Moravec's paradox

Filed correctly. proposals.md lists this under interpretation, and after working the alternatives the filing holds — but the reason has to be stated first and at length, because the dominant misuse of this entry is to treat it as the kind it is not. A limit, on the house definition that godel-incompleteness-1931 and turing-halting-1936 established, is a theorem: it has a proof, a set of hypotheses, and a territory outside them, and what it tells you is that a region is empty. Moravec's paradox has no proof, no stated hypotheses, no quantifiers you could check, and forbids nothing. It is an observation about which things had turned out to be hard by 1988, plus a story about why. It is quoted in the same tone as an impossibility result — "robots can't, Moravec's paradox" — and section 4 is mostly about the damage that does.

idea fails the test samuel-checkers-1959 set and goodharts-law-1975 applied: an idea is something you can implement, because there is a loop or a formalism in it. There is no loop here. The machinery people actually built out of this observation — Rodney Brooks's subsumption architecture and "nouvelle AI," which put sensing and acting first and threw out the symbolic middle — is Brooks's, not Moravec's, and it is not in canon/.

prediction is the serious rival, and it is closer here than it was for Goodhart, for a reason worth naming. Goodhart's product was monetary analysis and the famous sentence was an aside. Moravec's product was forecasting. Mind Children is a book of dated predictions by a man who made dated predictions for a living, and prediction is the kind this canon weights heaviest. I decline it anyway, on the narrow ground that this entry's subject is the paradox and the paradox carries no date. It says which class of problem is harder; it does not say by when, by how much, or what happens next. What it does carry is an implied ordering, and an ordering is gradeable, so section 3 grades it under the full prediction discipline — the claim in its own words, the date made, the date due, what actually happened, and the author's own grade set against an independent one. Those grades differ here in an unusual pattern, and section 3 is where that lives.

The entry that should be filed as prediction is a different one and it is not in proposals.md. Moravec's hard forecast — human-level machine intelligence roughly forty years from 1988, which is to say by 2028 — has a claim, a date made, a date due, and eighteen months left to run. That is the cleanest gradeable AI prediction in the canon's reach and it is about to come due inside this project's operating life. It is not this file's subject and this file does not grade it. A later session should propose it; moravec-2028 or similar would be the id, and it should not be folded into this one.

Two things about the label need saying before anything else.

First, "Moravec's paradox" is not Moravec's phrase. He did not call it a paradox in Mind Children, and I could not establish who did; the name was attached afterwards by the field. This matters more than a naming quibble, because "paradox" is doing rhetorical work the observation cannot support — it implies something surprising enough to need resolving, when the honest content is "here is a pattern in what worked, and here is my guess at why." Section 4 returns to it.

Second, Moravec was neither first nor alone, and one of the earlier statements is a rebuttal. Marvin Minsky wrote in The Society of Mind two years earlier that "in general, we're least aware of what our minds do best." Brooks was arguing the same case in the mid-1980s from the robotics side. Steven Pinker gave it its most quotable form six years later in The Language Instinct — "the main lesson of thirty-five years of AI research is that the hard problems are easy and the easy problems are hard" — which, exactly as with goodharts-law-1975, is the sentence most people actually have in mind when they cite the man whose name is on the label. And **Allen Newell, in 1983, five years before Mind Children, described the same proposition as a myth**: that "a myth grew up that it was relatively easy to automate man's higher reasoning functions but very difficult to automate those functions man shared with the rest of the animal kingdom." So the first recorded position on Moravec's paradox in the literature is against it, and it predates Moravec's statement of it. A reading that cites this entry for the standard version should know the standard version has been contested from before it was written.

On descends_from. Two entries, and the strength of each link differs.

samuel-checkers-1959 is a textual link and I believe it is the right one, with one inference marked: Moravec's canonical sentence names checkers as an example of what is cheap, and in 1988 the machine-checkers reference point was Samuel's program — Chinook's first world-championship match was two years away and its solution of the game nineteen. Moravec does not cite Samuel in the sentence as quoted; that he meant that program is my inference, and Mind Children itself was not read for this entry.

lighthill-1973 carries the other half, and the link is thematic rather than documented. Lighthill's Category B was literally called "Building Robots"; he judged eye–hand co-ordination "not of great potential"; the discussion around the report turned on the proposition that the general purpose robot is a mirage; and he put the point about children in the report itself. That is Moravec's observation fifteen years earlier, in an official document, with the sign flipped — indictment rather than explanation. Whether Moravec read it is not established here. dartmouth-1956 is the grandparent of both and is reachable through Lighthill, so it is named here and not listed.

