Unsteady Constants and Useful Mistakes · Morning Gist
Tuesday, October 135 stories · 13 min
Unsteady Constants and Useful Mistakes
A cosmic number that keeps drifting, a test for arguing fairly, and two studies on how AI shapes, and differs from, the way we think.
Picked from 45 new articles across 14 sources · arrow keys for other days
Today in 30 seconds
30s left
The Hubble constant keeps changing. Ethan Siegel says it only settles once matter has thinned so far that nothing is left to measure it. story 1 ↓
In Hesse's 1918 poem "Books", reading sends you back to yourself. Popova finds Baldwin and the young Proust reached the same idea. story 2 ↓
Dennett's four rules for fair criticism, taken from tit-for-tat's inventor Anatol Rapoport: three generous steps come before you may rebut. story 3 ↓
A Berkeley study found adults gave up more often after their ChatGPT was taken away, on math and SAT reading, within minutes. story 4 ↓
79 humans and two macaques experienced the motion aftereffect illusion. None of nine AI vision networks did, which suggests machines don't see like us. story 5 ↓
Also: Francis Halzen and the IceCube neutrino detector, buried in Antarctic ice, are reported to have won the 2026 physics Nobel.
1/5Big Think· Ethan Siegel· Oct 13
Why the Hubble constant isn't constant, and when it might become one
The universe's expansion rate, misleadingly called the Hubble constant, changes over time. Ethan Siegel explains that it only stops changing if dark energy dominates completely, and by then there will be no galaxies left to measure it.
Why it matters · It shows that one of science's best-known numbers is a snapshot, and that how well we can know the cosmos depends on when we happen to be living.
Know this first
Context from the editor, not the article · 30 sec
1 · Redshift and expansion
Light from distant galaxies arrives stretched toward redder wavelengths. In the 1920s, astronomers found that galaxies farther away are more redshifted, so they appear to be moving away faster. This was the first evidence that the universe is expanding.
2 · Measuring cosmic distances
Astronomers measure distance with "standard candles": objects whose true brightness is known, such as pulsing Cepheid stars or Type Ia supernovae. Comparing how bright one looks with how bright it really is tells you how far away it is.
3 · Dark energy
In 1998, studies of distant supernovae showed that the expansion is speeding up, not slowing down. The unknown cause is called dark energy. Its simplest form is Einstein's cosmological constant, an energy built into space itself.
2/5The Marginalian (Maria Popova)· Maria Popova· Oct 13
Hermann Hesse's 1918 poem argues books point you back to yourself
Maria Popova revisits "Books," a short poem Hermann Hesse wrote in 1918. It argues that reading can't give you wisdom directly, but it can lead you to wisdom you already carry.
Why it matters · It questions a quiet assumption behind any reading habit: that wisdom piles up from what we read, rather than being sharpened by it.
Know this first
Context from the editor, not the article · 27 sec
1 · Who Hermann Hesse was
A German-born Swiss novelist (1877–1962) best known for Demian, Siddhartha, Steppenwolf and The Glass Bead Game. His recurring subject is the lonely inward search for a self. He won the Nobel Prize in Literature in 1946.
2 · Hesse around 1918
Near the end of the First World War, Hesse was in a personal and creative crisis and had begun psychoanalysis. Soon after, his writing turned firmly inward. This is general background; the article itself doesn't cover it.
3 · What The Marginalian does
Maria Popova's long-running, reader-funded site, formerly called Brain Pickings. It usually takes one passage from an older writer, frames it with a theme, and links it to related thinkers.
3/5The Marginalian (Maria Popova)· Maria Popova· Oct 6
Daniel Dennett's four rules for criticizing an opponent fairly
Maria Popova revisits a recipe the philosopher Daniel Dennett borrowed from game theorist Anatol Rapoport. Restate the other side's view, name where you agree, credit what you learned, and only then criticize.
Why it matters · It turns "be charitable" from a vague virtue into a checklist you can test against your next argument, and Dennett defends it as tactics, not just good manners.
Know this first
Context from the editor, not the article · 24 sec
1 · Who Daniel Dennett was
An American philosopher of mind (1942–2024) known for books on consciousness, evolution and free will. His 2013 book Intuition Pumps and Other Tools for Thinking collects small mental techniques for reasoning well.
2 · Who Anatol Rapoport was
A Russian-born American mathematician and psychologist (1911–2007) who studied conflict and cooperation. His simple tit-for-tat program won Robert Axelrod's famous computer tournaments of the repeated Prisoner's Dilemma around 1980.
3 · Straw men and steelmen
A straw man is a weakened caricature of an opponent's view, built so it is easy to knock down. Steelmanning, the opposite habit, means arguing against the strongest version of the view.
4/5Nautilus· Jake Currie· Oct 12
Adults given ChatGPT gave up more often once it was taken away
Berkeley researchers let some adults use ChatGPT on math and reading problems, then took it away. Those people then solved fewer problems and skipped more than people who had worked without it all along.
