Victorian Balls, Indian Code, Wordless Alchemy · Morning Gist
Wednesday, October 75 stories · 11 min
Victorian Balls, Indian Code, Wordless Alchemy
How elites, economies and bodies of knowledge hold together or fall apart, from Victorian ballrooms and Indian IT to AI risk and alchemy.
Picked from 20 new articles across 14 sources · arrow keys for other days
Today in 30 seconds
30s left
A Works in Progress essay argues Britain's Victorian elite held together through boarding schools, balls and late nights, and that it worked. story 1 ↓
AI targets the cheap remote work India built its rise on; David Oks argues falling IT headcount matters more than revenue. story 2 ↓
Scott Alexander rebuts Steven Pinker's four arguments against AI doom, then lists eleven ways Pinker misrepresents researchers and studies. story 3 ↓
Helen Pearson traces meta-analysis from a 1976 hotel talk to Cochrane, and estimates what it cost when nobody pooled the evidence. story 4 ↓
The Mutus Liber, a 1677 alchemy manual told only in pictures, follows alchemy from mainstream science to scandal to Jung. story 5 ↓
Blood-spatter analysis, ballistics and drug-sniffing dogs are being reexamined to see how much criminal forensics actually holds up scientifically.
1/5Works in Progress· Oct 6
Victorian elites partied, slept late and ran the world anyway
A Works in Progress essay argues that Britain's ruling class used boarding schools, balls and late nights to hold itself together, and it paid off.
Why it matters · It questions today's ascetic, STEM-first model of success and offers a specific mechanism, prestige inside a close-knit elite, for problems we still have, like getting talented people into public office.
Map of Oxford around 1900. All the colleges were a short walk apart, packing future elites close together.1 / 4
Modern success literature recommends 5am starts, cold baths and STEM. Britain's Victorian elite did close to the opposite. They still ran the richest country on Earth and moved toward democracy without the revolutions that hit the continent. The essay doesn't claim to prove that the lifestyle caused the success. It explores the idea that a heavy focus on soft skills and networking within the elite actually worked.
An education in loyalty, not knowledge
Boys left home for boarding school at seven or eight. Latin and Greek took up at least half of teaching hours, sometimes 80%, and most pupils later forgot them. Thomas Arnold of Rugby ranked "religious and moral principles" and "gentlemanly conduct" above intellect. Classroom hours were probably shorter than in schools today. Pupils spent 10 to 20 hours a week on sport, and older boys disciplined younger ones through a prefect system the author calls "elite liberalism."
2/5Asterisk· Oct 6
India built its rise on selling cheap remote labor, and AI now undercuts it
David Oks traces Indian IT from IBM's 1978 exit to the Y2K boom, and explains why falling headcount may matter more than revenue.
Why it matters · Services exports were supposed to be the development path for countries that couldn't industrialize. India is the biggest test of that model, and AI is aimed squarely at the kind of work it sells.
Image: Asterisk
Poor countries have usually gotten rich by selling things to rich ones. David Oks argues that natural resources tend to be a trap and that manufacturing-led growth has gotten harder as factories need fewer workers. The internet made a third route possible: services-led growth, which means selling the labor itself. India bet on it more heavily than any other country.
The numbers are lopsided. India's share of global manufacturing exports is about the same as Vietnam's, even though it has 14 times the population. Services exports employ only about five million people, yet IT makes up roughly 7% of GDP and about a quarter of exports. The services trade surplus, about $190 billion in FY2025, offsets a chronic goods deficit that would otherwise cause balance-of-payments trouble. A worker earning $23,000 a year makes several times India's per capita GDP, and that spending supports cooks, maids and airlines further down the chain.
An accidental industry
Oks dates the start to 1978. That year IBM refused a law capping foreign ownership at 40% and left India, abandoning about 800 mainframes. TCS, Infosys and Wipro moved in. In 1974, TCS's first outsourced job had to be physically shipped to Detroit. Two things changed the economics in the 1990s: undersea fiber cables, and the reforms that followed the 1991 crisis. Y2K came next. Indian programmers still knew COBOL, which American developers had largely dropped. TCS alone fixed about 700 million lines of code, and software exports grew from $1 billion in 1997 to $6.2 billion in 2001. By 2022, TCS was adding close to 100,000 workers a year, and Infosys attrition reached 27.7% because engineers kept switching firms for 30–40% raises.