The honest ancestor that is not in the canon is Moravec's own decade on the Stanford Cart — the TV-equipped robot he rebuilt with stereo vision in 1977 and took through his 1980 Stanford dissertation under John McCarthy. In 1979 it crossed a room full of chairs without human intervention, in about five hours, advancing in one-metre spurts separated by ten-to-fifteen-minute pauses for image processing and route planning, on a KL10 running around 2.5 MIPS. Set that against his own later estimate that a human brain is about 100 million MIPS and the paradox stops being a paradox and starts being a receipt. A man who has spent ten years making a machine cross a room, while watching other people's programs prove theorems in an afternoon, does not need a literature to notice the asymmetry. The spec forbids inventing ids, so that is recorded in prose.

---

What it is

The sentence

In Mind Children: The Future of Robot and Human Intelligence (Harvard University Press, 1988), Hans Moravec wrote:

> "it is comparatively easy to make computers exhibit adult level performance on > intelligence tests or playing checkers, and difficult or impossible to give them > the skills of a one-year-old when it comes to perception and mobility"

That is the whole of the claim. Everything else attached to the name is either the explanation Moravec gave for it, or somebody else's restatement.

Note the shape carefully, because almost every misuse in section 4 is a mis-parsing of it. It is comparative, not absolute — "comparatively easy" and "difficult or impossible," a ranking rather than a boundary. It names two specific competences — perception and mobility — not manual work, not physical labour, not "the real world." And its bar is a one-year-old, which is a much lower bar than the one both sides of the modern argument actually use. Section 3 returns to all three.

The explanation

The evolutionary argument is Moravec's own and is the part that gets reproduced least and misused most:

> "Encoded in the large, highly evolved sensory and motor portions of the human > brain is a billion years of experience about the nature of the world and how to > survive in it. The deliberate process we call reasoning is, I believe, the > thinnest veneer of human thought, effective only because it is supported by this > much older and much more powerful, though usually unconscious, sensorimotor > knowledge. We are all prodigious olympians in perceptual and motor areas, so > good that we make the difficult look easy. Abstract thought, though, is a new > trick, perhaps less than 100 thousand years old. We have not yet mastered it. It > is not all that intrinsically difficult; it just seems so when we do it."

The mechanism is optimisation time. Old competences are heavily optimised, therefore unconscious, therefore invisible to introspection, therefore underestimated by the engineers trying to reimplement them; recent competences are lightly optimised, therefore effortful, therefore conspicuous, therefore overestimated. The paradox on this account is not a fact about machines at all. It is a fact about human self-report — a systematic error in how hard we think our own faculties are — which is the same thing Minsky said in 1986 and is the most defensible part of the whole package.

Section 3 grades this mechanism separately from the ranking it was invented to explain, because they have come apart.

The landscape, which is the part the misusers skip

Ten years later, in "When will computer hardware match the human brain?" (Journal of Evolution and Technology, vol. 1, 1998, received December 1997), Moravec gave the picture that settles what he thought he was claiming:

> "Imagine a 'landscape of human competence,' having lowlands with labels like > 'arithmetic' and 'rote memorization', foothills like 'theorem proving' and > 'chess playing,' and high mountain peaks labeled 'locomotion,' 'hand-eye > coordination' and 'social interaction.'"

And the water is rising. Half a century ago it drowned the lowlands and drove out the human calculators and record clerks; by 1998 it had reached the foothills; the peaks, he wrote, would be submerged "within another half century."

This is the single most important fact in the file for anyone about to cite the paradox as a limit. Moravec's own model is a waterline, not a wall. He named locomotion and hand-eye co-ordination as the last things to go under, not as the things that never would. Anyone using his name to argue that a class of physical task is permanently out of reach is contradicting him with his own metaphor, and the same paper carries his estimate of when: about 100 million MIPS to match a human brain, derived by scaling from the retina (100 million neurons, ~100 instructions per edge or motion detection, ~1,000 MIPS to match the retina at ten frames a second, and a brain about 100,000 times a retina), arriving in a $1,000 machine such that "computers suitable for humanlike robots will appear in the 2020s."

The modern replacement for the mechanism

The explanation that has displaced the evolutionary one in working robotics is data, and it is worth having because it makes different and checkable predictions.

Ken Goldberg put it in one number in Science Robotics on 27 August 2025: a 100,000-year data gap. The text and images that produced modern language and vision models amount to roughly 100,000 years of human reading and looking, and they existed already, for free, because people had spent a century putting them online. The equivalent corpus for manipulation — video paired with the motor commands that produced it — does not exist and has to be manufactured one robot trajectory at a time, by teleoperation, at human wages. Goldberg's own timing call, given to Berkeley's CDSS on 2 September 2025, is deliberately unexciting: "I'm not saying it's not going to happen, but I'm saying it's not going to happen in the next two years, or five years or even 10 years," and his stated reason for saying so is to head off "a bubble that could lead to a big backlash."