Why it matters · It tests a common worry, that leaning on AI wears down our willingness to struggle, and it puts the effect on a timescale of minutes rather than years.
Know this first
Context from the editor, not the article · 25 sec
1 · Productive struggle
Learning researchers have long argued that effortful difficulty, like getting answers wrong and retrying, is part of how skills stick. Tools that remove the effort can boost performance in the moment while weakening learning.
2 · Cognitive offloading
Handing mental work to an outside aid, such as a calculator, notes, GPS or now a chatbot. It frees attention, but people often get less practice at the skill they've handed off.
3 · The policy backdrop
Last month New York City restricted AI use for about 600,000 students from pre-K through eighth grade. The schools chancellor said children learn best by wrestling with hard problems.
5/5Nautilus· Kristen French· Oct 6
Humans and monkeys fall for a motion illusion that AI vision doesn't
York University researchers found that humans and macaque monkeys experience the motion aftereffect, an illusion in which still objects seem to drift after you watch steady motion. Nine AI vision systems did not, even when nudged.
Why it matters · It flips the usual test of machine vision: whether an AI sees like us may depend on whether it makes our mistakes, not just whether it gets the right answer.
Know this first
Context from the editor, not the article · 27 sec
1 · Illusions as clues
Vision scientists study illusions because they show the shortcuts the brain takes. A mistake that everyone makes in the same way points to a specific computation, not random noise.
2 · The waterfall illusion
Stare at a waterfall for a minute, then look at the rocks beside it, and they seem to creep upward. This is the classic motion aftereffect, usually explained by motion-sensing neurons tiring out and briefly losing their balance.
3 · Comparing brains and AI
Neuroscientists often test whether AI vision networks work like the primate visual system. Usually they check whether the networks recognize the same things or match recorded neural activity. Comparing their errors is a stricter test.
Hubble's 1929 distance–redshift plot beside a version about 70 years later: the line held to over a billion light-years.1 / 4
A constant that keeps changing
The 1 is the slope of one of astronomy's most famous graphs. Plot galaxies' distances against how fast they seem to be moving away, and the farther ones recede faster. Edwin Hubble published that roughly straight line in 1929, after Georges Lemaître (1927) and Howard Robertson (1928). In a new essay, astrophysicist Ethan Siegel argues the name is a misnomer: the expansion rate changes over time.
How the graph was made
Distances came from "2". Cepheid stars pulse on a cycle whose length reveals their true brightness, and comparing that with how bright they look gives their distance. Speeds came from spectroscopy: atoms leave lines at known wavelengths, and in distant galaxies those lines are shifted toward the red. The bigger the 3, the faster the galaxy is receding.
The line stayed straight for 70 years. Then, in the late 1990s, the most distant supernovae fell off it. This was the evidence for 4, the unknown cause of the accelerating expansion.
Why theory never promised a constant
Expansion follows the 5, Alexander Friedmann's 1922 solution of general relativity for a universe that looks the same everywhere on large scales. In these equations, the expansion rate depends on energy density. As space grows, matter and radiation thin out, so the rate falls; for both, H drops roughly as 1/time. Only a 6, an energy built into space itself whose density never thins, holds H steady.
Our universe contains all three ingredients, and each dominated in a different era. Looking far away means looking back in time, so the redshift–distance relation curves. Siegel calls this variation in space, but it is really time showing up as distance.
What the data say
Using standard candles alone, Siegel reports these best fits: today's rate is 71–76 7 (kilometers per second for each megaparsec, about 3.26 million light-years). Matter makes up 24–35% of the universe's energy and dark energy 65–76%, while radiation and spatial curvature are negligible. Dark energy behaves within about 20% of a cosmological constant. Recent hints that dark energy is weakening appear only when several datasets are combined, and Siegel calls that case weak.
The Catch-22
If dark energy truly is constant, it will eventually dominate. H will then level off at about 80–85% of today's value, roughly 55–60 km/s/Mpc, and become a true constant. But measuring expansion needs distant matter, and by then all of it will have moved out of view. Only our own merged Local Group will remain. Siegel's point is that the present is the brief window when the universe's history and fate can still be read.
A misnamed number
In 1929, Edwin Hubble plotted galaxy distances against how fast the galaxies appeared to be moving away from us, and he got a roughly straight line. A galaxy twice as far away receded twice as fast. The slope of that line became known as the 1. Georges Lemaître had found the same relationship in an obscure 1927 paper, and Howard Robertson found it independently in 1928, which is why it is now often called the Hubble–Lemaître Law. Getting an exact value for the slope was the main job of the Hubble Space Telescope.
In a new essay for Big Think, astrophysicist Ethan Siegel argues that "constant" is a complete misnomer. The expansion rate changes over time. Siegel adds that it isn't even uniform across space when we measure it today. And yet, he says, 2 may eventually make it a true constant, though only once no one can check.
