is a god. The tradition never extends that licence to a human claim that a speaking servant will stay a servant.

Plato's Phaedrus still maps. Thamus condemns writing because a text cannot be interrogated: it says the same thing, with great authority, to whoever picks it up. The LLM answers back, which the Phaedrus could not imagine. It does not answer for itself, and having no stake in being right does not stop it making you wrong.

The mute tradition is calibrated for a tool that is strong and silent, and the speaking tradition for a creature that has become a subject. The canon infers inner life from speech, and that inference breaks on a system that produces fluent answers without being owed an answer and without being a rival. How long the person who knows how to point it remains necessary is the smaller, harder question.

III. Two precedents that are not instances

The human computer: prestige is decided politically. Nearly two hundred young women, civilian and military, worked as human computers on the wartime ballistics-computation project. The figure is Jennifer Light's; Jeremy Bernstein put the Moore School computing section nearer one hundred, the usual gap between a programme headcount and a departmental one. ENIAC's six operators came out of that Philadelphia cohort. At the Jet Propulsion Laboratory, women computers wrote mission software for decades, and by around 1970 JPL had retired the title "computer" for "engineer," with new hires expected to hold engineering degrees.

They did deterministic arithmetic and were displaced by deterministic machines, and the lesson is political. Programming began as feminised clerical labour, and Nathan Ensmenger has shown how the work was reclassified upward in the same decade through professional associations, credential requirements, and personality testing that selected for a masculine culture of the "computer bum." Light's argument is that the wartime press framed the women as operators of a male invention. "AI literacy" is now sold as a new permanent class of work. The last time a cognitive tool created a job title, the title lasted until the machine absorbed the task and the survivors were renamed.

The oracle: fluent ambiguity has an economics. The Spartans asked Delphi whether to go to war before the Peloponnesian War, and the answer was close to a green light. The riddle about a great empire falling, the school example, went to Croesus, and it is the exception. Shang kings asked the cracks in oracle bones about harvests, ancestors, and campaigns, and the diviners who read them accumulated power over state activities.

Every oracle is a human medium plus an interpretive guild. The Pythia ran for about a thousand years, never displaced by a better oracle, and did not outlast pagan antiquity. What grew alongside it was exegesis, the craft of reading an authoritative, ambiguous utterance and deciding what it meant for the case at hand. An output ambiguous enough to confirm whatever the querent already wanted is the modern failure mode too, and a guild of interpreters can last a long time when the source does not improve. Language models get better on a cadence the oracle never had.

IV. The horse, and the measured case

The horse is the closest thing in this audit to a willed, unrepeatable power source pointed at human work. The modern domestic horse lineage originated on the western steppe, and horse-based mobility spread across Eurasia around 2200 BCE. For three millennia it replaced human legs for couriers and mounted warfare. Oxen, donkeys, mules, and camels took over the ordinary haulage, while walking and porting continued as before. Horsemanship defined aristocracies across much of Eurasia, and the failure mode, in every cavalry manual ever written, is a rider who treated a willed animal as a machine.

"You do not get replaced by the horse" is a slogan: grooms, ostlers, farriers, and coachmen were replaced when a better substitute arrived. What the horse teaches is a craft of handling with a documented expiry date, where the handler is an asset until the substitute no longer needs him. Horsemanship was a class for a long time because the animal did not start riding itself. Software does, and that expiry has one measured case.

In June 1998, in León, Kasparov and Fritz 5 drew 3-3 with Topalov and ChessBase 7.0 at sixty minutes a side. Kasparov's brief was to produce a quality of play neither man could reach alone, which is the pitch now made for AI literacy: the human plus the machine beats either alone. The format peaked in 2005. Two American amateurs, Steven Cramton and Zackary Stephen, with club-level ratings of 1685 and 1398 on the US Chess Federation scale, entered the PAL/CSS Freestyle Tournament on Playchess as ZackS. The event was run by the PAL Group, which also built Hydra. They used three ordinary computers running Fritz, Shredder, Junior, and Chess Tiger, won the qualifier at 6/7 and the main Swiss at 6½/8, a full point clear in the Swiss, and in the four-game knockout final beat the Russian grandmaster Vladimir Dobrov, working with a 2600-plus collaborator and computers, by 2½-1½.

Two corrections matter, because both propagate. Dobrov was twenty-one that summer; ChessBase's flash report called him fourteen, and he was born 13 April 1984, made IM in 2001 and GM in 2004. Hydra never reached the knockout: the engine cluster failed to qualify, with Hydra Chimera unattended on 3½/8 under the handle Zor_Champ and Hydra Scylla, with human help, on 4/8. Hydra was the sponsor's own machine, built by the PAL group in Abu Dhabi, which was underwriting the tournament. The 2017 winner was also called Zor; the names appear to be unrelated.