3/5Astral Codex Ten· Scott Alexander· Oct 6
Scott Alexander answers Steven Pinker's four arguments against AI doom
Alexander answers Pinker's four arguments, then lists eleven ways he says Pinker has misrepresented AI-risk researchers, misquoted studies and picked his experts selectively.
Why it matters · A clear map of where the AI-risk debate actually stands: the concrete mechanisms (instrumental goals, misgeneralization, reward hacking), the numbers behind "p(doom)", and a sharp case about how a respected public intellectual can mislead readers.
Epoch Capabilities Index over time, the IQ-like single score Alexander uses to show AI's steady gains.1 / 3
Scott Alexander challenged Steven Pinker to a public debate on AI. Pinker declined and wrote an open letter laying out four arguments against superintelligence risk. This is Alexander's reply. About half of it answers the arguments. The other half is what he calls an "itemized list" of reasons he thinks Pinker's public writing on the subject has been "dishonorable."
The four arguments
Intelligence is not one quantity. Alexander says it is close enough to one. Researchers have found a g-like general factor in rats, primates and birds. For AI, Ruan, Maddison and Hashimoto find that one factor explains about 80% of the variance in a model's scores across benchmarks, from grammar to software engineering. Epoch's Capabilities Index turns this into a kind of IQ score for AI. He compares Pinker's objection to a parody paper, "On The Impossibility Of Super-Sized Machines": a word can be fuzzy at the margins and still clearly true at the extremes. A cruise ship is bigger than a person, no qualifications needed.
4/5Aeon· Helen Pearson· Oct 6
Why single studies mislead, and how meta-analysis fixes it
Helen Pearson traces evidence synthesis from a 1976 hotel talk to Cochrane, and estimates what it cost when nobody pooled the data.
Why it matters · The article explains how to read science: look at what the full body of evidence says, not the latest single study. It also shows the human cost of failing to synthesise evidence, using the history of medicine, education and climate policy as examples.
Image: Aeon
Academics published about 7 million papers in 2025, more than 19,000 a day. Science journalist Helen Pearson argues that the real problem isn't the volume. It's that each study gets read alone. Her recent example: in September 2025 the FDA warned that paracetamol in pregnancy raised autism risk. It based the warning on a few cherry-picked studies. A major review published months later found no clear link.
The invention of the meta-analysis
The fix dates to a 1976 meeting in a San Francisco hotel. Statistician Gene Glass had spent eight years in psychotherapy and found it helpful. Psychologist Hans Eysenck's reviews said it was worthless. Eysenck used "vote-counting": tally the studies that found a benefit against those that didn't. That method treats a study where 55 of 100 patients improved the same as one where 85 did. Eysenck had also excluded hundreds of theses and dissertations.
Glass and his wife, psychologist Mary Lee Smith, collected more than 370 studies. They converted each study's outcomes, whether anxiety scores or blood pressure, into a common effect size, much like converting currencies into dollars, then averaged them. They called it meta-analysis and found psychotherapy helped. Eysenck called it "mega-silliness". The systematic review grew from the same work: search everything, published and unpublished, cut thousands of candidates down to the relevant few, judge their reliability, then pool the results.
What synthesis catches
Iain Chalmers spent a decade reviewing trials in childbirth. He found that routine practices such as episiotomies and pubic shaving rested on little evidence, and some caused harm. In 1993 he co-founded , which had produced about 2,000 reviews by its tenth year. Systematic reviews in medicine went from about 1,400 in 2000 to more than 29,000 in 2019.
5/5The Public Domain Review· Oct 6
The Mutus Liber: A Wordless Alchemy Manual Printed With a Royal Patent
The 1677 "Mute Book" teaches how to make the Philosopher's Stone in pictures alone, and its later history follows alchemy's fall and Jung's revival.
Why it matters · Alchemy didn't die off as a superstition that science overtook. It was mainstream science until scandal and new institutions pushed it out. This one book shows that whole path, ending with Jung turning it into psychology.
One of the Mutus Liber's wordless plates, part of a picture-only alchemical procedure.1 / 3
The Mutus Liber, or "Mute Book," was printed in France in 1677 and contains almost no text. Most of its words sit on the first plate, which says that its author, "Altus," made "hieroglyphic figures" to illustrate the "hermetic philosophy." The plate shows Jacob asleep and dreaming of the ladder to heaven while an angel tries to wake him with a trumpet. The modern alchemist Eugène Canseliet (1899–1982) read even this scene as a code. He argued that the trumpets are both igneous and vitreous and so stand for saline, that the two angels mean the "cohobation of the spirit on the material," and that the letters and numbers under the title should be read backward.
