← Back to the letter

Every source in this letter

Every factual claim in the letter, with the receipt it stands on. If you find one that doesn't hold, text me.

The opening

  1. OpenAI studied more than a million ChatGPT conversations and found roughly three quarters were people looking up information, getting practical advice, or getting help writing.National Bureau of Economic Research
  2. On June 9, Anthropic released Claude Fable 5, and said its capabilities passed anything it had made available to the public before.Anthropic
  3. Dario Amodei's 2024 essay argues AI could compress fifty to a hundred years of medical progress into five or ten, and put the possible drop in cancer deaths at 95 percent or more. The whole thing could arrive as early as 2026.Dario Amodei, Machines of Loving Grace
  4. A Commerce Department letter, sent Friday June 12 at 5:21pm, gave Anthropic ninety minutes to cut off every foreign national on earth, including its own employees, which left only one way to comply: turn the models off for everyone.CNBC
  5. The public version of Fable was already the restricted one, with its strongest capabilities locked behind safety filters.Fortune
  6. The unfenced model, Mythos, finds and exploits software flaws better than any other model and all but the most skilled human experts, and went only to a selected list of approved organizations, never the public.Anthropic, Project Glasswing
  7. A June 26 letter restored Mythos for the approved list. The public did not get Fable back until July 1. The club got it first.Bloomberg
  8. Fable came back only after Anthropic rebuilt its safety filters and the government, in the Commerce Secretary's own words, worked with the company to analyze and approve the model.The Hill
  9. While Fable was still dark, the White House asked OpenAI to restrict its newest models to a small circle of trusted partners before launch, and OpenAI did it while grumbling publicly that it shouldn't become the norm.OpenAI

What you sell

  1. On a Google-proof test of graduate-level science, PhDs in the exact field score in the high 60s, the top models score in the low 90s, and a PhD from another field with Google and half an hour scores about 34.International AI Safety Report
  2. Mathematicians built original, unpublished problems in late 2024 to stump AI, and Terence Tao expected them to hold for several years. Models cleared more than half of them in about eighteen months.BenchLM, FrontierMath tracker
  3. The best model went from fixing under 2 percent of real software bugs in 2023 to roughly 95 percent, and the benchmark was retired in early 2026 because the machines maxed it out.BenchLM, SWE-bench Verified tracker

What an agent actually is

  1. The creator of Claude Code says there are days he manages tens of thousands of AI agents at once.Fortune
  2. The same person says he hasn't hand-written a line of code in eight months.Fortune
  3. Andrej Karpathy, an OpenAI co-founder, joined Anthropic in spring 2026 to lead a research team.CNBC
  4. Karpathy calls today's agents glitchy, says they forget what you tell them between sessions, and puts fully working ones about a decade away.Dwarkesh Patel podcast
  5. Two months later, the day after Christmas, Karpathy posted that he'd never felt this much behind as a programmer, that he could be ten times more powerful stringing existing tools together, and that failing to feels like a skill issue.Andrej Karpathy, on X
  6. Pew runs the biggest survey in America on how people use AI, and its 2026 edition doesn't contain a single question about agents.Pew Research Center
  7. The president of Goldman Sachs described the firm as a human assembly line about to change the way factories did when the robots came.Bloomberg
  8. Goldman's CEO says AI can draft 95 percent of an IPO prospectus in minutes, work that used to take a six-person team weeks, and that the last 5 percent now matters because the rest is a commodity.Fortune
  9. Goldman, JPMorgan and Citi are shrinking their junior analyst classes by as much as two-thirds as AI absorbs entry-level work.Fortune
  10. Stanford tracked payroll data covering about one in six US workers and found that since late 2022, employment for 22-to-25-year-olds in the most AI-exposed jobs fell 16 percent relative to everyone else, while older workers in the same jobs held steady.Stanford Digital Economy Lab, Canaries in the Coal Mine
  11. McKinsey now puts an AI in the room for final-round interviews and has candidates run it live.Management Consulted
  12. AI mentions in entry-level job postings nearly doubled in a single year.CNBC

Your job, specifically

  1. For the first time ever, AI became the number one reason American companies gave for cutting jobs, and it has held that spot every month since.CNBC
  2. The CEO of Salesforce cut his support team from about nine thousand people to five thousand, explaining it in four words: I need less heads.CNBC
  3. The club whose access came back first is a real program, Project Glasswing, whose stated job is defending the systems everyone depends on, from power grids to hospitals to banks.Anthropic, Project Glasswing
  4. Anthropic put two hundred million dollars into studying what AI does to workers.U.S. News & World Report

