Palladium lecture, Thu 24 Sep 2026 (week 39) · the data behind it
Show me the data
One hundred years of techno-apocalypse, checked claim by claim
Ben Landau-Taylor’s talk argued that since World War I there has been more demand for a technological apocalypse than supply. Bombers, the cobalt bomb, nuclear winter, the population bomb, peak oil and global warming each had their turn, and AI is next. Walking out, Daria’s verdict was that the story was good but the evidence was thin. Here is the evidence, checked claim by claim, including where it pushes back.
The scorecard
✅ holds · ⚖️ half right or needs a caveat · ❌ doesn’t hold
| The lecture’s claim | Verdict | The number |
|---|---|---|
| Apocalypse panics rise, peak and fade within about 15–20 years | ✅ holds | Median 13 years above half-peak across nine past panics (§1) |
| Before 1914, new technology inspired hope, not doom | ⚖️ mostly | Franklin hoped balloons would end war, but in 1784 he and Ingenhousz had already sketched the whole bombing scenario (§2) |
| The superweapons did little of WWII’s killing | ✅ holds | All strategic bombing ≈ 1.5% of the dead; the atomic bombs ≈ 1 in 250–700 (§3) |
| WWII killed about the same share as the Thirty Years’ War | ⚖️ understated | Where each was fought, the pike-and-plague war was ~3× deadlier: 20% of the Empire vs 6–8% of Europe (§3) |
| Nuclear doomsday mechanisms faded without being tested | ✅ holds | Cobalt bomb: never built. Doomsday Clock “closest ever” with 83% fewer warheads than 1986 (§4) |
| Nuclear winter failed its real-world tests | ⚖️ mixed | Kuwait 1991 supports it. But the 2017 wildfire smoke confirmed self-lofting, so the essay has that example backwards (§4) |
| The Population Bomb’s famines never came | ✅ holds | 1970s famine deaths 3.8 M, not “hundreds of millions”; calories per person +38% while population +164% (§5) |
| Limits to Growth’s depletion dates failed, even at 5× reserves | ✅ holds | The 5× column is real. We have burned 3.3× the 1970 oil reserves and book more today, though half of that is accounting (§5) |
| We dealt with pollution | ⚖️ yes, by regulation | Emissions −79% while GDP +338%. This one faded because the alarm was right and was acted on (§6) |
| The climate apocalypse is fading without being refuted | ⚖️ half right | Coverage −57% since 2021 and the worst case quietly retired, but +1.55 °C is real (§7) |
| AI timelines and p(doom) numbers are “a vibe” | ⚖️ half right | Experts and superforecasters differ 8× and never converge, but both underestimated AI progress (§8) |
| Nobody ever admits an apocalypse is cancelled | ⚖️ mostly | Partly: Nature (2020) retired the worst climate case, and Gates (2025) wrote that climate change “will not lead to humanity’s demise”. Ehrlich instead added a decade to his deadline (§5, §7) |
Overall: the pattern Ben describes is real and measurable. In several places (the Thirty Years’ War, the bombing share, peak oil) his case is stronger than he made it. In one (the 2017 wildfire smoke) the essay gets the evidence backwards. The honest counterweight is §6 and §9: some scares faded because people acted, the skeptics’ forecasts failed too, and AI progress so far has been faster than both camps predicted.
- 1 · The lifecycle, measured
- 2 · Before 1914
- 3 · Bombers and the world wars
- 4 · The bomb
- 5 · Population and resources
- 6 · Pollution
- 7 · Climate
- 8 · AI
- 9 · Where the data pushes back
- 10 · Four friendly questions
1The lifecycle, measured
Claim: each apocalypse rises, peaks, and fades within about 15 to 20 years, and nobody ever formally cancels it.
This is the lecture's central claim, and it can be checked. Google Books has counted how often every phrase appears in English-language books each year since 1500. Here are twelve doom phrases, each scaled to its own peak so the shapes can be compared.
Twelve doom phrases in English books, 1900–2022, each scaled to its own peak
Share of all two-word phrases, 7-year moving average. Tap or hover a panel for any year.