Physical Intelligence made the same point from inside the effort to beat it, in a post published 22 December 2025 titled, in so many words, "Moravec's Paradox and the Robot Olympics": "we can't program physical intelligence because we don't actually understand it at a conscious level," and language models do not supply it because "they are trained on human communication (i.e., text from the web), which does not communicate physical skills." Their reported result is the part a reading should keep: their π0.6 model scored an average of 72% task progress across ten domestic tasks on under nine hours of data per task, while a vision-language model without robot pre-training scored 9%.

The two explanations agree on the observation and disagree about what follows. Moravec's says the difficulty is intrinsic to the competence and indexed to evolutionary age. The data account says the difficulty is indexed to how much recorded experience of the task exists in a form a learner can eat. They make opposite predictions about which physical tasks fall first, and section 3 shows the record has been answering.

---

Why a reading would cite it

The admission test is easy here — this project has a standing lens for robotics and embodiment, and the paradox is that lens's oldest question. The discipline is the hard part, so most of this section is the discipline.

The occasion the proposal named, and it happened

proposals.md says: cite when olympiad golds land the same month as robots that still fumble doors. That month is on the record.

On 16 July 2026, at the International Mathematical Olympiad in Shanghai, AI systems from Huawei and Xiaohongshu each took an officially graded 42/42 — the first perfect machine scores, against seven of 666 human contestants who managed the same. Two weeks later, on 30 July 2026, DeepMind released Gemini Robotics 2 with whole-body humanoid control and five-finger dexterity, and published the success rates: 76.3% picking an object off a shelf, 68.4% off a table, 45.7% off the floor.

Fourteen days apart: a machine that is perfect at the hardest examination a seventeen-year-old mathematician will ever sit, and a machine that picks something up off the floor slightly less than half the time. That is the sentence from 1988 with 2026 numbers in it, and it is why the entry exists.

But the citation has to be made carefully, and the care is the point. The comparison is only worth making because the robotics number has a denominator. A published 45.7% is a measurement; a video of a robot fumbling a door is not. The project's own robotics ledger makes exactly this call — it treats DeepMind publishing an unflattering rate as the notable event, because it is the first window in which humanoid capability claims came with a denominator. A reading that reaches for Moravec should reach for the number, not the fumble.

Where the paradox is being falsified in the same record, which is the better citation

A reading that only ever cites this entry on the AImageddon-adjacent side of robotics is misreading its own ledger. Driving is a perception-and-mobility task — the exact pair Moravec named — and it is deployed.

Waymo went fully driverless in Las Vegas including the Strip on 8 July 2026, on a fleet then running roughly half a million paid rides a week, with San Diego, Tampa and Denver queued behind it. On 23 July 2026 the Insurance Institute for Highway Safety — an insurance body with no stake in the technology's success — compared about 50 million Waymo driverless miles against 222 billion human-driven miles in the same four cities and found 68% fewer crashes, 81% fewer injury crashes and 85% fewer single-vehicle crashes per mile. Aurora and Kodiak are running driverless freight commercially; 255,000 acres of Florida sugarcane are farmed by unmanned tractors under one supervising operator; Volvo and Boliden moved close to 700,000 tonnes of rock autonomously.

None of that is a machine with the skills of a one-year-old. It is something stranger and more useful to a reading: machines with the sensorimotor competence of a professional adult in a narrow domain that has a data flywheel, and nothing like the general competence of a toddler in a domain that does not. A fleet carrying 500,000 people a week is a data-collection apparatus. A humanoid in a kitchen is not, yet. The 2026 record separates by data availability, not by evolutionary age, and a reading that cites the paradox without saying which of those it thinks is operating has cited a mood.

The economic occasion, which is where the misuse is expensive

The most common load-bearing use of this entry in 2026 is not about robots. It is about jobs: reasoning is cheap for machines and hands are expensive, therefore the office goes first and the trades are safe.

The first half has a filing behind it. Oracle's FY2026 10-K, filed 22 June 2026, shows headcount down from 162,000 to 141,000 across the fiscal year, $1.84bn in severance against $374m the year before, and the sentence "the adoption and deployment of AI technologies across our operations have resulted, and may continue to result, in reductions to our workforce." Cloudflare's 8-K of 7 May 2026 cuts about 1,100 people — 20% of the company — to move to an "agentic AI-first operating model," announced alongside record quarterly revenue. Those are securities disclosures, where lawyers price the cost of being wrong.

The second half — therefore the trades are safe — does not follow from anything in Moravec, and section 4 takes it apart. What a reading may legitimately say is narrower and still worth saying: the observed ordering of displacement in 2026 matches the ordering the paradox implies, and the paradox is not the reason. The reason available in the record is that a cognitive task's training data is the work product itself, already recorded, already digital — the 15 August 2026 reporting on defunct startups' Slack archives and Jira tickets selling at $10,000–$100,000 a company is that mechanism becoming a market — and a manipulation task's training data has to be manufactured. Cite the paradox for the shape; cite the data gap for the cause; do not let the first stand in for the second.

