Keen On America

It’s always fun to have Nobel Prize winners on the show. Recently, we’ve hosted the journalist-dissident Maria Ressa, the economist Simon Johnson and the gloomy godfather of AI Geoffrey Hinton. Today it’s the Delhi-based Kailash Satyarthi, who won the 2014 Peace Prize for his fight against child exploitation.
He is now trying to globalize the ideal of compassion. But for Satyarthi, compassion doesn’t mean kindness, empathy, mercy, or even pity. In his new book Karuna (a Sanskrit word that loosely translates into compassion), he defines it as a disruptive force. It means feeling somebody else’s problem as your own and then doing something about it. Which is why, keynoting the UN’s AI for Good summit in Geneva this July, he challenged the executives in the room to send their people to witness suffering. To see the 470 million children living in war zones, the 138 million working as labourers and slaves, the 270 million out of school. That, he guarantees, would make do-gooders genuinely compassionate.
Which brings us to artificial intelligence. Satyarthi is no fan of the AI arms race, which he thinks has left Big Tech’s engineers unable to identify the ethical point of their products. But he insists he’s not against the companies themselves and says he remains a believer in their potential humanity. Satyarthi certainly sees the power of AI to change the world. That’s why, last week, he came to Silicon Valley to talk about AI.
Geoffrey Hinton told me that he wants AI to acquire a female sensibility. Kailash Satyarthi, in contrast, wants compassion to define AI ethics. One day, no doubt, a compassionate bot, maybe named Karuna, will win a Nobel Prize. If so, we’ll certainly have them on the show.

FURTHER READING

•   Kailash Satyarthi, Karuna: The Power of Compassion (HarperCollins India, 2026) — also published in Hindi as Karuna: Samaadhaan ki Shakti

•   Globalising Compassion — Satyarthi’s own account of the idea, including the four components he sets out in this conversation: awareness, connectedness, feeling and action
•   The Satyarthi Movement for Global Compassion, launched in New Delhi in 2024, which is now the vehicle for the compassionate-leadership work in technology, governance, academia and the judiciary
•   Geoffrey Hinton on the show last year, warning about “alien beings that could take over” — the episode behind the question about building different values into the machines: listen or watch

•   Anthropic’s Matthew Botvinick on Thursday, arguing that a safe AI system dropped into an authoritarian setting is not a safe system — the mirror image of Satyarthi’s case: listen or watch

•   The Atlantic’s Nicholas Thompson, whose own conversation with Satyarthi about building compassion into AI models comes up here: listen or watch

