Episode 9: facts, sources and further reading
Checkpoint California — Sean Forbes and Charles McGuinness on AI: how it got here so fast, what it can and can’t do, and what California owes the people whose jobs it touches.
This page goes through the checkable claims made in the episode — the ones that were right as well as the ones that weren’t — with what the sources say and links to them. Times are where each point comes up in the published episode. The episode was recorded on October 20, 2026. At the end is the reading list Charles gathered while preparing the show.
Our policy: if something said on the show is plainly false, we cut it before the episode goes out. If it’s off — a wrong number on a true point, a detail misremembered — we leave it in and correct it here and in the show notes. Corrections to this page: [email protected].
Corrections at a glance
- 10:58 — No report says half of proposed data centers were shut down. The tracker Data Center Watch counts $64 billion of projects blocked or delayed over two years, and about $130 billion in the first quarter of 2026 alone.
- 14:16 — Britain began transporting convicts to Australia in 1788, because American independence had closed the old destination and the prison hulks were full — a generation before the Luddites. No measure puts unemployment then at 30–40%.
- 21:24 — Manuel Urrutia was a magistrate in Santiago de Cuba, not a Supreme Court justice. The rest is right: president for six and a half months of 1959, and a ceremonial president (Osvaldo Dorticós) until 1976.
- 31:32 and 32:42 — Amazon has about 1.1–1.2 million US employees, the country’s second-largest employer. Its warehouse workers are employees; it’s the delivery drivers who work for contractors.
- 43:14 — The party’s website is votecnp.org, not CNP.org. The Discord is at discord.gg/5f4naVZxaz.
The claims, in order
2:08 — Iran’s Lego videos
Right in substance. The Lego-style AI videos of the Iran war — dying children, fighter jets, Trump and Netanyahu as warmongers — were made by a group called Explosive Media, which first told the BBC it was “totally independent” and then admitted the videos were made on commission from Tehran. A Canadian government analysis calls them Iranian state-linked; the reporting doesn’t say which arm of the state paid. Explosive Media claims 900 million views for the series.
Sources: Christian Science Monitor, May 2026 · Wikipedia: Explosive Media · Global Affairs Canada analysis
3:57 — Moore’s law, and buying the future with money
Right. Moore’s law, as usually stated, has computing power doubling every 18 months to two years. The money side is what the research group Epoch AI measures: the compute used to train notable AI models has grown about 4x a year since 2010, far faster than the hardware alone, because the spending grew too.
Sources: Epoch AI: training compute grows 4–5x a year · Wikipedia: Moore’s law
7:07 and 7:57 — The moon: over and over, then once more
Right. Six Apollo crews landed on the moon between July 1969 and December 1972; Apollo 17 was the last. No one went back to the moon’s neighborhood until Artemis II, which flew four astronauts around the moon without landing, April 1–10, 2026 — once, as Charles says.
Sources: NASA: Artemis II launch · Wikipedia: Apollo 17
8:46 — The Canada lynx and the snowshoe hare
Right. The Hudson’s Bay Company’s fur records, which run for most of a century, show hares and lynx booming and crashing together about every ten years, the lynx peak trailing the hare peak by two or three years. The lynx eats almost nothing else. One wrinkle for the analogy: where lynx died out, the hares kept cycling anyway.
Sources: Memorial University: the lynx–hare cycle · Krebs et al., “What drives the 10-year cycle of snowshoe hares?”
10:20 and 12:28 — No transformers, no memory chips
Right. Lead times for large power transformers in the US have gone from about a year in 2020 to two to four years, and over 160 weeks for generator step-up units in early 2026, according to Wood Mackenzie and PwC — so an order placed now does land in 2028. Memory makers have shifted production to the high-bandwidth memory that AI chips use, and DRAM contract prices roughly doubled in the first quarter of 2026; J.P. Morgan expects prices up more than 400% from the start of 2024 to the end of 2026.
Sources: pv magazine: transformer lead times extend to four years · Reuters via AJOT: utilities scramble for equipment · J.P. Morgan: the DRAM shortage · The Register: DRAM prices expected to double in Q1
10:58 — “Fifty percent of proposed data centers already shut down”
Correction. We found no report with that figure. The closest is Data Center Watch, a tracker of local opposition: its May 2025 report counted $18 billion of projects blocked and $46 billion delayed over the previous two years, and its first-quarter 2026 report put the quarter’s stalled or delayed projects at about $130 billion across 75 projects. That is a lot of money, but it isn’t a share of all proposals, and the tracker itself says attributing every delay to opposition oversimplifies.