How the line was built
The relationship needed two kinds of measurement. The first was distance. In 1923, Hubble identified a Cepheid variable star in the Andromeda galaxy. A Cepheid brightens and dims on a regular cycle, and the length of that cycle tells you how bright the star really is. Compare that true brightness with how bright it looks from Earth and you get its distance, the same way you could judge the distance of a light bulb of known wattage. This "standard candle" method has since been extended to many other kinds of objects, most importantly Type Ia supernovae.
The second was spectroscopy: splitting light into its separate wavelengths. Atoms, ions and molecules leave dark absorption lines and bright emission lines at precise wavelengths. Vesto Slipher noticed in the 1910s that in many galaxies these lines sit shifted toward the red end of the spectrum. The size of the shift gives a 3, which can be converted into a recession speed. Put distance and redshift on the same graph and you get Hubble's line.
Astronomers kept extending that line throughout the 20th century, and it stayed straight. Then, in the late 1990s, the most distant supernovae fell off it. The modern Pantheon+ sample of 1,550 supernovae traces that curve. The bend was taken as evidence for dark energy.
Why the theory never promised a constant
Siegel's key point is that theory never required the line to be straight. In Einstein's general relativity, gravity is the curvature of spacetime produced by matter and energy. Its field equations have no single general solution: you have to specify what fills the universe first. In 1922, Alexander Friedmann solved them for a universe that is isotropic (the same in every direction) and homogeneous (the same at every location), averaged over large scales. The resulting 4 link the expansion rate to the total energy density.
In these equations, the separation between any two distant points is tracked by a scale factor, a. The Hubble parameter is how fast a is growing as a fraction of its current size: H ≡ ȧ/a. How H changes depends on what dominates the universe:
Radiation: H falls as 1/t, and a grows as t^½.
Matter (ordinary and dark): H also falls as 1/t, and a grows as t^⅔.
A 5 (Λ): H stays fixed, set by √Λ, and a grows exponentially.
The reason is dilution. As space expands, the number of photons or atoms stays the same but the volume grows, so the density drops and the expansion rate drops with it. Radiation thins out even faster, because its wavelengths also stretch, so each photon carries less energy. A cosmological constant is energy belonging to space itself, so its density never falls. Only a universe dominated by Λ has a Hubble parameter that doesn't change.
For decades the default model was the Einstein–de Sitter universe, proposed in 1932, made entirely of matter. In that model, the value measured today, written H₀, is simply one moment's reading on a dial that keeps falling. Our actual universe contains radiation, matter and dark energy, and each dominated during a different era. When we look at distant objects we are seeing them as they were long ago, when the mix was different. So even a measurement made today does not produce a straight line.
A note on Siegel's claim about space: what the curved line really records is time. Distant galaxies show the expansion rate as it was when their light set out. Siegel calls this variation "in space", but that is a way of describing how distance and lookback time are tied together, not a claim that the expansion is faster in some regions.
Running the logic backwards
Cosmologists turn this around. They plot redshift against distance for objects near and far, then test models with different values for local H₀, radiation, matter, spatial curvature, dark energy density and dark energy's equation of state. The equation of state is a single parameter, w, describing how the energy's pressure relates to its density; a pure cosmological constant has w = −1. Using only the standard-candle distance ladder, Siegel reports these best fits:
H₀ ≈ 71–76 6 (kilometers per second for each megaparsec of distance, about 3.26 million light-years)
matter: 24–35% of the total energy density
dark energy: 65–76%
radiation: below 0.1%
spatial curvature: below 0.3%
dark energy's equation of state: within roughly ±20% of a cosmological constant
Is dark energy really constant?
Results from 7 (the Dark Energy Spectroscopic Instrument, which maps galaxies in 3D) have made headlines suggesting dark energy is weakening. Siegel says the evidence isn't strong. DESI's data on its own is consistent with constant dark energy. The case for evolution appears only when DESI is combined with cosmic microwave background and supernova datasets. He expects large-scale structure surveys from Euclid, the Vera Rubin Observatory and the Nancy Grace Roman Space Telescope to settle the question.
The Catch-22
If dark energy is a true cosmological constant, matter will keep thinning until Λ makes up essentially all of the energy. H will then level off at about 80–85% of today's value, roughly 55–60 km/s/Mpc. That figure matches √0.7 ≈ 0.84: in a Λ-only universe, the expansion rate depends only on the dark energy fraction. At that point the Hubble constant finally deserves its name.
But expansion can only be measured with matter: objects at some distance that give off or absorb light. A universe where H has stopped changing is one where the matter has effectively thinned to nothing. Siegel notes that by then the only galaxy left in the observable universe will be the merged remains of our Local Group, our galaxy's small gravitationally bound cluster. Everything else will have moved beyond view. His conclusion is that our own era is the lucky one. It is now, not then, that we can reconstruct the Big Bang, detect dark energy and predict a fate whose evidence will later disappear.