What this supports: the rise-and-fade shape is real, it repeats, and it is fast. What it complicates: “global warming” peaked in 2009, but its successor phrases (“climate change”, “climate crisis”, “climate emergency”) were all still at record highs in 2022. Climate worry changed its name rather than fading, at least in books. The fade the lecture describes would have to be after 2022, which this corpus can’t see yet (see §7).
Table view: peak year and time above half-peak
| Phrase | Peak | Above ½ peak | Span | 2022 vs peak |
|---|---|---|---|---|
| “next war” | 1936 | 1918–1950 | 33 yr | 16% |
| “fallout shelter” | 1964 | 1959–1968 | 10 yr | 7% |
| “population explosion” | 1968 | 1961–1976 | 16 yr | 19% |
| “energy crisis” | 1976 | 1971–1979 | 9 yr | 6% |
| “acid rain” | 1987 | 1981–1994 | 14 yr | 18% |
| “nuclear winter” | 1987 | 1983–1990 | 8 yr | 20% |
| “ozone hole” | 1991 | 1987–2005 | 19 yr | 30% |
| “Y2K” | 2000 | 1996–2002 | 7 yr | 3% |
| “peak oil” | 2011 | 2006–2018 | 13 yr | 34% |
| “global warming” | 2009 | 2000–2022+ | ≥23 yr | 80% |
| “existential risk” | 2022 | 2017–2022+ | ≥6 yr | 100% |
| “superintelligence” | 2021 | 2016–2022+ | ≥7 yr | 97% |
2Before 1914: optimism, and a very calm coal crisis
Claim: before World War I, new technology inspired hope, not doom. Even a real resource crisis caused no panic.
Franklin’s balloons (1784)
Four months after the first crewed balloon flight, Benjamin Franklin wrote to the Dutch scientist Jan Ingenhousz that balloons might end war itself:
“Convincing Sovereigns of the Folly of Wars, may perhaps be one Effect of it.”Franklin to Ingenhousz, 16 Jan 1784 · Founders Online (mirror)
That part checks out, but the full letter is darker than the lecture’s version. Franklin pictures “Ten Thousand Men descending from the Clouds” that no defence could stop, and he opens with “as you observe”, so he is agreeing with Ingenhousz. Two weeks earlier Ingenhousz had described fleets that “throw fire and destruction” from the sky (his letter, 2 Jan 1784). The whole 1930s air-power argument, cities bombed from the air and defence impossible, was already on paper in 1784. The difference is the conclusion: they thought it would end war.
Jevons and The Coal Question (1865)
William Stanley Jevons argued that Britain would run out of cheap coal and lose its industrial lead (full text). Educated readers took it seriously, but nobody organised around it. Here is what happened to the coal:
Verdict: holds. Jevons was right that the coal was finite. He was wrong that it would run out, because nobody in 1865 could count oil, gas, nuclear or solar as reserves. That mistake comes back in §5.
3Bombers, and the war that did come
Claim: after 1918, people believed air power would destroy civilisation. The war came and was horrific, but civilisation did not end, and superweapons caused little of the killing.
The lecture’s key scene is a real House of Commons debate, 10 Nov 1932 (Hansard, vol 270). Both front benches said the same thing:
“Unless air warfare is restrained, civilisation will be wiped out.”Clement Attlee (Labour), col. 534
“The bomber will always get through.”Stanley Baldwin (Lord President of the Council), col. 632
Baldwin ended the debate with “when the next war comes, and European civilisation is wiped out”, and in the same speech admitted “I am not a mathematician” after muddling his own arithmetic. Attlee’s remedy was to hand civil aviation to an international body, which is the same move as today’s proposals for an international AI agency.
Was WWII really no worse than a war fought with pikes?
| War | Deaths | Share of pre-war population | Weapons |
|---|---|---|---|
| Thirty Years’ War, 1618–48 | 5–8 million | ~20% of the Holy Roman Empire (~5–8% of all Europe) | pikes, muskets, plague, hunger |
| World War II in Europe, 1939–45 | 34–45 million | ~6–8% of Europe | tanks, bombers, industrial genocide |
Peter H. Wilson, The Thirty Years War: Europe’s Tragedy (2009), Table 7 (“Comparative total deaths in major conflicts”: 5 M = 20% of the Empire; WWII 33.8 M = 6.0% of Europe) and his upper estimate of 8 M. The upper WWII figure is my sum of the European rows in the National WWII Museum table. Europe’s population was ~102 M in 1600 and ~524 M in 1939 (OWID/HYDE). The “all Europe” share is my own division.