ChessBase drew the conclusion from the field. Of the eight quarter-finalists, six were titled players working with computers, and the organisers concluded that "this combination is in general much stronger than even vastly superior hardware that is not supported by human intervention." The only machine running without a human, Dieb Fritz, "was completely outclassed by a GM working with a lesser machine." That is the centaur thesis in the primary record in 2005.

The obvious suspicion was that a hidden grandmaster must be feeding ZackS moves. ChessBase asked Kasparov, who confirmed he was not that grandmaster, then declined to treat the question as load-bearing: "Even if they were assisted by the devil, that would probably be covered by the rules. Only the moves they played count."

Process skill, in 2005, beat chess skill: the amateurs managed the tool better than the grandmaster managed being a grandmaster with the tool beside him. Two people with club ratings and a method beat titled players, and the method was the scarce resource. Then the window closed. In 2017 a pure engine, Zor, won the InfinityChess Ultimate Challenge outright, with the leading centaur team third. Engines had passed humans years earlier, and the centaurs kept winning; what ended the format was human intervention starting to hurt. Once the engine's line outruns the human's ability to check it, the human's edits start subtracting value. Strength is one threshold, utility the other, and utility is the one that ends the centaur.

The timeline is the whole argument: seven years from advanced chess in 1998 to the freestyle peak in 2005, then twelve more to an engine finishing ahead of the centaurs at the 2017 Ultimate Challenge. Nineteen years across three formats, which is a long product cycle and a short career for the idea they share.

AlphaZero belongs to a different argument. DeepMind's paper went to arXiv on 5 December 2017, and the closed-door match against Stockfish 8 came earlier, with 28 wins, no losses, and 72 draws under conditions that remain disputed. That is engine versus engine, and it says something narrower: games with a cheap, objective outcome function close fast.

Chess had a fixed, cheap, objective evaluation, and many paid tasks have ground truth as well: a test suite, a compiler, a reconciled ledger. The centaur window stays open where verification is expensive. That describes where leases remain; it does not promise that prompting is a profession.

V. What is new

Non-determinism is old. Humans have used willed, non-repeatable agents for as long as we have kept animals. Fluency in isolation is also old: ELIZA, statistical machine translation, search, recommenders, and automated trading already pointed probabilistic systems at work people used to do. Printing, radio, recorded music, and software already bent the cost curve of copies. Each of those is old, and the combination is new mostly as a matter of degree.

What is new is that the four arrive in one system: manufactured where the horse was grown, fluent in a way ELIZA never reached, cheap at a price no previous substitute offered, and one-to-many where a horse stands in one field. The category is not empty; one cell within it is. The handler rents the gap between what the tool can do and what the work needs, and each release erodes the rent.

The horse could not update itself overnight. Nor did the oracle get cheaper every quarter. Chess did, and the rider's advantage lasted one professional generation. Raymond picked a saw, and the question is whether the analogy holds: a hand tool does not absorb the judgement the work needs, and the chess record says a judging tool does.

VI. The horror plot needs a subject

The closest historical analogue to a general-purpose, adaptable substitute that can be directed at any task is people and animals. The horse is half of that. The other half is chattel slavery, which is where the labour-relations horror story comes from.

Everyone keeps retelling it: the created thing that wakes up, resents its condition, and turns on the maker. It is the oldest story we have, a person treated as a tool. Čapek named the genre, and Westworld and Ex Machina are the same plot with better production design. That plot requires a subject. A language model can be directed at any task describable in language without being a person and without being an animal, so the slave-rebellion alarm misfires.

The distinction is moral and it settles nobody's job. Treating the model as a slave harms no one. Treating the people whose tasks it absorbs as if "learn to use it" had settled their fate is the error that matters, and history does not excuse it. The people at risk are the riders, and the horse was never in danger.

VII. Two ways this goes

The post takes a deterministic tool, hands one demonstrator four objections, answers none of them, and treats the resulting fork as a test of allegiance. Refuse the saw as the analogy and the fork collapses, because a saw does not take over the trade and a judging tool does, on a clock, once intervention starts to hurt.

Three complaints are live: confident errors at scale, the use of writers' and artists' work to train commercial models without payment or consent, and deployment into work where the error is invisible to the person who needed the answer. None of those apply to a saw. A centaur might have been an honest analogy in 2005, and by 2017 it had already lost.

Nobody knows which fork this is, and the answer depends on a variable that can be watched. Two scenarios, and a list of signals, roughly in the order they will arrive.

Scenario A: the lease renews. Verification stays expensive, or gets more expensive, across most of the domains you work in. The marginal human in the loop keeps adding measurable value, and intervention overhead stays flat or falls. The 2005 result is the template, and handling behaves like every previous augmentation: compilers, IDEs, search, the spreadsheet. It is a real trade for a real stretch of time.