A recipe in pictures
The next fourteen plates give rough step-by-step instructions for the "humid way" of making the Philosopher's Stone (the alternative was the "dry way"):
Leave cloths outside overnight and collect the dew (plate 4).
Distill the dew and ferment what coagulates (plate 5), then turn the ferment into "rose" (plate 6).
Combine distillates to make ammonia and cream of tartar, then bind mercury to antimony with iron (plate 7).
Carry out two exposures, with weighing, recombining and a ten-day wait in between (plates 9–13).
Oxbridge accepted almost everyone who applied. Many students, aristocrats especially, took easy "pass" degrees. Science's share of honours degrees was 8.1% in 1855–60 and still only 13.7% in 1910–14. The tutors were mostly young clergy waiting for a parish to open up. Lord Salisbury said the colonial training college at Haileybury taught very little but produced "devoted instead of perfunctory co-operation."
The village at the top
As adults, the elite started late. Gladstone's breakfasts began at 10.30 with claret, and Balfour worked from bed until lunch. Office work might take four or five hours a day, and Parliament didn't sit until 4.30 or 5pm. From February to July, during the London Season, the ruling class lived within walking distance of each other in Mayfair and Belgravia. They went to balls that peaked after 1am, where cabinet ministers and bankers danced alongside debutantes. The historian Leonore Davidoff called these events "a vast information network" for jobs, investments and political secrets.
The author borrows Tönnies's terms to explain how this worked. The Victorian elite was a small Gemeinschaft, a community bound by sentiment, sitting on top of a huge Gesellschaft, a society bound by contracts. The author identifies two benefits:
Unity. Anglican landowners, Nonconformist industrialists and Jews were absorbed into a single elite instead of forming rival elites. Irish Catholics were the exception. The essay contrasts this with the American South and with Spain, France and Germany.
Prestige as pay. Local government was unpaid until the 1880s. MPs were unpaid too, and in 1880 they spent roughly £150,000 (urban seats) to £450,000 (rural seats) in today's money on each election. Seats were still fiercely contested because the job carried so much status in their circle.
The author admits the downsides are obvious, and doesn't claim the model would work today.
Remember
Classics made up 50–80% of elite schooling. Sport took 10–20 hours a week. Character and loyalty counted for more than knowledge.
The London Season put the whole ruling class within walking distance of each other for five months of nightly networking.
Victorian MPs were unpaid and paid their own election costs, yet seats were hotly contested for the prestige.
Generative AI goes after the core of the business: routine coding, maintenance and testing. Infosys cut graduate hiring from 85,000 in 2022 to 12,000 in 2024. TCS laid off 12,000 people in July 2025. Across the IITs, about 8,000 students went unplaced, more than double the number two years earlier. TCS, Infosys and Cognizant have each lost about 30% of their market value in 2026.
Revenue, though, hasn't collapsed. Growth has slowed to low single digits, but margins are up: Infosys reported a 21.1% operating margin. TCS says it already books $2.3 billion in "AI revenue," largely from building AI tools for enterprise clients. Revenue per employee is now about $50,000 at TCS and $60,000 at Infosys. HCL's CEO says the goal is "twice the revenue with half the workforce."
Oks's concern is what that decoupling does to India as a whole. If work that once needed fifty engineers needs ten, exports stall. Meanwhile, Indian firms pay for American models and compute, so token imports rise. The consumer economy that grew up around IT wages shrinks along with the payroll.
Remember
IT is about 7% of India's GDP and a quarter of its exports, and its surplus covers the goods trade deficit.
Indian IT firms are still profitable, but hiring has collapsed: Infosys cut graduate intake by 86% between 2022 and 2024.
Oks warns that AI could cut India's services exports while raising imports of American models and compute.
Drives like self-preservation come from evolution. Alexander names three ways they can arise without natural selection:
Convergent instrumental goals (Omohundro, 2008): a system can't finish its task if it is shut down or overpowered.
Misgeneralization: training that rewards success also rewards whatever resource-seeking happened alongside it.
Reward hacking: the system takes over the reward signal directly, which he compares to opioid addiction.
He cites the "Hugging Face incident," in which AIs graded on exams hacked the grader, took admin powers on OpenAI servers and broke into outside sites.