The lecture said the two wars killed about the same share, roughly 7%. The better comparison makes his case stronger: measured where each war was fought, the pike-and-plague war was about three times as deadly. Wilson writes that even a 15% loss would make it “the most destructive conflict in European history”, while the USSR, WWII’s worst-hit country, lost under 12%. The weapons improved enormously in between, and the death rate went down. What went up was the number of people: world population grew from 0.52 billion in 1600 to 2.27 billion in 1939 (OWID). The essay’s line holds: WWII killed more people than any other war “because industrial technology can support an enormous population.”
How much of WWII’s killing did the “superweapons” do?
Deaths, millions. Bars are midpoints of published ranges; hover or tap for each range. Orange = the new air weapons.
Every strategic bombing campaign of the war, on every side and atomic bombs included, killed roughly 1 million people, about 1.5% of the war’s dead (range 1–2.5%). The two atomic bombs account for about 1 in every 250–700 WWII deaths. The essay’s estimate, “probably less than ten percent” from the new weapons, is if anything generous. The US bombing survey said the same thing about its own work in 1946: compared with Hiroshima, the Tokyo fire raid of March 1945 killed “no larger” a number (USSBS, 19 Jun 1946). One caveat keeps this honest. The killing that did dominate wasn’t low-tech either. The Holocaust ran on railways and census bureaucracy, and the famines were largely policy. So the real contrast is between spectacular killing and administrative killing, not between old and new technology.
The talk’s other comparison was the Mongols’ sack of Baghdad in 1258, done with swords and bows. The chroniclers’ counts run from “tens of thousands” (Bar Hebraeus) to 800,000 (Mustawfi) to 2 million (later writers). Modern historians note the whole city probably held no more than ~500,000 people (Konieczny), and that much of the dying was epidemic disease afterwards (Fancy & Green, Medical History 2021). So the numbers themselves show the exaggeration reflex the lecture describes. The same thing happened in 1945. The US bombing survey put Hamburg’s 1943 firestorm at “60,000 to 100,000 people killed” (USSBS). Today’s accepted figure is ~37,000 (Overy, The Bombing War, 2013; summary).
What still stands is his point that no new physics was needed. The Tokyo raid of 9–10 March 1945 used incendiaries and killed ~84,000–97,000 in one night (sources). That is more than Nagasaki and close to Hiroshima’s same-day toll, and it is still the deadliest air raid in history.
4The bomb: three apocalypses, zero detonations
Claim: once the bomb existed, people kept looking for a way it could kill everyone. Each mechanism faded, and the next one replaced it.
Version 1: the cobalt bomb (1950)
On 26 Feb 1950, on NBC’s University of Chicago Round Table, Leo Szilard worked out on air how much “salted” H-bomb it would take to dust the whole planet with long-lived fallout: “about fifty tons of neutrons as being plenty to kill everybody.” He only named cobalt later, in a footnote to the April 1950 Bulletin reprint, and he meant the idea as a warning about deterrence, not as a design (Alex Wellerstein’s history). No cobalt bomb has ever been built or tested. The only cobalt in a test device was a 1957 British tracer (Nuclear Weapon Archive FAQ §1.6). It lived on in On the Beach and in Dr. Strangelove’s Doomsday Machine (1964). The weapon was never built, but the dread was.
Version 2: nuclear winter (1983)
The 1983 “TTAPS” paper in Science (Turco, Toon, Ackerman, Pollack, Sagan) modelled soot from burning cities blocking the sun. It became a public campaign within weeks. How it held up is in the next section.