Why LENSES.md already contains this entry

LENSES.md instructs the robotics lens that "what was demonstrated under controlled conditions and what is actually deployed and working are different findings; say which." That rule is Moravec's paradox in operational form. The gap between a demo and a deployment is exactly the gap the paradox describes, because a demonstration removes the variance that sensorimotor competence exists to absorb. This entry is the citation for why that lens has a rule the other lenses do not.

Where a reading must not cite it

Not to dismiss a robotics result. If a machine did the thing, it did the thing. The paradox bears on what is comparatively hard to build, never on whether something that has been built is real.

Not to date anything. There is no rate in it, no threshold and no horizon. "Humanoids are ten years away" is a forecast; under LENSES.md it goes in the prose labelled as judgment, and Moravec's name does not convert it into a finding.

Not to decide who keeps their job. See above and section 4.

And not to make an empty lens mean something. This is the specific temptation in this room. The robotics lens read "Nothing notable found since the last reading" in both the 15 August midday and 16 August midnight readings. The paradox is a ready-made explanation for that emptiness and it is a bad one: a nine-hour Saturday window with nothing in it is a fact about the window. An explanation is not a finding, a quiet lens is not evidence for the difficulty of embodiment, and reaching for a 1988 sentence to make a blank section feel meaningful is padding with a citation on it.

Nothing in this file is evidence. The 2026 items above are named as citation occasions, not re-argued; each carries its own primary source in the genesis ledger or in section 5. Nothing here is deposited, nothing here grades a lab or a model, and nothing here touches the needle.

---

What it got right, and what it got wrong

Not required for interpretation. Done in full anyway, under prediction discipline, because the claim is an ordering and an ordering is gradeable, because the author is a professional forecaster, and because a canon that grades Kurzweil and Turing should grade the observation it uses most.

The claim, the dates

Claim, in its own words: "it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility."

Date made: 1988, Mind Children, Harvard University Press.

Date due: none stated, and that is a defect in it as a prediction, graded as such below. The nearest thing to a due date is supplied by the author elsewhere rather than by the claim: the 1998 landscape essay puts the peaks — locomotion, hand-eye co-ordination, social interaction — under water "within another half century," which is 2048; and his separate forty-year forecast falls due in 2028. The world has supplied a de facto grading window anyway, because the claim's two halves have both been tested hard since 2012.

What actually happened: the claim has three separable parts and they have graded differently. Taking them apart is the whole of this section's value.

(a) The ranking: right, for thirty-eight years, and still right at the frontier

The ordering held. Games fell first and completely — checkers, chess, Go, poker. Theorem proving followed. By 2026 the abstract end of the landscape is under water past anything Moravec described: officially graded perfect IMO papers on 16 July 2026, and open problems in mathematics falling to machine search. Meanwhile the best published number for a general-purpose humanoid picking an object off the floor is 45.7%.

Thirty-eight years is an extraordinary run for an uninstrumented observation made by one person about the state of a young field, and the honest grade on the ranking is: right, and right for longer than the person making it had any way to justify.

(b) The magnitude: wrong, and the wrong word is "impossible"

"Difficult or impossible" was a live disjunct in 1988 and it is dead. Perception in Moravec's sense — the retina calculation that anchors his whole estimate — is substantially solved; alexnet-2012 is the entry that dates it. Mobility in unstructured public space is not merely possible but commercially deployed and, on an insurance body's independent measurement of 50 million miles, safer per mile than the humans it replaces. Locomotion on legs, one of Moravec's three named peaks, is a solved engineering problem to a standard that would have been science fiction when he wrote.

There is no honest reading of the 2026 record on which perception and mobility are impossible. Anyone still quoting the sentence with that clause attached is quoting a claim that has been falsified in public, at scale, with a denominator.

(c) The mechanism: probably wrong, and this is the most useful failure

Moravec's explanation predicts that difficulty for machines tracks evolutionary age: the older the competence, the more optimised, the harder to reimplement. Run that against the order things actually fell.

Vision is among the very oldest competences — eyes are half a billion years old, and Moravec's own retina calculation makes vision the anchor of his estimate. It fell substantially in 2012, before dexterous manipulation, before locomotion in the wild, before almost everything else on his list.

Language is recent on Moravec's own clock — his sentence about abstract thought puts the recent stuff at "less than 100 thousand years," and language belongs there. On his mechanism it should have been comparatively cheap. It was the single most stubborn problem in AI for fifty years and fell only in 2020–2022, a decade after vision.

Locomotion is very old and largely fell; contact-rich manipulation in clutter is the same age and did not.