FIVE TAKEAWAYS

1. Machines cannot feel, so the compassion has to come from the people building them. This is where he starts, and it is an engineer’s starting point rather than a moralist’s. Codes and algorithms cannot be compassionate and never will be, and he has no interest in pretending otherwise. What he objects to is the companion assumption that the machine is therefore neutral. It is not neutral, because it is made by people who arrive with their own biases, apprehensions, flaws and mindset, and all of that flows into the thing they build. The conclusion he draws is that compassion has to be a design principle at the foundation stage, carried through development and deployment, rather than a correction applied after the harm has been exposed.
2. Compassion, in his usage, is not kindness, and the distinction is the whole argument. He is emphatic about it, and he repeats the list twice in the hour: not kindness, not empathy, not mercy, not pity, not charity, not welfare. Those are good human traits and he does not disown them, but they are not what he means. Compassion as he has spent five decades redefining it is a disruptive, dynamic and transformative force, born from feeling the problems of others as your own and driving you to mindful action to solve them. Shorter: mindful and pragmatic problem solving. It has four components — awareness, connectedness, feeling, and then action — and the fourth is the one that separates his definition from every sentimental version of the word.
3. “AI for Good” invites the obvious follow-up question, which he put to the room in Geneva. Good for whom? He was keynoting the inaugural session of the UN’s AI for Good summit this July, in front of delegations from 177 countries and twelve thousand people from the tech sector, and the challenge he laid down was whether the good in question belonged to world leaders, governments and eight or nine companies. His numbers: 470 million children living in war zones, 138 million working as child labourers and child slaves, 270 million out of school. His practical ask is disarmingly small — send the engineers to spend a week or ten days in sub-Saharan Africa, in South Asia, in the war zones of the Middle East, and he will organise the trip himself. His verdict on the race as it stands is that it is an arms race, and that in two years’ time only the regret will remain in our hands.
4. Morality is borrowed from outside; compassion is already inside. This is his quarrel with the responsible-AI vocabulary, and he extends it, politely, to the Pope, whose interventions on technology he otherwise endorses. Ethics, morality and religious sermons all come from outside, he argues, and their priorities keep changing, which is why the world still has its problems despite all the temples, mosques, churches and synagogues. Consciousness and compassion, by contrast, are twin sisters, biological and neurological and native to us. So the task is not to import a code of conduct but to ignite something that is already there — which is also why he insists that his argument is not a theological one, even when the conversation strays into Genesis.
5. The young people in his own office write better and think less. He calls the measurable version of all this the compassion quotient, and his claim is that most of us run a low one, caged by the layers that accumulate from identity to purpose to the prevailing definition of success; exposure to actual suffering raises it, as he has watched happen to the volunteers, researchers and journalists who spend a week in his ashrams. The observation that lands hardest, though, is domestic and unflattering to the technology. The youngsters who work for him can now produce a speech or an article better than he could think of — and there is not a single new creative idea in any of it, just the previous data and the old biases that came with it. The creative power and the inclusivity of the young, he says, are now the matter of concern.

ABOUT THE GUEST

Kailash Satyarthi was awarded the Nobel Peace Prize in 2014, jointly with Malala Yousafzai, for his...

“A human being is worth more than his opinions”, Hannah Arendt wrote to her Zionist friend Gershom Scholem in 1946, after he read her critical essay on Zionism and accused her of betraying the Jewish people. So how to value a human being? What, according to Arendt, are we worth? According to the Arendt scholar Roger Berkowitz, our value lies in the power of friendship. That’s what he calls, in his new book, A World We Share. It’s what makes us human.
Berkowitz, who runs the Hannah Arendt Center at Bard College, reminds us that Arendt wasn’t a postmodern relativist who believed everyone’s opinions were equally credible. Facts mattered to her enormously — but a thousand facts about immigration will not verify migration policy, and two people with Harvard doctorates, especially if they are friends, can still disagree about how many people to let into America.
Nor is Berkowitz presenting Hannah Arendt as an anodyne centrist. For Arendt, he argues, people are plural and always will be so. We can, and should, have strong opinions that sometimes cause us to raise our voices in anger. But an argument, she believed, should end in disagreement rather than discord. So, when she and Scholem fell out all over again seventeen years later, this time over her argument about the banality of Nazi evil in Eichmann in Jerusalem, she hoped it wouldn’t affect their friendship. It’s our ability to disagree and reconcile (which isn’t the same as forgiveness) that makes us human, Arendt says. That’s what we should — indeed, must — share if we are to rebuild our broken world.

FURTHER READING

•   Roger Berkowitz, A World We Share: Hannah Arendt and the Power of Friendship in a Broken World (Yale University Press, 2026)
•   The Hannah Arendt Center for Politics and Humanities at Bard College, which Berkowitz founded and directs
•   Yesterday’s episode featuring Anthropic’s Matthew Botvinick, author of AI and Political Freedom, articulating an engineer’s view of politics — listen or watch

•   Sherry Turkle on Artificial Intimacy — last week’s episode Berkowitz praises: listen or watch

•   Keri Leigh Merritt on the American dissident Lillian Smith, who, like Hannah Arendt, was An Inconvenient Woman: listen or watch