Sources: Data Center Watch report · Data Center Dynamics on the $64 billion figure · Trending Topics on the 2026 quarterly figures
10:58 — The 69th district and Josh Lowenthal
Right. Josh Lowenthal, a Democrat, has represented the 69th Assembly District — nearly all of Long Beach, plus Signal Hill and part of Carson — since December 2022. Sean’s story about a Long Beach data center that became an industrial park is his own account; we couldn’t find a public record of the proposal either way.
Sources: California State Assembly: District 69 · Wikipedia: Josh Lowenthal
12:10 — Long Beach gets its power from Southern California Edison
Right. Long Beach is in Edison’s territory; the city’s own utility handles water and gas. The Los Angeles Department of Water and Power serves the City of Los Angeles.
Sources: SCE service area fact sheet · Solar.com: Long Beach utility
12:28 — California makes data centers pay for their own grid
Right. On September 21, 2026, Governor Newsom signed a package of seven data center bills. The governor’s office says they make data centers “pay their fair share of grid update costs” and that “any upgrades needed to supply water would be paid for by the data center”; the bills also require disclosure of electricity and water use to local governments and remove data centers from blanket environmental exemptions. The ratepayer provisions are mainly SB 886 and SB 887 (Senator Steve Padilla) and SB 1168 (Senator Jerry McNerney).
Sources: Governor’s office, September 21, 2026 · Utility Dive · CalMatters
13:40 and 14:07 — Tropical storms are new to Southern California
Supported. The National Weather Service issued its first-ever tropical storm watch for Southern California on August 18, 2023, as Hurricane Hilary approached; no tropical storm had reached the region since September 1939. So “tropical storm” and “hurricane” are indeed new words in the local forecast.
Sources: Al Jazeera, August 18, 2023 · KTLA
14:16 — The 1997–98 El Niño
Right. The 1997–98 season brought downtown Los Angeles 31.28 inches of rain against an average of about 12, with 13.68 inches in February 1998 alone, and a December 1997 storm dropped more than seven inches on parts of Orange County in a day. Seventeen people died in California’s storms that winter.
Sources: GV Wire on the 1997–98 deluge · Rebuild SoCal: 1997–98 rainfall
14:16 — The Industrial Revolution, unemployment and Australia
Correction. Transportation to Australia began with the First Fleet in 1788. Britain had sent its convicts to the American colonies; American independence closed that door in 1783, the prison hulks filled, and Botany Bay was chosen in 1785 — a generation before the Luddites of 1811–16. Some Luddites were transported there later, which is probably the connection Sean had in mind. The years after 1815 did bring falling wages and mass unemployment, with some 300,000 soldiers and sailors demobilised into a mechanising economy, but there are no unemployment statistics from the period and no basis for a 30–40% figure.
Sources: National Museum of Australia: convict transportation · History Skills: why convicts were sent to Australia · VoxDev: AI and the Industrial Revolution
16:40 — The Luddites weren’t against technology
Right. The Nottinghamshire framework knitters who began the movement in 1811 objected to wide stocking frames that turned out cheap “cut-ups” — cloth cut and sewn into stockings instead of knitted whole — and to employers using unapprenticed labour. They chose their targets, breaking the machines of masters who made shoddy goods and leaving others alone. The word’s modern meaning came later.
Sources: Encyclopedia.com: Luddites · Kevin Binfield, Luddite history · Wikipedia: Luddite
17:23 — “A ticket to Australia or we’ll hang you”
Right. The Frame Breaking Act of February 1812 made machine-breaking a capital crime; Lord Byron’s maiden speech in the Lords was against it. At the York trials of January 1813, seventeen men were hanged and others were transported to Australia, with more executions at Lancaster and Chester.
Sources: History.com: who were the Luddites · Smithsonian: Byron defends the Luddites
17:46 — “Stochastic parrot”
Right. The phrase is the title of a 2021 paper by Emily Bender, Timnit Gebru, Angelina McMillan-Major and Margaret Mitchell, “On the Dangers of Stochastic Parrots”.