The mechanism does not order the observations. Something else does, and it is available in the same record: how much recorded experience of the task exists in a form a learner can consume. Text, trillions of tokens, already online, free. Images, billions, already online, free. Video, abundant but weakly coupled to the actions that produced it. Manipulation trajectories, thousands of hours, produced one at a time by a human holding a controller. Goldberg's 100,000-year gap is that ranking stated as a number, and the ordering it predicts is the ordering the world produced, including the two cases — vision early, language late — where Moravec's mechanism gets the sign wrong.

This is not a small correction, because the two accounts license opposite forecasts. On the evolutionary account, the remaining hard tasks are hard intrinsically, and a reading should expect slow grinding progress indefinitely. On the data account, the remaining hard tasks are hard contingently, and a reading should expect them to fall in the order that data flywheels appear — which is why driving, where a fleet generates data at 500,000 rides a week, is deployed, and general household manipulation, where nothing generates data, posts 45.7%. The data account is checkable and the evolutionary account is not. A reading that wants to say something falsifiable about robotics should use the first and cite the second only for the observation.

(d) The bar: both sides of the modern argument are arguing past the sentence

This one is small, unremarked elsewhere as far as I could find, and it decides a lot of bad arguments.

Moravec's bar is "the skills of a one-year-old." A one-year-old pulls to stand, walks a few steps, grasps clumsily with a whole hand, recognises faces, and puts things in its mouth. It does not fold laundry, make a sandwich, scrub a greasy pan or turn a sock right side out.

Those four are on Physical Intelligence's Robot Olympics list — which is to say the standard modern demonstration that the paradox is being beaten is aimed at adult domestic competence, well above Moravec's stated bar. And the standard modern demonstration that the paradox holds — a humanoid falling off a stage, a robot fumbling a door handle — is aimed at balance and gross motor control, at or below it. Both sides collect evidence about a bar Moravec did not set.

Graded against the bar he actually set, the position in 2026 is genuinely ambiguous and a reading should say so rather than pick a side: machines exceed a one-year-old at recognition by an enormous margin, match or exceed it at locomotion, and remain below it at general grasping in an unfamiliar cluttered scene — which is roughly what 45.7% off the floor means.

The creator's own grade, and independent ones — and they differ, in two directions

This is what the prediction rules exist for, and this case is more interesting than the Kurzweil pattern the rule was written for.

Moravec's grade of Moravec: deflationary. Now 77, retired from the field by illness, he told New York magazine in July 2026 that he is "a spectator at this point," and when pressed on his record as a forecaster answered, "You get lucky sometimes." He has not claimed the paradox, did not name it, and did not build a research programme on it — he went and co-founded a warehouse-robotics company in Pittsburgh in 2003 instead, which is a more interesting response to your own pessimism than restating it.

The independent grades point opposite ways, and it is the observation rather than the forecast that comes off worse. On the timeline, outside opinion is generous: New York's framing is that he was right about AI, and gwern's verdict — "It's Hans Moravec's world, and the rest of us were just living in a fool's paradise" — is about as strong as an outside grade gets. On the paradox, outside opinion is harsh. Arvind Narayanan's "Fact checking Moravec's paradox" (29 January 2026) states flatly that it "has never been fact checked" despite decades of repetition by serious researchers, argues that the apparent inverse correlation is selection bias — you get it by ignoring two quadrants of the 2×2, the tasks easy for both (tic-tac-toe, "how bright is this image") and the tasks hard for both (predicting stock prices, picking hit songs) — and concludes that the paradox is "really a statement about what the AI community finds it worthwhile to work on" rather than a claim about difficulty. He calls the evolutionary reasoning "highly dubious" on the ground that reasoning in open-ended settings is not separable from the sensorimotor substrate it runs on.

So the pattern is: the man grades himself low, the world grades his forecast high, and the specialists grade his famous observation low. Three grades, three directions, and the one thing everybody quotes is the one that reviews worst.

Two things follow that a reading can use. First — and this repeats the point goodharts-law-1975 made about deflationary self-grades — a base rate for AI forecasting assembled only from people who promote their own hit rate is a base rate about promoters. Moravec is the second entry in this canon whose author scores himself below the outside view, and both are worth more to the base rate than a confident self-assessment. Second, and practically: Narayanan's selection objection is the right first question to ask of any instance of the paradox in the wild. What are the tasks that are hard for both, and why are they not in your argument?

What the objection does not reach

Narayanan's case is the strongest in the file and it is not total, and section 2 would be dishonest without saying where it stops. The selection argument explains away the correlation — why the pattern looks like a law — but it does not explain away the residue, which is real, measured, and has a mechanism. Manipulation in unstructured environments is genuinely behind, the gap has a number attached (72% task progress on nine hours of data per task, against 9% with no robot pre-training; 45.7% off the floor), and the reason offered for it by the people closest to the work is not evolutionary mysticism but a missing corpus. "The community chose to work on other things" does not by itself explain a 100,000-year data gap that the community did not choose and cannot wish away. The correct 2026 position is that the paradox's explanation is wrong, its magnitude clause is dead, its correlation is inflated by selection, and its residual observation is nonetheless the live constraint on embodied AI — which is a useful thing to be able to say in one sentence, and is what this entry is for.