Remember
Matter and radiation thin out as space expands, so the expansion rate falls. Only constant dark energy keeps it steady.
Using standard candles alone, the best fits are 71–76 km/s/Mpc today, about 30% matter and 70% dark energy.
Hints that dark energy is weakening appear only in combined datasets; Euclid, Rubin and Roman should settle the question.
Cover of The Seasons of the Soul, the posthumous Hesse collection where the poem appears.
The newest post on 1, Maria Popova's reader-funded site of literary commentary, is built around a short poem by Hermann Hesse, the German-born Swiss novelist behind Siddhartha and Steppenwolf. The poem is called "Books." Hesse wrote it in 1918, as he entered his forties, nearly three decades before his 1946 Nobel Prize. It appears in the posthumous collection 2.
The paradox
Popova starts by listing what reading gives us. Books show us what it's like to be someone else and then return us to ourselves. They teach us how to love, grieve and be glad. They also offer company in hard times. She quotes James Baldwin: "You think your pain and your heartbreak are unprecedented in the history of the world, but then you read."
Then she turns. "No example, no parallel, no borrowed wisdom is a substitute for life itself," she argues. Wisdom comes only from "the deepening truth of our own experience." Reading helps us live, but it can't do the living for us.
The poem
Hesse's three short stanzas work through that tension:
All the books of the world / will not bring you happiness, / but build a secret path / toward your heart.
What you need is in you: / the sun, the stars, the moon, / the illumination you were seeking / shines up from within you.
The quest for wisdom / made you comb the libraries. / Now every page speaks the truth / that flashes forth from you.
The poem moves in three steps. Books are a path, not a destination. The light the reader was looking for was already inside them. And in the last stanza the direction reverses: the seeker who searched libraries for truth now finds that every page reflects back a truth the reader brings. The book hasn't changed; the reader has.
Proust got there too
Popova pairs the poem with the young Marcel Proust's reflections on why we read. Proust wrote that "the end of a book's wisdom appears to us as merely the start of our own," because "the essential book, the one true book… already exists in every one of us." On this view, a good book takes you to the edge of its insight and leaves the rest to you.
What's left open
This is a short anthology entry, not a long argument. Popova supplies the framing and the pairings but doesn't test the claim. One question she leaves alone: if the inner light shows up only after combing the libraries, the reading may be doing more work than the poem admits. She also doesn't connect the poem to Hesse's situation in 1918. He was coming out of a wartime crisis and had recently begun psychoanalysis, which fits a poem about turning inward.
The newest post on 1, Maria Popova's reader-funded site of literary and philosophical commentary, is built around "Books," a twelve-line poem by Hermann Hesse. Hesse (1877–1962) was the German-born Swiss novelist behind Siddhartha and Steppenwolf. He wrote the poem in 1918, just as he was entering his forties. It appears in the posthumous collection 2: The Poetic Guidance and Spiritual Wisdom of Hermann Hesse. The piece is short: a framing argument, the poem itself, and a related passage from Marcel Proust. Most of what it offers is the argument, so that is what this summary follows.
The paradox Popova sets up
Popova starts by listing what reading does for us. Books "show us what it is like to be another and at the same time return us to ourselves." We read to learn how to love, suffer, grieve and be glad, and to "3 our values." Annealing is a metalworking term: heating metal and cooling it slowly so it becomes tougher and less brittle. We also read for company. She quotes James Baldwin: "You think your pain and your heartbreak are unprecedented in the history of the world, but then you read."
Then she turns. Books may give us "a foothold" when life is disorienting, and a remedy for loneliness. But, she argues, "no example, no parallel, no borrowed wisdom is a substitute for life itself." The story of a life is written only "on the blank page of living," and real wisdom comes only from "the deepening truth of our own experience." So the tension is this: reading helps us live, but it can't do the living for us.
The poem
According to Popova, Hesse put this tension into verse more than a decade before his essay on "the timeless magic of books" and nearly three decades before his 1946 Nobel Prize. The poem has three stanzas (the post doesn't name the translator):
All the books of the world / will not bring you happiness, / but build a secret path / toward your heart.
What you need is in you: / the sun, the stars, the moon, / the illumination you were seeking / shines up from within you.
The quest for wisdom / made you comb the libraries. / Now every page speaks the truth / that flashes forth from you.
The poem works in three steps:
Books don't deliver happiness, but they build a path. In Hesse's image, reading is a road, not a destination. Its value lies in where it leads, and it leads inward ("toward your heart").
The light was always yours. The second stanza moves the light from the page to the reader. The light "shines up from within you."
The reading changes. In the last stanza the direction flips. The seeker combed libraries looking for truth. Now each page "speaks the truth / that flashes forth from you." The book hasn't changed; the reader has. Pages that once seemed to hold answers now reflect back something the reader brings to them.
One way to picture this (an analogy, not Hesse's): the book works like a tuning fork. It doesn't supply the note. It makes something in you ring that was already able to ring.