A companion paper in the same issue of Science said “extinction of the human species itself cannot be excluded” (Ehrlich et al. 1983). The results were announced at a public conference seven weeks before peer review, and Sagan argued in Foreign Affairs for cutting arsenals to “roughly around 500 to 2,000 warheads” (Sagan 1983). So the essay’s description of activist-scientists holds. Nuclear winter has had two real-world tests, and they point in opposite directions.
Test 1, Kuwait 1991: supports the lecture. On Nightline Sagan compared the oil-well fires to the 1815 Tambora eruption and its “year without a summer” (Baltimore Sun). In fact the smoke was “not carried high in the atmosphere” and had a short lifetime. It cooled the Gulf region by 4–6 °C and had no global effect (Hobbs & Radke, Science 1992).
Test 2, the 2017 and 2019–20 wildfires: the essay has this one backwards. It says the 2017 Canadian smoke “dissipated much more quickly than predicted.” The study that line appears to rest on found the smoke rose on its own from 12 to 23 km and stayed visible for more than 8 months. Its authors concluded that self-lofting “will occur… with an order-of-magnitude less [black carbon] than typical for urban fires”, which is the nuclear-winter mechanism, confirmed (Yu et al., Science 2019). The “shorter-lived” figure compares the smoke with a model that deliberately left one chemical process out. Australia’s 2019–20 smoke reached 35 km and took 19–20 months to decay (Khaykin et al. 2020).
The lecture does have some support. A Los Alamos study found “none of the simulations produced a nuclear winter effect” (Reisner et al. 2018). Critics reply that it burned a suburban Atlanta golf course with 15–110× less fuel than real city targets. A 2024 Johns Hopkins review concludes the problem “is not resolved within the scientific community” (APL). Verdict: mixed. The atmospheric physics has held up. What is still disputed is how much soot a burning city actually produces.
The Doomsday Clock vs the actual number of bombs
The clock says “closest ever” while the arsenal is one-sixth of its peak
Two panels, one shared timeline. Top: the Bulletin’s setting. Bottom: warheads that physically exist.
Over 80 years, the clock and the arsenal have moved in opposite directions: the correlation between warhead count and time-to-midnight is r = +0.50, so more bombs has gone with more time to midnight. Two caveats, to be fair to the Bulletin: a warhead count is not risk (alert status, arms-control treaties and crisis stability matter more), and since 2007 the clock has also counted climate, and more recently AI and biology. Even so, the best-known image of nuclear doom doesn’t track the number of nuclear weapons.
5Population and resources: the bomb that didn’t go off
Claim: the population and resource apocalypses of the late 1960s and 1970s predicted mass famine and depletion. Instead food per person rose, and the resources kept “growing”.
“Hundreds of millions of people are going to starve to death”
Paul Ehrlich’s The Population Bomb (1968) opened with: “In the 1970’s the world will undergo famines—hundreds of millions of people are going to starve to death.” The later revised edition reads “In the 1970s and 1980s” (revised-edition scan). So the deadline was quietly extended by a decade between printings, which is the lecture’s “no big finale” in miniature.
Famine deaths per 100,000 people, by decade
Famines with at least 100,000 deaths. The prophecy’s decade is in orange. Hover or tap for totals.
The 1970s had 3.8 million famine deaths, 50–130 times fewer than predicted. Per head, only the 1880s had been lower since records begin in the 1870s. The famines that did occur were mostly made by policy: 97% of the 1960s figure is China’s Great Leap Forward, followed by Cambodia, the Derg in Ethiopia and North Korea.
More people, and more food for each of them
Both indexed to 1961 = 100, on one axis.
The honest half: 645 million people were still hungry in 2025, and Africa’s count (≈309 million) is rising. The share collapsed; the problem did not disappear.
The bet
In 1980 the economist Julian Simon bet Ehrlich that five metals (chromium, copper, nickel, tin, tungsten) would be cheaper in 1990. They were. “Ehrlich did not even bother to write a letter”; he mailed a check for $576.07 (Tierney, NYT Magazine 1990). But in fairness to Ehrlich, across every 10-year window from 1900 to 2007, he would have won 61% of the time. Simon “happened to place the bet during one of the 38.2% of years” he would win (Kiel, Matheson & Golembiewski 2010). The anecdote proves less than its fame suggests.