FIVE TAKEAWAYS

1. Better facts will not fix politics. This is the claim that puts Arendt at odds with almost everyone currently worrying about democracy. We tend to assume that misinformation is the disease and that better education, better journalism and better facts are the cure. Arendt did not believe it. Facts mattered to her enormously, but she insisted they have very little power to tell us what to do: twenty facts about immigration still will not settle immigration, and two people who have both been as well educated as it is possible to be can still disagree about Trump. Berkowitz traces the opposite assumption back to Plato, who watched Athens execute Socrates, concluded the many could not be trusted, and invented the philosopher king to rule in their place.
2. Reconciliation is not forgiveness. Arendt’s distinction here is precise and worldly rather than moral. Forgiveness, as she understood it, is Christian: it implies that you might have done the same, so you should not throw stones. Reconciliation claims nothing of the kind. It says that what you did was wrong, that I would not have done it, that I do not forgive you — and that I am nonetheless willing to love a world with you in it. The condition is that it runs both ways. Heidegger came to her after the war, went down on his knees and turned himself in, and she reconciled. Karl Jaspers never did, and the two of them had their worst row over it in Switzerland in the fifties, ending with each accepting that the other had made a different judgment.
3. And there are people you cannot have in your world. The last two chapters of the book are about the limits, which Berkowitz is at pains to point out are not the same thing as civility. Eichmann is the case that settles it. Arendt’s argument was never that he was a monster; it was that in a normal age he would have been an unremarkable civil servant whom nobody would have heard of. But he did not live in a normal age, and he became a willing and obedient servant of bureaucratic genocide. Her conclusion was not moral outrage but a judgment about what she could live alongside: she could not love a world in which such a person was simply readmitted to it, and so he had to die. The useful corollary, in Berkowitz’s formulation, is that the uncle whose opinions you cannot bear is generally not Eichmann.
4. Totalitarianism is organised loneliness. The line is Arendt’s, and Berkowitz keeps it on the Arendt Center’s website: if totalitarianism is organised loneliness, then the loneliness of modern life means totalitarianism is hardly a phenomenon of the past. His gloss is the darkest stretch of the conversation. Religion has lost most of its public meaning, the family has broken down, and the Western tradition — the banisters we lean on precisely because nobody can think all the time — has shattered. What is left is the atomised individual, and the movements that offer him meaning. Which is why Arendt held that the fundamental political problem of the age is not housing or immigration or even the environment, but meaninglessness, and why anyone hoping to see off fascism has to offer people a life that feels worth living.
5. There is no such thing as a two-week-old friendship. The third element of friendship, after conversation and respect, is simply duration, and Arendt was unsentimental about it. She knew W. H. Auden for twelve years, at one point turning up with Mary McCarthy to clean his notoriously filthy flat, and he proposed marriage; she still said they had met too late in life to become friends. Berkowitz, at fifty-eight, found himself making a list and ringing people up once a month because the friends he had were leaving New York or dying. The argument he draws from it is that friendship takes time nobody has any more, and that if machines are going to take the labour, it is the capacity most worth protecting on the other side.

ABOUT THE GUEST

Roger Berkowitz is Founder and Academic Director of the Hannah Arendt Center for Politics and Humanities at Bard College, where he is Professor of Politics, Philosophy and Human Rights. A World We Share: Hannah Arendt and the Power of Friendship in a Broken World is published by Yale University Press. He lives in New York.

ABOUT KEEN ON AMERICA

Keen On America is a daily conversation about the ideas, books and arguments shaping the United States, hosted by Andrew Keen, author of The Internet Is Not the Answer and How to Fix the Future. New episodes every day.