Proust's version
Popova pairs the poem with the young Marcel Proust's reflections on why we read. Proust wrote that "the end of a book's wisdom appears to us as merely the start of our own," because "the essential book, the one true book… already exists in every one of us." The passage comes from his essay "4," written as a preface to his translation of John Ruskin's Sesame and Lilies. Proust's point closely matches Hesse's: a good book takes you to the edge of its own insight and leaves you there, and the rest of the work is yours. The post doesn't say whether Hesse knew Proust's essay, so the agreement looks like two writers reaching the same idea independently.
What to make of it
This is a short anthology entry, not a sustained essay. Popova's contribution is the framing and the pairings, and she doesn't argue the point beyond them. Some questions she leaves open:
Is the claim true, or just comforting? "What you need is in you" can read as a reason to stop reading. Hesse's own first stanza resists that reading, since the books are what build the path. Still, the poem doesn't say what someone lacking good books, or good experiences, would find inside.
Where does the inner light come from? If wisdom only "flashes forth" after combing the libraries, the reading may be doing more than the poem admits. A more modest version of the claim would be that books don't add wisdom so much as give you words for experience you already have.
Why 1918? Popova notes Hesse's age but not his situation. He was near the end of a wartime crisis and newly in psychoanalysis, and soon afterward wrote the novel . That background fits a poem about turning inward, but the article doesn't draw the connection.
The post ends with links to related Marginalian pieces: Pythagoras on wisdom, Nick Cave on trusting yourself, Rebecca Solnit on how books console, and earlier entries on Hesse's ideas about courage and the inner voice.
Remember
Hesse treats reading as a path inward: books don't hand over wisdom, they lead toward it.
The poem's last stanza flips the direction, so the reader stops taking truth from the pages and brings it to them.
Proust independently argued that a book's wisdom ends exactly where the reader's own begins.
Maria Popova, author of the long-running essay blog The Marginalian, has returned to a short passage by the late philosopher Daniel Dennett (1942–2024). It answers his own question: "Just how charitable are you supposed to be when criticizing the views of an opponent?"
The problem
Most arguments aim to win, not to understand. Popova sets two older voices against each other. Arthur Martine's 1866 guide to conversation advised: "let your aim be to come at truth, not to conquer your opponent." Mark Twain's verdict was harsher: the critic is a dung beetle that lays its egg in somebody else's dung. The failure Dennett targets is caricature, or attacking a 1, a weakened version of a view built to be easy to knock down. "Be charitable" is easy to agree with and hard to do.
The four rules
In Intuition Pumps and Other Tools for Thinking (2013), Dennett gives what he calls "the best antidote." He credits it to Anatol Rapoport, the game theorist who devised 2, the strategy of cooperating first and then copying your opponent's last move. To criticize well:
Restate your target's position so clearly and fairly that they say, "Thanks, I wish I'd thought of putting it that way."
List points of agreement, especially ones that are not widely shared.
Mention anything you learned from them.
Only then say "so much as a word of rebuttal or criticism."
The first rule is the hardest. Your opponent, not you, decides whether you have understood them. That is a stricter test than 3, which means arguing against the strongest version of a view. The second rule's qualifier matters too. Agreeing on easy points proves nothing, while agreeing on contested ones shows where the real disagreement lies.
Tactics, not just manners
Dennett insists this is not naive niceness. According to Popova, the steps make your opponent "a more receptive audience for your criticism." People defend against critics who have misread them and listen to those who clearly understood them. The order also echoes Rapoport's game theory, though the article doesn't make this link: cooperate first, then deal with real conflict.
What's left open
The piece is brief. It does not give Rapoport's original wording or where he first set out the rules. It also does not deal with the obvious limits: opponents arguing in bad faith, the time the steps take, and views too muddled to restate fairly.
Maria Popova, who writes The Marginalian (a long-running, one-person essay blog), has gone back to a short passage by the late philosopher Daniel Dennett. It answers a question he put plainly: "Just how charitable are you supposed to be when criticizing the views of an opponent?" His answer is four rules, in a fixed order, that he credits to the social psychologist and game theorist Anatol Rapoport.
The problem: arguing to win
Popova frames the piece with two older voices. The first is Arthur Martine's 1866 guide to conversation: "let your aim be to come at truth, not to conquer your opponent. So you never shall be at a loss in losing the argument, and gaining a new discovery." The second is Mark Twain, who said the critic's symbol should be the tumble-bug (a dung beetle), which "deposits his egg in somebody else's dung, otherwise he could not hatch it."
Most arguments sit closer to Twain than to Martine. That is especially true online, where Popova says people "deploy the artillery of our righteousness from behind the comfortable shield of the keyboard." She calls this reacting rather than responding. The failure Dennett names is narrower: the "tendency to caricature one's opponent." Philosophers call this attacking a 1, a weakened version of a view built so that it is easy to knock down.