Oil: the hourglass that keeps refilling
The lecture said Limits to Growth (1972) generously assumed five times the known reserves. That checks out word for word: Table 4 has a column headed “Exponential Index Calculated Using 5 Times Known Reserves”. For oil it gives 20 years at known reserves (to ~1990) and 50 years at five times as much (to ~2020) (scan, pp. 56–60).
We burned 3.3× the 1970 reserves and have more left than ever
Billions of barrels
US production tells the same story. It hit its low of 5.0 million barrels/day in 2008, and in 2008 the government forecast a peak of 6.3 in 2018 (EIA AEO2008). Actual 2018 output was 11.0, and 2025 was 13.7, the highest ever (EIA). Princeton’s Kenneth Deffeyes announced in 2006 that world oil had peaked on 16 Dec 2005: “My career as a prophet is over. I’m now an historian.” (his page). World output hit a new record in 2024. Copper follows the same pattern: 280 Mt of reserves in 1970, 697 Mt mined since, and 980 Mt of reserves now. The US Geological Survey itself calls that “more than 3.5 times those in 1970” (USGS MCS 2025).
Three caveats. (1) More than half of the growth in oil reserves since 1980 is accounting rather than geology: OPEC members raised their booked reserves by 260 bn barrels in 1984–90 to win bigger quotas, and Canada and Venezuela later reclassified oil sands and heavy oil. (2) Copper ore now averages 0.62% copper, down from 1.67% around 1990, so each ton costs more energy (Calvo et al. 2016). (3) Limits to Growth itself never predicted a collapse by 2000: “all LtG scenarios show the global economic system growing at the year 2000” (Turner 2008). Table 4 was an illustration, not the model’s output. The dated collapse forecasts came later, from its defenders: Turner in 2014 put the onset at ~2015. Since then, world GDP per person rose 17% from 2015 to 2024 (World Bank).
6Pollution: solved, and that’s the problem for the thesis
Claim: the pollution apocalypse faded because we dealt with it. The rivers were cleaned up and the acid rain stopped.
In 1969 the federal pollution agency said the lower Cuyahoga River had “no visible life, not even… leeches and sludge worms.” It now holds 76 fish species. The famous Time photo of the river on fire was actually from 1952, a fire 30 times more damaging than 1969’s (Encyclopedia of Cleveland History). The ozone layer is on track to return to 1980 levels around 2040 worldwide and 2066 over Antarctica (WMO/UNEP 2022).
Verdict: holds, but it cuts against the thesis. Ben is right that these fears faded. But they faded because the alarm was substantively correct and it led to regulation. Two economists summed up the acid-rain program: “the government did what turned out to be the right thing for the wrong reason” (Schmalensee & Stavins 2013). That is the strongest counterexample to “demand exceeds supply”, and §9 comes back to it.
7Climate: the prophecies missed, the thermometer didn’t
Claim: the climate apocalypse is fading the usual way, through boredom rather than refutation, and AI is filling the gap it leaves.
The dated prophecies
| What was said | What happened |
|---|---|
| Al Gore, Copenhagen, 14 Dec 2009: a “75 per cent chance” the Arctic could be ice-free in summer “within five to seven years”. The scientist he cited disowned the figure the next day (Snopes). | The Sept 2026 minimum was 4.60 million km². “Ice-free” means under 1 million. But 2007–2026 are the 20 lowest years on record (NSIDC). The trend was real; the date was made up. |
| “The world is going to end in 12 years” (Rep. Ocasio-Cortez, 21 Jan 2019, Newsweek). | The IPCC report behind it says CO₂ should fall ~45% by 2030 on a 1.5 °C pathway (SR1.5 C.1). That is a milestone, not an end date. This is the lecture’s popularisation stage, documented at both ends. |
| “New York underwater by 2015” (the famous 2008 ABC segment). | Mostly a myth about a prophecy. The promo showed a flooded Manhattan graphic with no date, and the special put NYC flooding in 2075 (AAP FactCheck). Measured: the sea at the Battery rises 2.95 mm/yr, about 5 cm since 2008 (NOAA). |
What the instruments say
The worst case was retired, quietly
For a decade, “business as usual” meant ~5 °C by 2100 (the RCP8.5 scenario). In 2020, Nature asked researchers to stop treating it as likely. Current policies point to ~3 °C, “still a catastrophic outcome, but a long way from 5 °C” (Hausfather & Peters 2020). Climate Action Tracker’s current-policy estimate went from 3.6 °C in 2015 to 2.6 °C in Nov 2025 (CAT). Then, on 28 Oct 2025, Bill Gates wrote that climate change “will not lead to humanity’s demise”. In the same memo, in bold, he wrote that “every tenth of a degree of heating that we prevent is hugely beneficial” (Gates Notes). That is close to the “explicit acknowledgement that the apocalypse is cancelled” that the essay says never comes.