About those hands

  1. Tesla announced a humanoid robot in 2021 with no hardware to show, so a person in a white spandex suit danced on stage.TechCrunch
  2. Musk now tells shareholders the robot will be the biggest product of all time, bigger than the cell phone, with eventually tens of billions of them.CNBC
  3. Morgan Stanley projects a billion humanoid robots on Earth by 2050.Morgan Stanley
  4. The bank tracking this closest multiplied its own forecast several times over in about a year as robot costs collapsed, and the cheapest humanoid has fallen by roughly three quarters in two years.Goldman Sachs
  5. The New York Times scored fifteen years of Musk's public promises, 602 with real deadlines, and found he delivered about one in five on time.The Seattle Times (New York Times syndication)
  6. Self-driving was a year away for close to ten straight years.The Autopian
  7. Five thousand robots were promised for 2025, and the program was still in R&D as of early 2026.Electrek
  8. Against that record, a SpaceX booster landed itself for the thirty-fifth time, and the company went on to the largest IPO in history. The destination arrives even when the deadline slips.CNBC
  9. The much-celebrated humanoid fleet at BMW turned out to be a single robot on a pilot.BMW Group
  10. A robot hand still can't come close to the roughly seventeen thousand touch receptors in a human one.Rodney Brooks
  11. China's government ordered more than ten thousand humanoids into real jobs by the end of the year. An order, not a forecast.eWeek

The fork

  1. The legal authority behind the shutdown letter had never been used that way before. No published standard, no hearing, no appeal.Center for Strategic and International Studies
  2. Nearly eighty security executives signed an open letter demanding an open, transparent process for decisions like this, and got silence.Cybersecurity Dive
  3. The government's case was a reported jailbreak; the White House says it asked for a fix first and got refused, and the company disputes that flatly.Snyk
  4. Two days before the letter, Amodei published an essay asking the government to build exactly this kind of safety authority done properly, with testing, written standards, and a guarantee that citizens keep access to AI when a government acts badly.Dario Amodei, Policy on the AI Exponential
  5. Roughly three quarters of Americans tell pollsters they worry about AI concentrating too much power.Pew Research Center
  6. Ten days before the letter, an executive order created a system where AI companies volunteer their newest models to the government for review, up to thirty days before the public sees them, with the capability bar set by a classified process.The White House
  7. OpenAI sent its newest models to a small group of trusted partners shared with the government, and said the same day that the arrangement keeps the best tools from the people who need them.OpenAI
  8. Anthropic's own approved list for its top model includes Apple, Google, Microsoft, and JPMorganChase, and the company says it does not plan to make that model generally available.Anthropic, Project Glasswing
  9. As the floor of who's-needed shifts, the people building AI are already buying insurance. Anthropic put two hundred million dollars into studying AI's effect on workers, and its leaders talk more about everyone owning a piece.U.S. News & World Report

What's truly coming

  1. Sam Altman wrote that we are past the event horizon and the takeoff has started.Sam Altman, The Gentle Singularity
  2. OpenAI set public goals of an AI research intern by September 2026 and a fully automated AI researcher by March 2028, while admitting it might miss those dates.Sam Altman, on X
  3. Amodei wrote in January 2026 that powerful AI could be one to two years away, that there's a very strong chance it comes in the next few, and that the loop where AI builds the next AI may be only one to two years off.Dario Amodei, The Adolescence of Technology
  4. Amodei's term for that machine, smarter than a Nobel Prize winner across most fields, with millions of copies running at once, is a country of geniuses in a datacenter.Dario Amodei, Machines of Loving Grace
  5. An independent research group, METR, measures how long a stretch of human work an AI agent can finish on its own; the doubling sped from about seven months to about four, the best models now run around five hours at a coin-flip success rate, and the trend points to month-long projects before the decade is out.METR
  6. Google says three quarters of its new code is now written by AI and approved by engineers, up from half the previous fall.Google
  7. The four biggest tech companies spent around four hundred billion dollars on AI infrastructure in 2025 and have guided investors to between six and seven hundred billion for 2026.Fortune
  8. Seven datacenter sites are under construction for OpenAI alone, planned to draw about as much power as New York City at its peak.Epoch AI

The bubble

  1. Leaked financial statements verified by the Financial Times show OpenAI brought in about thirteen billion dollars in 2025 and spent thirty-four billion doing it, a roughly twenty-one billion dollar operating loss, and its own documents don't project profitability until the end of the decade.Fortune
  2. Anthropic told investors in May 2026 that it had crossed a forty-seven billion dollar run rate.Anthropic
  3. Half the population started using this technology within three years of its existing, faster than the PC and faster than the internet spread.Stanford HAI, AI Index 2026
  4. The price of yesterday's best AI keeps collapsing, by roughly forty-fold per year on some measures.Epoch AI
  5. In the late 1990s, telecom companies buried more than five hundred billion dollars of mostly borrowed money in fiber. Twenty-three of them went bankrupt, and in 2002 only about three percent of that fiber was even lit.Paul Starr, The Great Telecom Implosion

Your move

  1. Since February 2026, ChatGPT shows ads in its free version.TechCrunch
  2. Roughly half of the new articles on the internet are now written by AI.Graphite
  3. Merriam-Webster's word of the year was slop.Merriam-Webster
  4. When Google answers your question with AI at the top of the page, clicks to the actual sources get cut roughly in half.Pew Research Center
  5. The number one skill employers say they want is analytical thinking, and about seven in ten companies call it essential.World Economic Forum, Future of Jobs 2025
  6. AI literacy is the fastest-growing skill in American job postings.LinkedIn
  7. Harvard Business Review named workslop, AI output that looks like work and says nothing, and found it takes about two hours to untangle every time it lands on someone's desk.Harvard Business Review

← Back to the letter