Scenario B: the lease expires. Intervention overhead crosses zero in your domain, and your review starts costing more than it saves, the way a 2017 centaur team finished behind a lone engine. Handling was a real skill with a finite life, and it ran out before the people holding it could convert it into anything else.

The signals, ranked by how early you would see them:

1. Intervention overhead per task. Does reviewing the output take longer than producing it yourself? This is the earliest and cheapest signal, measurable in a week on your own work, and it is the quantity that killed the centaur in chess. Track it per task type; the aggregate hides the domains that have already flipped.
2. Marginal contribution of a human. In a blind comparison, does adding a person improve the artefact or degrade it? Chess ran this twice, with a field in 2005 where the combination won and a result in 2017 where it lost. It is testable in most domains, and almost nobody runs it.
3. The entry-level rung. Junior hiring and headcount in the affected job families, over a full hiring cycle. Slower, two to five years, but it decides whether the craft can reproduce itself, and it is the one objection Raymond's demonstrator got right.
4. Verification cost inflation. If catching errors costs more as output volume rises, the arithmetic inverts regardless of how capable the model gets. This is the mechanism behind Scenario B, visible before the outcome.
5. Who gets to own the skill. The JPL precedent: is the renamed job credentialled, and re-staffed by a different population, in the same decade as the tool? It arrives last, and it decides who captures the value those signals describe.

None of this argues for a pause, for a licence to use a model, or for a regulator deciding who may run one. A wrong analogy is answered with a better argument, and the signals above exist to make the lease measurable, not to stop the buildout.

Whether the tool in your hand needs you at all is testable. A saw does. A horse eventually does not, and the difference is discoverable, per task, this quarter, by anyone willing to measure the overhead.

Chess ran the experiment to completion and published the results. The amateurs won the lease. The engine kept the house.

Sources

Where a claim rests on a primary text, the primary is cited first.

• The parable: Eric S. Raymond, X, 11 September 2026, 10:48 UTC. Mirrored at Digg, "AI Coding Tools Spark Carpenter-Power Tool Debate" (digg.com/tech/86730cc2-6173-49f6-9826-089ce25ad471). The demonstrator's objections are quoted from the captured screenshot.
• Mythology: Sanhedrin 65b and Gittin 68a-b, Babylonian Talmud; YIVO Encyclopedia, "Golem Legend" (encyclopedia.yivo.org/article.aspx/Golem_Legend); Iliad 18.373-377 and 18.417-421, A. T. Murray translation (theoi.com/Text/HomerIliad18.html); Shelley, Frankenstein (1818); Čapek, R.U.R. (1920); Plato, Phaedrus 274b-279c (perseus.tufts.edu).
• Fluent error: Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023), sanctioning a lawyer for filing chatbot-fabricated case citations.
• Human computers: Jennifer S. Light, "When Computers Were Women," Technology and Culture 40:3 (1999), 455-483 (cs.brown.edu/courses/cs1951i/lightWhenComputersWereWomen.pdf); Nathan Ensmenger, The Computer Boys Take Over (2010); JPL's title change around 1970, Smithsonian American Women's History Museum.
• Oracles: Thucydides 1.118.3; Herodotus 1.53; "Divination and Power," Current Anthropology (journals.uchicago.edu/doi/10.1086/588495).
• Horses: Librado et al., Nature 598 (2021) and Nature 631 (2024).
• Advanced chess: ChessBase reports of 11 and 19 June 2005, covering ZackS, Hydra, the field-wide result, the winners' own account (en.chessbase.com/post/pal-c-report-from-the-dark-horse-s-mouth), and Kasparov's statement; León 1998, Kasparov and Fritz 5 against Topalov and ChessBase 7.0, 3-3, in the contemporaneous ChessBase reporting; the 2017 InfinityChess Ultimate Challenge, won by the engine Zor (mindmatters.ai/2018/12/can-big-data-beat-the-humans-who-compile-it; organiser Arno Nickel, chessprogramming.org/Arno_Nickel). Hydra Chimera's 2005 handle Zor_Champ is unrelated to the 2017 winner Zor. Vladimir Dobrov, born 13 April 1984, IM 2001, GM 2004; the ChessBase flash report's "14-year-old" is a contemporaneous error.
• Luddites: "What the Luddites Really Fought Against," Smithsonian Magazine; The National Archives (UK), "Why did the Luddites protest?"

Transparency Note

The ideas, arguments, and structure in this essay originated with the author. AI tools were used to assist with drafting, research, and revision. All claims, sources, framing, and final wording reflect the author's own thinking and were reviewed for accuracy before publication.

This essay is for informational and educational purposes only and does not constitute professional advice of any kind, including financial, legal, medical, or otherwise. The author makes no guarantees regarding accuracy or completeness. Readers should consult a qualified professional before acting on any information contained here. The author accepts no liability for decisions made based on this content.