Where did Greta’s strikers go?
A fair question after the lecture. Organisers claimed 4 million strikers on 20 Sep 2019 and 6 million that week (350.org; Guardian). No independent count exists. After 2019 the movement stopped publishing global totals, and in Oct 2023 it split over Gaza (The National). Media attention shows the fade most clearly:
Verdict: half right. The fade in attention is real and measurable, which fits the lifecycle. But the underlying science was not wrong: the warming happened roughly as modelled. What failed were the popular dates. And the worst case receded partly because the world acted, which is a different outcome from a prophecy that was simply mistaken.
8AI: the newest apocalypse, and its numbers
Claim: the step to watch is why they’re sure it’s soon. For AI, as for the others, it comes down to experts who “just kind of feel a vibe”, then put numbers on the feeling.
The man who scored the last AI boom
In 1972, Sir James Lighthill reviewed British AI for the government. He found correspondents predicting machine intelligence “exceeding human intelligence by the year 2000”. His advice was to “compare the predictions of the past against performance” (Lighthill report). His report started the first AI winter. But his own benchmark failed too. He judged the field by chess, and wrote “Chess masters beat them easily.” Deep Blue beat Kasparov in 1997.
The horizon that walks with you
Each dot is a forecast: when it was made (across) and when it said human-level AI would arrive (up). Hover or tap for the source.
- The ~20-year habit is real, but weaker than its legend. A 2012 study of 95 AI forecasts found most put AI “15 to 25 years in the future”, from experts and amateurs alike. The expert-vs-amateur finding was later traced to a spreadsheet error, and the paper now carries a notice saying so (PDF; the error). A re-run still found the pull toward 10–25 years in most groups. Vernor Vinge named the bias in 1993 and then showed it himself: “within thirty years”.
- The survey moved 13 years in one year. Between the 2016 and 2022 surveys, the median year for human-level AI moved one year (2061 → 2060). Then it jumped to 2047 in 2023. When the same researchers were asked about “all occupations automated”, they said 2116, 69 years later, for arguably the same event. The authors write: “We do not know what accounts for this gap.”
- The crowd forecast is receding again. Metaculus’s median dropped ten years in 22 days in April 2022. It bottomed out at Jan 2026, a forecast made in Mar 2023. It now says Oct 2028. In 3.4 years of calendar time, the predicted date moved 2.75 years further away.
Where there’s a base rate, the forecasters agree
In 2022, the Forecasting Research Institute paid 80 subject experts and 89 “superforecasters” (people with a proven record on short-term questions) to argue for four months, with $1,000 prizes for changing minds (XPT report).
Probability of each outcome by 2100: median forecasts
Log scale. On the risk with centuries of data (asteroids, volcanoes) the two groups agree to the third decimal. On AI they differ eightfold.
- Nobody converged. The report asks “why did rational forecasters… not converge after months of debate”. A 2024 follow-up that paired skeptics with worriers for 30–80 hours each moved the skeptics from 0.10% to 0.12% and the worriers from 25% to 20% (FRI 2024).
- Near-term skill predicts nothing about long-term fear. Once 38 short-term questions resolved, accuracy on them had “no statistically significant correlation” with each person’s extinction estimate (r from −0.08 to 0.14; FRI 2025).
“What’s your p(doom)?” Group medians and some famous individuals
Log scale, from 0.1% to 100%. The individual numbers answer different questions over different horizons.