Sources: Wikipedia: Stochastic parrot
19:32 — Models can’t hold a book word for word
Right, with one exception that tests the rule. Pride and Prejudice (1813) is in the public domain, and Charles’s account of what a model stores — relationships and predictions, not copies — is the standard one. A 2025 study by A. Feder Cooper, Mark Lemley and colleagues at Stanford tested 50 books against 17 open models and found that “most LLMs do not memorize most books”, in whole or in part. The exception was Llama 3.1 70B, from which they recovered nearly all of Harry Potter and the Sorcerer’s Stone and 1984. Two caveats the paper itself makes: the extraction isn’t ordinary prompting but a search over the model’s likeliest continuations, and the books it worked on are ones that appear on the web many thousands of times through popularity and piracy — the case Charles allows for when he says things seen over and over may be memorized by accident. It doesn’t show a model can produce a random book, any more than knowing the words to a hit single means knowing the B-side. The authors say the finding cuts both ways in the copyright cases.
Sources: Cooper et al., “Extracting memorized pieces of (copyrighted) books from open-weight language models” · Stanford RegLab summary
21:24 — Cuba’s first president after the revolution
Right on the story — correction on the judge. Manuel Urrutia Lleó, a magistrate in Santiago de Cuba who had refused to convict the Moncada attackers in 1953, was president from January 2 to July 18, 1959 — about six and a half months — before Castro forced him out. Osvaldo Dorticós then held the presidency until December 1976, while Fidel Castro was prime minister from February 1959 and took the presidency in 1976. Urrutia was not a Supreme Court justice. We didn’t test which chatbots get this wrong.
Sources: Wikipedia: Manuel Urrutia Lleó · Wikipedia: Osvaldo Dorticós · Cubanet on Urrutia
25:12 — AI is “the only real driver in the economy”
Depends on the measure. Harvard economist Jason Furman calculated that investment in information processing equipment and software accounted for 92% of US GDP growth in the first half of 2025, and ING puts AI investment at about a third of growth in 2026. Goldman Sachs’s Jan Hatzius counters that AI’s net contribution in 2025 was “basically zero”, because most of the equipment is imported and imports subtract from GDP. Big either way; “only driver” is the gross view.
Sources: ING: how much is AI contributing to US growth? · Goldman Sachs says AI added basically zero growth · Goldman Sachs: AI investment over $1 trillion in 2026
30:36 — Trade barriers against Chinese cars
Right. The Biden administration raised the Section 301 tariff on Chinese electric vehicles to 100% in 2024, which the industry treats as a ban, and the rate remains in force in 2026. The Supreme Court’s February 2026 ruling against the separate “reciprocal” tariffs didn’t touch it.
Sources: Gulf News: US finalizes 100% duty on Chinese EVs · The Charge Port: EV tariffs in 2026
31:32 and 32:42 — Amazon’s workforce and its contractors
Correction. Amazon is a huge employer: about 1.1–1.2 million people in the United States, the second-largest private employer after Walmart, out of some 1.5 million worldwide. Warehouse workers are Amazon employees. It’s the delivery drivers in Amazon vans and uniforms who work for “Delivery Service Partners”, small contractors — as Charles says — though two regional offices of the National Labor Relations Board found in 2024 that Amazon is their joint employer, findings Amazon is contesting.
Sources: New York Times via JoeMyGod: Amazon’s 1.2 million US workers · Statista: Amazon employees · Jacobin: the NLRB rulings on DSP drivers
32:42 — Bottles, and the machine over your shoulder
Right on the bottles — partly on the rest. In March 2021 Representative Mark Pocan said Amazon workers urinate in bottles; Amazon’s account called the claim unbelievable, then apologised a week later and admitted its drivers “can and do” have trouble finding restrooms. The warehouse tracking is real: a system that measures “time off task” and, per documents in a 2019 labor board case, could issue warnings and terminations automatically — more than 300 people at one Baltimore warehouse in a year. We found no “three strikes” rule in the reporting.
Sources: TechXplore, April 2021: Amazon admits drivers pee in bottles · Business Insider on the Verge report · Reuters via Insurance Journal, 2022
34:11 — Milton Friedman’s negative income tax
Right. Friedman proposed a negative income tax in Capitalism and Freedom (1962): below a set income, the tax runs the other way and the government pays you.
Sources: Wikipedia: Negative income tax
35:35 — Earning a little on SSI gets you kicked off
Right in substance. Supplemental Security Income cuts the benefit by roughly a dollar for every two earned after small exclusions, so wages above about $2,070 a month end it. The sharper cliff is the resource limit: $2,000 in countable assets for an individual, $3,000 for a couple, unchanged since 1989.