---

Commonly misused as

Not required for interpretation, and included because it is why the entry earns its place. This one is misused in a specific and predictable way: it is deployed as a limit by people who would not defend it as one if asked.

1. As a law that forecloses something — "robots will never…"

The commonest and the worst. There is no proof, no hypotheses, and no entailment. The specific tell is the form "Moravec's paradox says X, therefore Y." Nothing follows from Moravec's paradox, because it is a comparative observation with no rate.

The decisive rebuttal is the author's own: the landscape metaphor is a rising waterline, not a wall. Moravec named locomotion and hand-eye co-ordination as the last peaks to be submerged and gave a date for the submerging. Using his name to argue permanence contradicts him with his own picture, and the 2026 record contradicts it with 50 million driverless miles. Note the parallel this canon already documents: this is the pattern godel-incompleteness-1931 and goodharts-law-1975 describe, running here in a third register. The field abuses its favourite theorem, its favourite economics aside and its favourite robotics observation by the identical move — converting a scoped claim into a general licence to declare something foreclosed.

2. "Therefore blue-collar jobs are safe."

The expensive one, and live right now. The chain is: reasoning is cheap for machines, hands are expensive, therefore the office is exposed and the trades are not. The premise is roughly true, the conclusion does not follow from it, and three separate steps are missing.

First, "hard to automate" is not "safe." A trade can be devastated by partial automation of the parts around it — scheduling, estimating, diagnosis, dispatch — without a robot ever touching a wrench.

Second, the paradox says nothing about timing, and timing is the entire question for a person deciding what to train for. A displacement that arrives in forty years is a different life from one that arrives in eight, and Moravec supplies no basis for choosing between them. Goldberg does supply one — not two, five or ten years for general-purpose humanoids — and he is a roboticist offering it explicitly to deflate a bubble, which is the opposite of the usual incentive and worth more for it.

Third, and decisively, the mechanism is wrong, so the ranking it produces is not the ranking that will happen. On the data account the exposed physical jobs are the ones inside a data flywheel — driving, warehouse picking, long-haul freight, row-crop agriculture — regardless of how ancient the underlying skill is. Every one of those is in this project's ledger as already deployed and paid for, while the humanoid-in-a-home case that the "trades are safe" argument is really about is still shipping pilots. The paradox, used as career advice, gets the ordering backwards for exactly the jobs it is most often used to reassure.

3. The reverse misuse — "vision is solved, so the paradox is dead."

Aimed at people who have read a section like this one and overcorrected. The magnitude clause is dead and the mechanism is wrong; the residue is not. Two numbers hold the line: 45.7%, and the 9% a strong vision-language model scores on domestic manipulation without robot pre-training. Whatever is missing is not supplied by scaling the text-and-image stack, and everyone actually building these systems says so. "The paradox is debunked" and "the paradox is a law" are the same error in opposite directions — both replace a measured gap with a slogan.

4. Attributed wrongly, or quoted as Pinker's sentence under Moravec's name.

"The hard problems are easy and the easy problems are hard" is Pinker, 1994, and it is a strengthening rather than a paraphrase: Moravec's sentence is comparative and hedged ("comparatively easy," "difficult or impossible"), Pinker's is a clean inversion with no hedge. Minsky's 1986 formulation is about self-knowledge, not about machines, and is the one that has aged best. Newell called the whole proposition a myth in 1983. A reading that wants the strongest version of the claim should quote Pinker and say so; a reading that wants the defensible version should quote Minsky. Quoting the strong version and attributing it to Moravec gets both the author and the claim wrong at once — the same failure goodharts-law-1975 documents for Strathern, in a canon that now has two instances of it and should probably expect more.

5. "It proves intelligence requires a body."

The embodied-cognition overreach. Moravec's claim is about engineering difficulty, not about the constitution of intelligence, and nothing in it entails that a disembodied system cannot be intelligent — a proposition the 2026 record does not support anyway. The related and better claim, which Narayanan makes and which is a genuine dispute rather than a corollary, is that abstract reasoning may not be separable from the sensorimotor substrate it evolved on top of. That is an argument about architecture, it is contested, and it is not what the paradox says.