Advice to "be charitable" is easy to agree with and hard to put into practice. There is also a real tension. If you are too charitable, you never get to say what is wrong with the view. If you are too little charitable, you end up refuting a position nobody holds.
The four rules
Dennett gives the rules in Intuition Pumps and Other Tools for Thinking (2013), a book of small techniques for clear reasoning. He calls them "the best antidote" to caricature. To compose a successful critical commentary:
Re-express your target's position "so clearly, vividly, and fairly that your target says, 'Thanks, I wish I'd thought of putting it that way.'"
List any points of agreement, "especially if they are not matters of general or widespread agreement."
Mention anything you have learned from your target.
Only then "are you permitted to say so much as a word of rebuttal or criticism."
Each step does a different job. The first sets a test that only your opponent can mark. You have not understood a view until its owner would accept your summary of it, and preferably prefer it to their own. That is a stricter bar than 2 (arguing against the strongest version of a view), because it is judged by the other person.
The second step has a sharp qualifier. Agreeing that "cruelty is bad" costs nothing. The rule asks for agreement on points that are actually contested, and those show where the real disagreement starts and stops. The third step forces you to admit the exchange was worth something to you. The fourth is a gate. Criticism is not banned, it is postponed until it has been earned.
Here is an illustration of my own, not one from the article. Suppose you think a colleague's proposal to rewrite a system from scratch is wrong. Under the rules you first state their case better than they did: the old code makes every change slow, and patching it only spreads that cost over time. Then you note that you also think the current design has a fundamental flaw, which is not a common view on the team. Then you say what you learned, for example that the flaw costs more than you had estimated. Only after that do you argue that a gradual migration would get the same benefits with less risk. By then the colleague knows you are arguing with their actual position.
Why it's strategy, not niceness
Popova stresses that Dennett does not present this as "naively utopian" politeness. His claim is psychological. The steps turn your opponent "into a more receptive audience for your criticism," and that moves the discussion forward. People defend themselves against critics who have misread them. They listen to critics who have plainly understood them.
The author of the rules adds something the article only mentions in passing. Rapoport is best known for 3, a strategy for repeated games like the 4: cooperate on the first move, then copy whatever the other player did last. In Robert Axelrod's computer tournaments around 1980, this very simple strategy beat far more elaborate ones. It did so by being nice first, hitting back when provoked, and forgiving quickly. The four rules have a similar shape: open with cooperation, then deal with real disagreement. Popova does not draw this parallel herself. It is an interpretation, but a natural one.
What the piece leaves open
The article is short. It gives the rules and Dennett's justification, but not Rapoport's original wording or where he first set the rules out. (Dennett says only that they were formulated "decades ago.") It also does not deal with the obvious limits. The rules assume an opponent arguing in good faith, someone who might actually say "thanks." They take time that a comment thread rarely allows. And in some cases step one is impossible, because a position is incoherent and cannot be restated fairly in a way its holder would recognize. Whether that should count as a failure of the critic or of the position, the rules do not say.
Popova points readers to Susan Sontag's three steps for refuting an argument as a comparison, but she does not describe them here.
Remember
The first rule's test belongs to your opponent: you understand their view only when they endorse your restatement of it.
Agreement only counts if it's on contested points; agreeing on easy points shows nothing about where the disagreement lies.
Dennett defends charity as tactics: an opponent who feels understood actually listens to the criticism that follows.
A study presented this week at the 1 (COLM, a research venue for large language models) found that a few minutes of ChatGPT help made adults more likely to give up once the help was gone. The work comes from a team led by Brian Christian of the University of California, Berkeley.
The experiment
Adults solved fraction problems such as 5/6 minus 1/3. One group had a ChatGPT window open, and the other worked alone. Partway through the test, the chatbot disappeared and everyone finished unaided. The researchers measured two things: how often people solved problems, and how often they skipped them. The 2 works as a rough gauge of persistence.
What happened
With ChatGPT available, the AI group solved more problems and skipped fewer. Once the tool was gone, the groups swapped places: the people who had worked alone all along now solved more and skipped less. The team repeated the design with SAT reading-comprehension questions and got the same pattern.
"You almost couldn't ask for a clearer story," Christian said. The tools, he argued, are "often helping us in ways that are kind of unhelpful."
The bigger claim
The authors warn of a "3" problem. If regular AI use slowly wears down the motivation behind long-term learning, the damage could build up over years and be hard to reverse by the time anyone notices. That is an extrapolation. The study itself covered sessions lasting minutes.
The finding fits a live policy debate. Last month New York City restricted AI for about 600,000 students from pre-K through eighth grade, on the grounds that children learn by wrestling with hard problems. The study, however, tested adults.
Caveats
The Nautilus write-up gives no sample sizes or effect sizes, so we know the direction of the effect but not its size. Other explanations also fit what it describes:
Less practice: the control group spent the first block warming up on the problems, and the AI group didn't.
Contrast: losing a tool you just had may be more frustrating than never having it.