Sources: NerdWallet: SSI income limits for 2026 · Justice in Aging: the SSI resource limit
36:03 — Falling birth rates
Right. Provisional CDC figures put the US general fertility rate at a record low in 2025, with 3.61 million births; the total fertility rate works out to about 1.57 children per woman. California’s rate fell from 2.21 in 2007 to 1.48 in 2023, the lowest in its recorded history, and the state’s 2025 birth rate was about half its 1992 level.
Sources: US fertility rates drop to record low in 2025 · PPIC: California’s plunging birth rates · PPIC: trends in California’s birth rates
36:03 — Free school lunches for everybody
Right. California’s Universal Meals Program, which began with the 2022–23 school year, gives every public school student a free breakfast and lunch regardless of family income — the first statewide program of its kind.
Sources: CBS News: California first state to provide free meals to all students · ABC7
36:03 — “A thousand edits in an hour”
Right as a round number. Comparing the recordings with the published episodes, our last three full-length shows had 872, 968 and 687 cuts on recordings of 76–88 minutes — about 650–700 an hour, so “a thousand” is the right order of magnitude. This episode had 473 cuts in 49 minutes. The rest of Charles’s description of how the show is cut is how it’s actually done; the figures are in the show’s own records.
43:14 — “CNP.org”
Correction. The party’s website is votecnp.org; the older address californianational.party redirects there. The Discord server Charles mentions is at discord.gg/5f4naVZxaz.
What we didn’t check
Opinions and predictions — about the AI bubble, diminishing returns, what CEOs tell Wall Street, what jobs go first, universal health care, basic income and expert representation — aren’t fact-checked. Charles’s degree and work history, and Sean’s accounts of his former employer in property management, his grandfather and Los Alamitos High School, are their own. The thought experiment about a manufacturer’s CEO at 28:02 is a hypothetical.
Further reading
Links Charles gathered while preparing the episode, kept where the show went. A few are working papers or secondhand summaries; those are marked.
Boosters and skeptics, both wrong
- AI skeptics and AI boosters are both wrong — Timothy B. Lee, Understanding AI
- The phony comforts of AI skepticism — Casey Newton, Platformer
- Dario Amodei: “We are near the end of the exponential” — the Anthropic CEO’s timeline, in his own words (Dwarkesh Patel interview)
- The AI Con — Emily Bender and Alex Hanna’s book; a review (OLR Supplement)
The money
- The AI buildout enters the show-me-the-money phase — Axios
- AI spending is surging faster than revenue — Forbes
- Hyperscalers are spending 102% of cloud revenue on capex — 24/7 Wall St., citing UBS
- Global AI investment forecast to exceed $1 trillion in 2026 — Goldman Sachs, the less worried view
- California’s budget projections rely heavily on a continued AI boom — CapRadio
- Betting on the bubble: legislative analysts say Newsom is overestimating AI revenue — EdSource
The S-curve, and the things that ran out
- International AI Safety Report 2026 — on rising costs per gain
- Transformers in 2026: shortage, scramble, or self-inflicted crisis? — POWER magazine
- Four bottlenecks shake AI data center boom — Seoul Economic Daily, on the New York Times report
- $130 billion in AI data centers stalled: the bottleneck is consent — Forbes
- Data Center Watch report — the tracker of projects blocked or delayed by local opposition
- AI’s water strain includes surging demand from electricity use — Insurance Journal
- METR time horizons: now 10x a year — LessWrong, the strongest evidence against the slowdown
- The 10x-a-year clock: what METR’s AI time horizons really show — AGI Tracker
Jobs and productivity
- AI, productivity and labor markets: a review of the evidence — International Center for Law & Economics (summarizes the Stanford/ADP and 21,000-firm working papers)
- Has AI impacted the labor market yet? — Alex Imas
- Radiologists are the canaries in the coal mine — Joshua Gans, who was there when Hinton said “stop training radiologists”
- METR didn’t reverse its 19% AI slowdown — on the developer study and its 2026 follow-up
- Handling hallucinations: a reflection on the last three years — UNC Law Library, on AI-invented citations in court
Who gains: the small and the new
- Generative AI at Work — Brynjolfsson, Li and Raymond (NBER): an AI assistant raised customer-support output 14% on average, 34% for novices, barely at all for experts
- GPT-4 could be the great skill leveler for consultants — The Decoder, on the Harvard/BCG “jagged frontier” study, including where AI made consultants worse
- AI startups, solopreneurs and actual numbers — Forbes