6. Confused with the AI effect, and with the bitter lesson.

Three distinct claims that arrive in the same conversation. Moravec: sensorimotor competence is comparatively harder to build than abstract reasoning. The AI effect: whatever machines can do stops counting as intelligence, so the goalposts move — a claim about how people react to results, not about difficulty. The bitter lesson (Sutton, 2019; proposed as bitter-lesson-2019, unwritten): general methods that scale with compute beat handcrafted structure. The third is the sharpest citation for why the paradox's residue is shrinking at all, and it is routinely mislabelled as the first. If the argument is "they said it was hard and then it wasn't," that is Sutton or the AI effect, not Moravec.

7. The one this room has to watch in itself.

Rule 3 requires both cases before the needle. This entry is unusually convenient for one of them: an empty robotics lens plus a 1988 sentence produces a respectable-sounding AImageddon paragraph out of nothing having happened. It is equally available to the other side — "the paradox is falling, look at Waymo" is an AItopia paragraph obtainable at the same zero cost.

That symmetry is the tell. A citation available to both sides of a reading decides nothing, and a reading that lets it decide something has used a canon entry as a mood. The defence is the one this file has been applying throughout: the paradox is a demand for two facts — what was demonstrated, and under what conditions — and a reading that cannot supply them has no business invoking it. Which is, word for word, the rule LENSES.md already sets for the robotics lens.

---

Sources

A general caveat that applies to everything below. Web pages in this session were retrieved through a fetch layer that returns a summarising model's rendering of the page rather than the raw text. Passages marked as quotations were returned as quotations by that layer and agree across independent sources where I could check them, but no quotation in this file has been verified against a page I read character by character, and the entry says so once here rather than hedging every sentence.

**Primary, and the honest state of it: I did not read Mind Children. The 1988 book was not obtained. A borrowable scan exists at archive.org/details/mindchildren00hans and a later session with library access should check this file against it — in particular the page attributions and whether the word "paradox" appears anywhere near the passage. An excerpts page at cse.iitk.ac.in was fetched and did not contain any of the quoted passages**; it did carry a secondary note that Moravec expected human-level robots "within next 50 years i.e by 2038," which conflicts with the forty-years-from-1988 framing used elsewhere and is one reason the 2028 forecast is left to its own future entry rather than graded here.

• **Hans Moravec, Mind Children: The Future of Robot and Human Intelligence, Harvard University Press, 1988. Both quoted passages — the "one-year-old" sentence at p. 15 and the "billion years of experience / prodigious olympians / thinnest veneer" passage at pp. 15–16 — are taken from the Wikipedia article on Moravec's paradox, fetched directly, which supplies those page numbers; the long passage was independently returned by a separate search against three unrelated pages with identical wording. The page numbers are Wikipedia's, not mine.**
• **Hans Moravec, "When will computer hardware match the human brain?", Journal of Evolution and Technology vol. 1, 1998 (received December 1997)**, fetched at jetpress.org. This is the one Moravec text reached in something close to its original form, and it is the source for the landscape-of-human-competence passage, the flooding metaphor and the half-century horizon on the peaks, the 100-million-MIPS estimate and its derivation from the retina, and "computers suitable for humanlike robots will appear in the 2020s."
• **Marvin Minsky, The Society of Mind, 1986, p. 2; Steven Pinker, The Language Instinct, 1994 (page cited as p. 190 of the 2007 edition); Rodney Brooks (2002, quoted via McCorduck 2004, p. 456); Allen Newell (1983) on the "myth". All four are from the same Wikipedia article, which is the only source consulted for them. None of the four originals was read**, and the Newell attribution in particular is carried here on one source's presentation of a single sentence; a session that wants to lean on "the first published position was against it" should find the Newell text first. Wikipedia's own summary of Narayanan is also the source for the sentence quoted from it there.
• Hans Moravec, biographical. Born 30 November 1948 in Kautzen, Austria; Stanford PhD 1980 under John McCarthy on the Stanford Cart; at Carnegie Mellon's Robotics Institute from 1980, research professor from 1995, adjunct from 2005; co-founder of Seegrid Corporation, Pittsburgh, 2003. From the Wikipedia article on Moravec, fetched directly. The Stanford Cart figures used in the preamble — rebuilt with stereo vision in 1977, crossed a chair-filled room without human intervention in 1979 in about five hours, one-metre spurts with ten-to-fifteen-minute pauses, on a KL10 of roughly 2.5 MIPS — are from search-result summaries of the Computer History Museum's catalogue entry, a Stanford AI timeline and cyberneticzoo.com, which agree with each other; none of those pages was fetched directly, and the 2.5 MIPS figure in particular rests on a single summary.