Either would mean something different from a lasting loss of grit. The article doesn't say whether the researchers ruled them out.
What stands out is the speed. The shift in behaviour appeared within a single sitting and held across two different kinds of task.
The news
A study presented this week at the 1 (COLM, an academic venue for research on large language models, now in its third year) reports that a few minutes of help from ChatGPT made adults more likely to give up once the help was gone. The work comes from a team led by Brian Christian of the University of California, Berkeley, who is also known for his books on AI and human decision-making.
The timing fits a wider argument. Last month New York City schools restricted AI for about 600,000 students from pre-K through eighth grade. Chancellor Kamar Samuels defended the move by saying children "learn best when they are wrestling with hard problems, when they're getting it wrong, when they're trying again and finally getting it right." The Berkeley study puts a version of that claim to an experimental test, though in adults.
The experimental design
The setup is simple, and that is what makes it useful. Adult participants worked through a series of fraction problems, such as 5/6 minus 1/3. They were split into two groups:
The AI group had a ChatGPT window open beside the problems for the first stretch of the test.
The control group worked unaided the whole time.
Partway through, ChatGPT vanished, and the AI group finished the remaining problems on their own. The researchers tracked two measures: the solve rate (how often people got problems right) and the 2 (how often they gave up on a problem and moved on). The skip rate is the more interesting one, because it is a rough behavioural measure of persistence rather than of skill.
What happened
While they had the chatbot, the AI group did better on both measures. They solved more problems and skipped fewer. Once the chatbot was removed, the two groups traded places. The group that had worked alone throughout now solved more and skipped less, and the formerly assisted group's solve rate fell while its skip rate rose.
To check that this wasn't specific to arithmetic, the team ran the same design with reading-comprehension questions taken from SAT exams. The result was the same: accuracy and persistence in the AI group fell after the tool disappeared.
"You almost couldn't ask for a clearer story," Christian said. "We've built these systems to be helpful, but are they being helpful to us in a way that's actually helping us? We showed, ironically, that they're often helping us in ways that are kind of unhelpful."
The "3" argument
The authors go well beyond a single session. They write that "if sustained AI use erodes the motivation and persistence that drive long-term learning, these effects will accumulate over years, and by the time they are visible, they will be difficult to reverse." The image is the boiling frog: a decline too gradual to notice until it is entrenched. Note the "if". This is an extrapolation from a short lab task, not something the study measured.
The Nautilus piece pushes the worry further, to 4 (software that carries out multi-step tasks on a user's behalf, such as Meta's Muse) that could remove friction from everyday life, not just from homework. That is the writer's speculation, not a finding.
What the article doesn't tell us
The write-up is short and leaves out much of what you'd want before weighing the result:
No numbers. It gives no sample sizes, effect sizes or statistical tests. We know the direction of the effect, not its size.
Short-term only. The tests lasted minutes. Whether the effect lasts, builds up or fades by the next day is unknown.
Adults, not children. The New York policy concerns young children, and the study doesn't test them.
Other explanations worth considering
A sharp drop in persistence fits the authors' reading: the AI eroded motivation. Other readings also fit the description, and the article doesn't say whether the researchers ruled them out:
Less practice. The control group spent the first block working fraction problems by hand, so it was warmed up when the second block began. The AI group may simply have practised less. That would be lost skill rather than lost grit, though more skipping does suggest a motivational part too.
A contrast effect. Losing a tool you've just been using can feel frustrating in a way that never having had it does not. An analogy: a cyclist whose e-bike battery dies mid-ride probably struggles more on the next hill than someone who set out on a plain bike. That reflects the shift, not a lasting change in the rider.
Recalibrated expectations. After a stretch where answers came easily, an ordinary hard problem may feel unreasonably hard, and skipping it feels justified.
Each of these points to a different fix. If the problem is practice, AI that coaches rather than answers might avoid it. If it is contrast, the issue is depending on the tool, not using it.
Why it's still worth attention
Even with those caveats, the finding is notable for its speed and its consistency across two quite different tasks. Most debate about AI and learning concerns slow effects over semesters. This study suggests the shift in behaviour can start within a single sitting. The open question is whether those minutes add up.
Remember
The groups switched places: AI users did better with the chatbot but did worse and quit more often once it was gone.
The same pattern appeared in fraction arithmetic and SAT reading questions, so it isn't specific to math.
The article reports no effect sizes. Less practice or post-removal frustration could explain the result as well as lost motivation.
Researchers at York University in Toronto have found a gap between human and machine sight. People and monkeys fall for a well-known illusion. The AI vision systems they tested do not. Senior author Kohitij Kar, a neuroscientist, argues that AI meant to see like us may need to make our mistakes.
A useful mistake
The illusion is the 1. After you watch something move steadily in one direction, a still object can seem shifted or drifting the opposite way. It is a bit like stepping off a rocking boat and feeling the ground sway.