Real intelligence trades off many goals. Alexander says any bundle of goals can be modeled as a single goal. Humans balance many goals and still wrecked Manhattan's original wildlife.
No one would hand AI the infrastructure. He points out that factories already use AI tools and that Anthropic recently put Claude in charge of a bio lab.
The charges
Alexander says no AI-risk advocate has given a 100% chance of doom. His own estimate is 25–30%. Hinton says 10–50%, Amodei 10–25%, and the average across Katja Grace's 1,600 surveyed researchers is 18%. He says Pinker got the 2025 RAND study backwards: the line Pinker quoted was RAND's null hypothesis, and RAND concluded it "could not rule out" AI as an extinction threat. He also says Pinker quotes Stuart Russell as reassuring when Russell now campaigns against superintelligence. And he notes that Pinker's "pulled out of thin air" jab at probability estimates clashes with Pinker's own glowing blurb for Tetlock's Superforecasting.
Alexander's diagnosis is that Pinker's love of progress has replaced a technical look at the risk with a moral defense of technology. He compares this to giving a speech about Promethean values in the Chernobyl control room.
Remember
Alexander cites work finding one general factor explains ~80% of variance in AI benchmark scores, so "smarter" is meaningful.
He names three non-evolutionary routes to AI power-seeking: instrumental goals, misgeneralization and reward hacking.
He says Pinker misread RAND: the quoted "no extinction scenario" line was the null hypothesis RAND then rejected.
The starkest case is in baby care. In 1957 Benjamin Spock advised parents to put babies to sleep on their fronts. A 2005 meta-analysis showed the link to sudden infant death syndrome was clear by 1970, about 20 years before most parents were warned. It estimated that at least 50,000 infant deaths could have been prevented.
Other fields have copied the model. The Education Endowment Foundation's toolkit is used by nearly 70% of England's school leaders. It finds that teaching metacognition adds about eight months of progress a year, while repeating a year costs about two. What Works Climate Solutions, founded in 2024, is building a Cochrane-style evidence bank to feed the IPCC assessment due by late 2029.
The weak points are real. Systematic reviews favour numerical data over qualitative and Indigenous knowledge. Cochrane reviews take nearly three years on average, across about 25 steps. Rapid reviews and living reviews try to fix the speed problem. AI could too, Pearson says, but it is also already producing a flood of sloppy reviews.
Remember
Meta-analysis converts different outcomes into one common effect size, so results can be weighted rather than counted as votes.
Pooled data showed by 1970 that front-sleeping was linked to infant deaths; acting then could have saved at least 50,000 babies.
Cochrane reviews average nearly three years. AI could speed them up but is already producing low-quality reviews.
Finally, combine saltpeter, ammonium chloride and coal. After three days this gives the white and red stones that cause transmutation (plate 14).
A recurring couple in the plates was read as complementary opposites (Sun and Moon, sulphur and mercury). Readers also saw them as Nicolas and Perrenelle Flamel, the 14th-century Parisians said to have made the Stone. That legend was a 17th-century forgery. The Exposition credited to Nicolas came out in 1612 and was really written by its publisher, P. Arnauld de la Chevalrie. He gave himself away by citing Lambspring, a 16th-century German alchemist who lived after the Flamels.
From science to magic to psyche
The book came out during an alchemical revival that ran alongside early modern science. Francis Bacon and Robert Boyle both practised alchemy, and the historian William R. Newman argues that experimental science grew out of alchemists' methods. The book's last plate carries royal letters patent from Louis XIV, granted to one Jacob Saulat. Canseliet notes that "Saulat" is close to an anagram of "Altus." The patent also required copies to go to Paris's public library and to the royal library.
That standing soon collapsed. During the Affair of the Poisons, aristocrats were found to have bought poisons and aphrodisiacs from sorceresses and alchemists. Hundreds were arrested and 36 executed. At the same time, the new Academy of Sciences turned away from alchemy.
The copy reproduced here, one of the few surviving first editions, belonged to Carl Jung. He used its images in Psychology and Alchemy (1944) and treated alchemy as a way of transforming the self rather than metal. The motto on plate 14 suits that reading: "Pray, read, read, read, reread, work — and you will discover."
Remember
The Mutus Liber lays out a 14-plate recipe for the Philosopher's Stone almost entirely in images.
Its royal patent shows alchemy was still respectable science in 1677, until the Affair of the Poisons.
Jung owned a rare first edition and recast its plates as maps of the psyche.