On Moravec's own grade, which is the weakest sourcing in the file. Dan Kagan-Kans's interview with Moravec ran in New York magazine's Intelligencer on 24 July 2026 at nymag.com/intelligencer/article/hans-moravec-interview.html. The article could not be fetched — this machine's retrieval layer refuses nymag.com — and I did not read it. The quotations attributed to Moravec above ("a spectator at this point," "You get lucky sometimes"), the detail that he retired from the field after developing myasthenia gravis, and gwern's "It's Hans Moravec's world, and the rest of us were just living in a fool's paradise" all come from search-result summaries of an MSN syndication of that article, at one remove, and should be re-verified before any of them is quoted in a digest. AI Weekly's summary (25 July 2026), fetched directly, confirms the article's existence, author, URL and framing but carries no quotations. Gwern's "30 Questions for Hans Moravec" (22 April 2026), fetched directly, turns out to be a list of proposed questions, not an interview — it records that another journalist secured the interview instead, which is how the New York piece came to exist. That page contains no Moravec answers and was not used for any claim above.

On the independent grade of the paradox. Arvind Narayanan, "Fact checking Moravec's paradox", 29 January 2026, fetched twice with different prompts and consistent results — source for the selection-bias argument, the two ignored quadrants and their examples, "has never been fact checked," "really a statement about what the AI community finds it worthwhile to work on," and the "highly dubious" verdict on the evolutionary reasoning. This is the entry's most important secondary source and it is a blog post rather than a peer-reviewed result; it is treated above as a strong argument, not as a finding. Ege Erdil, "Moravec's paradox and its implications", Epoch AI, 27 December 2024, fetched directly — read for the opposite case, which is that the paradox is a usable predictor of automation order via evolutionary optimisation pressure plus data availability. Its brain-compute figures (1e14–1e16 FLOP/s, 1e23–1e25 FLOP of "training") are noted here and not used above, because nothing in this entry should turn on a brain-equivalence estimate.

On the data-gap account. Ken Goldberg's Science Robotics editorial of 27 August 2025 on closing the "100,000-year data gap" is known here through Berkeley CDSS's write-up (2 September 2025), fetched directly, which is the source for both Goldberg quotations and for the framing of the gap. The editorial itself was not read, and neither was the Science Robotics debate "Data will solve robotics and automation: True or false?" that appeared alongside it — a later session with journal access should read both. Physical Intelligence, "Moravec's Paradox and the Robot Olympics", 22 December 2025, fetched directly — source for the π0.6 results (72% average task progress, under nine hours of data per task, 9% for a VLM baseline without robot pre-training), the ten-task list, and both quoted sentences. This is a vendor publishing its own numbers about its own model, which is exactly the claimed-versus-demonstrated distinction the readings maintain, and it is used above only for the shape of the argument and for the 9%-versus-72% contrast, never as an independent measurement. Nathan Lambert, "Can robotics take off like GenAI? Moravec's paradox vs. scaling laws", 29 September 2023, fetched directly — read as corroboration that the data-bottleneck reading of the paradox predates the current humanoid cycle; not quoted above. Aaron Prather, "The Year of the Hand", 9 January 2026, fetched directly — read and not used, because its claims about hands are qualitative and this file's argument is built on numbers; recorded here so a later session does not spend the fetch again. Beren Millidge's review of Mind Children, 12 January 2025, fetched directly — the only reader's grade of the book found; source for the observation that Moravec's hardware extrapolation held up well and his robotics-deployment expectations did not. Used for context, not quoted.

On the 2026 record, all of which is in this project's own genesis ledger with primary links, and none of which is evidence in this file. The IMO results of 16 July 2026 (Taipei Times, 24 July); Gemini Robotics 2 and the 76.3% / 68.4% / 45.7% pick rates on Apptronik's Apollo 2, 30 July 2026 (DeepMind); Waymo's driverless expansion and roughly half a million paid rides a week, 8 July 2026 (Waymo); the IIHS comparison of about 50 million driverless miles against 222 billion human miles, 23 July 2026 (IIHS); Aurora, Kodiak, U.S. Sugar and Volvo/Boliden, all carrying their own primary links in history/genesis/robotics-and-embodiment.jsonl; Oracle's FY2026 10-K of 22 June 2026 and Cloudflare's 8-K of 7 May 2026, in history/genesis/work-and-the-economy.md; and the Slack-archive training-data market reported 15 August 2026, in digests/2026-08-16-00.md. The two "Nothing notable found" robotics lenses cited in section 2 are from digests/2026-08-15-12.md and digests/2026-08-16-00.md. All of these are named as citation occasions. This file deposits nothing in the ledger, grades no lab and no model, and places no needle.

Two notes for whoever writes here next. First, proposals.md has no entry for Moravec's 2028 forecast and should — it comes due inside this project's operating life and it is the cleanest gradeable prediction available to the canon. Second, canon/short-circuit-1986.md and canon/terminator-1984.md both contain paragraphs written against this file before it existed, and both explicitly say that moravecs-paradox-1988 is the entry that should be cited alongside them. It now exists. The commission for this run forbids touching another entry's file, so those two descends_from lines are left unedited deliberately rather than overlooked; they are one line each, and they are the owner's or a later session's to make.