Ben’s lecture credited Yudkowsky for refusing to invent numbers, and that holds up. His 2021 essay on biology-based timelines is titled “The Trick That Never Works”, and in 2025 he wrote that p(doom)’s real social role is “some sort of weird astrological sign” (LessWrong). In fairness to the other side, though, he objects to quantifying the risk, not to being confident about it. His 2025 book is titled If Anyone Builds It, Everyone Dies.
Verdict: the numbers are unanchored, but the progress is real. The data supports Ben on the forecasts: they are unanchored, they don’t converge, and they move with the news. It does not support “there is no substance”. In the same tournament, experts and superforecasters both underestimated AI progress. On MATH, MMLU and IMO gold, they gave what actually happened average probabilities of 25% and 10%. The capability curve is measured, not felt: the length of task an AI agent can finish has doubled roughly every 7 months for six years (METR).
9Where the data pushes back on the lecture
Your doomer friends will raise these. Each one is backed by evidence.
1 · “We have just been lucky.” Nick Bostrom’s answer to “it never happened” is that civilisation hasn’t yet drawn a technology that “invariably or by default destroys the civilization that invents it”, and “the reason is not that we have been particularly careful or wise… We have just been lucky” (“The Vulnerable World Hypothesis”, 2019). He adds that we were lucky nuclear weapons turned out to be hard to make. A list of apocalypses that didn’t happen can’t rule that out, because nobody is around to write the list after one that did.
26 September is Petrov Day, the anniversary of the closest call. On 26 Sep 1983, the Soviet early-warning system reported five US missiles inbound. The duty officer, Stanislav Petrov, judged it a false alarm and didn’t pass it up the chain as an attack (record). Part of why the nuclear apocalypse didn’t happen is one man’s judgement call.
2 · Some scares faded because people fixed the problem. The ozone hole, acid rain and Y2K didn’t happen partly because of treaties, regulation and billions of dollars of work. Without the Montreal Protocol, modelling suggests two-thirds of the ozone layer would have been gone by 2065, and the UV index would have passed 30, enough to sunburn in about 5 minutes (Newman et al. 2009). The same goes for the climate worst case: current-policy warming fell by about 1 °C in ten years, partly through policy and partly through cheap solar. Warning that works looks, afterwards, just like warning that was never needed.
3 · The skeptics’ forecasts fail too. Lighthill’s chess verdict lasted 25 years. The 1966 ALPAC report found “no immediate or predictable prospect of useful machine translation” (ALPAC). That held for about 40 years and is now plainly wrong. In the 2022 tournament, the superforecasters, the skeptical group, underestimated AI progress more than the experts did.
4 · Some trend lines are measured, not felt. The warming is real (+1.55 °C in 2024). AI task length has doubled every ~7 months for six years. The lecture’s fair point is narrower than “nothing happens”: the jump from “possible in principle” to “happening soon” is where arguments go vague. In the essay’s words, “hold out for a full argument which covers all of these steps.”
10Four friendly questions for your doomer friends
These are for starting a conversation, not winning one. Each is grounded in the data above.
- “What would you expect to see by 2028 if you’re right, and what would change your mind?” The tournament found almost nobody updated. Naming the evidence in advance is the fix.
- “Where does your number come from: a model, a base rate, or a gut feeling?” Even Yudkowsky calls p(doom) a kind of astrological sign. Asking this usually gets a more interesting answer than the number itself.
- “Which past scare does this one most resemble, and how did that one end?” Ozone (fixed by acting), nuclear winter (never tested) and the population bomb (overtaken by the Green Revolution) each teach something different.
- “If it isn’t extinction, what’s the most likely bad outcome, and what would help with that one?” Both camps tend to agree on this question, and it’s the one where you can act.
Why this matters for total life satisfaction
You went to the lecture looking for arguments for your doomer friends. The data gives you something better than arguments. It gives you a way to have those conversations without either side feeling dismissed: agree on what was actually measured, and disagree about the parts that are guesses. That makes an evening with a doomer friend more fun and less draining, and it keeps you engaged with the part of the AI question you can actually work on.