The article suggests the illusion comes from efficiency. The brain starts to expect continued motion and overshoots when it has to correct. The standard explanation, which the article doesn't spell out, is that neurons tuned to one direction of motion get tired from long use. When the motion stops, their rested opposite-direction neighbors briefly win out. A system with that kind of adjustment over time should show the illusion. A system that judges each image on its own should not.
The test
The team showed 79 humans, two 2 monkeys and nine 3 a series of moving and stationary images. The macaques responded much as humans did. The AI networks did not, "even when nudged in the right direction."
The article doesn't name the networks or explain how they were nudged. The effect was also very small.
What it does and doesn't show
Kar's team suggests ways to train the effect into AI, but they haven't tested any of them. They also haven't shown that adding the quirk would help a network perform better. So far the finding is a difference between primates and machines, not a fix.
Kar argues that getting the right answer isn't enough. To build AI that understands the world in "human-compatible ways," researchers need to understand the computations behind human perception, and neuroscience can uncover them.
Uses, and a warning
The possible uses include:
AI that stands in for the human visual system in neuroscience experiments
systems that predict what people will actually see, for medical image displays, driver-assistance screens and 4
The article also flags a risk. A model that knows exactly where human perception fails won't necessarily share human values, and it could be used to deceive people more effectively.
A team at York University in Toronto has found a gap between human and machine sight. People and monkeys fall for a well-known motion illusion. The AI vision systems the team tested do not, even when the researchers nudged them toward it. The senior author, neuroscientist Kohitij Kar, argues that if we want AI to perceive the world the way we do, it may need to make our mistakes too.
Imperfect in different ways
Human vision is full of quirks. We have a blind spot where the optic nerve leaves the eye. We see after-images, and illusions fool us. The Nautilus piece suggests that some of these quirks may be side effects of an efficient design rather than plain defects.
AI vision systems are imperfect in a different way. They are good at locating and mapping objects in physical space. What they don't do is distort that picture the way human experience does. So a network can match human accuracy on a task and still reach its answers by a different route. Kar puts the question this way: will more capable systems become "more like us or increasingly different from us"?
The illusion
The team picked the 1. After you watch something move steadily in one direction, a still object can seem shifted or drifting the opposite way. The article compares it to stepping off a rocking boat and feeling the dock sway for a while.
The article's explanation is about efficiency. The brain starts to expect the ongoing motion, then has to correct when the motion stops, and it overshoots. The standard textbook account, which the article doesn't spell out, is similar. Neurons tuned to one direction of motion fire hard during long viewing and then adapt, turning down their response. When the scene stops, the tired "this way" neurons fire less than their rested "that way" counterparts. The imbalance reads as motion in the opposite direction.
That makes the illusion a useful probe. Any system that has the effect probably adjusts to recent input over time, the way neurons do. A system that judges each frame on its own, with no memory of what came before, should be immune to it.
The experiment
The researchers tested 79 humans, two 2 monkeys and nine 3 on a series of moving and stationary images. The results suggested that the macaques respond to the aftereffect much as humans do. The AI networks did not, "even when nudged in the right direction."
The article leaves out a lot. It doesn't name the nine networks or describe their architectures. It doesn't say how they were nudged or how the monkeys' responses were measured. It also doesn't name the journal. Readers wanting those details will need the paper itself.
How small, and what it proves
The measured effect was very small, which is worth keeping in mind. Kar's team proposes specific ways to build the effect into AI through training and feedback. But those proposals were not tested, so they remain hypothetical. The researchers also did not show that giving a network this quirk would make it perform better at anything.
So the result is a difference, not a fix. Primates share a perceptual signature that current AI networks lack. Whether that gap matters for practical systems is still open.
Why make AI see wrongly?
Kar's argument is that getting the right answer is not enough. "We also need to understand the computations that produce human perception and behavior," he said. In his view, neuroscience can uncover those computations and possibly build them into AI.
The article names several possible payoffs:
Better models for neuroscience. A network that has human illusions could stand in for the visual system in experiments.
Predicting what people will see. A system that knows how a display will look to a human, distortions included, could improve medical image displays, driver-assistance screens and 4, where virtual images are laid over the real world.
The other edge
The article also notes a risk. A model that can predict human perceptual errors won't necessarily share human values or be easier to control. It could just as easily help design visuals that deceive or manipulate people, and the article observes that AI is already fairly good at that. Knowing exactly where human perception slips is useful to the designer of a safer dashboard and to a manipulator alike.
What's left open
The study leaves open how much human likeness is worth pursuing in machine vision, and of what kind. The motion aftereffect is one narrow test with a small effect. The untested proposals for building it into AI are the obvious next step. The deeper question is whether copying our quirks leads to AI that understands us better, or only to AI that exploits us better.
Remember
A shared illusion can show a shared mechanism. Matching our mistakes is a stricter test of 'human-like' than matching our answers.
The effect was small, and the proposed fixes for AI haven't been tested.
AI that can predict human perceptual errors could make displays safer, or make deception more effective.