Matthew Botvinick works at Anthropic, where he leads the Anthropic Institute’s research on AI and the rule of law. His new book, AI and Political Freedom, focuses on AI’s risk to democracy and how we should respond to it. It’s important to note that for this Princeton University Press book, Botvinick doesn’t represent Anthropic. That said, the fact that he just happens to be a senior policy guy at Dario’s trillion dollar startup makes both Matt and his book particularly intriguing.
He obviously knows his AI stuff. He was previously senior director of research at Google DeepMind, and before that ran neural computation labs at Princeton and Penn. He has been doing what we now call deep learning since the early nineteen nineties, when it was called connectionist modelling and few imagined it would ever amount to a commercial product. Then he went back to school for a law degree. Yale, no less. So Matt is clearly an Ivy League-trained, smart logical guy. Somebody to be (mis)trusted.
His argument is that AI doesn’t invent a new threat to democracy so much as raise the stakes on the demagogue’s playbook. It’s that familiar Trump/Modi/Orbán/Bolsonaro game of splitting the people from the elites, then, once in office, rewiring the state to entrench power and muzzling anyone who objects. What AI adds, Matt worries, is populist fuel at both ends. Thus the dangerous flammability of our current politics. Job displacement manufactures exactly the insecurity demagogues feed on, and AI hands whoever wins a surveillance and censorship apparatus that no previous authoritarian could dream of. He also expects AI-driven national security crises, and notes that states of emergency are designed to expand executive power, which makes them very easy to abuse.
I pushed Matt on whether anyone should believe any of this from a senior Anthropic policy maker. He said no, and probably meant it. There is no reason, he argued, to trust anybody from Anthropic, or any newspaper either, because a marketplace of ideas doesn’t run on trust but on the strength of the argument.
Perhaps. But my bigger concern is the ethic of the all-too-logical engineer that Matt brings to politics. As I will discuss tomorrow with the Hannah Arendt scholar Roger Berkowitz, Arendt argued that politics was anything but an engineering problem with quantifiable fixes. So when Matt promises AI enabling ancient Athens “at scale” through citizen assemblies, we should be wary. Not because this will end humanity or make Anthropic employees exorbitantly rich. But because politics isn’t code, and can’t be fixed by engineers.

FURTHER READING

•   Matthew Botvinick, AI and Political Freedom: The Risk to Democracy and How to Respond (Princeton University Press, 2026)
•   The author’s own page for the book: mattbotvinick.com

•   Adam Jentleson on Supermajority — the Democratic strategist on rebuilding a governing coalition: listen to that episode

•   Susan Neiman on whether Trump is evil — the philosopher on how half of America accepted him twice: listen to that episode

•   Jonathan Rauch and Peter Wehner on the patrimonial turn in American government: listen to that episode

•   Roger Berkowitz, A World We Share: Hannah Arendt and the Power of Friendship in a Broken World (Yale University Press, 2026) — Berkowitz is on the show tomorrow

FIVE TAKEAWAYS

1. AI doesn’t invent the threat. It raises the stakes. Botvinick is careful to say that AI does not create some exotic new danger to democracy. What it does is sharpen a vulnerability that democracies have always had, and which the American founders worried about explicitly. The pattern he calls the demagogue’s playbook begins with populist appeals that divide the people from an allegedly hijacking elite, and proceeds, once the demagogue is in office, to two objectives: rewiring the machinery of government to entrench power, and muzzling dissent by attacking the press and civil society. AI feeds the first stage and arms the second.
2. Job displacement is the fuel. The mechanism by which AI raises the risk is economic rather than technological. People who feel insecure about their livelihoods and their standing in society look to political leaders to solve problems they cannot solve privately, and that is precisely the demand that demagogues exist to meet. Botvinick is candid that he cannot predict the timeline, because how fast AI diffuses through the economy and how fast employers actually use it to replace labour are harder to forecast than the technology itself. His position is simply that serious displacement is coming and that we should prepare rather than argue about the date.
3. Emergencies and surveillance are the weapons. Two tools in particular worry him. The first is the state of emergency: he expects AI to cause national security crises, cybersecurity above all, and points out that states of emergency are not inherently anti-democratic but are intrinsically designed to expand executive power, which makes them very easy to exploit. The second, and the one he considers more serious, is that AI is an extraordinary instrument of domestic surveillance and censorship — for tracking political opponents, silencing them, and preventing the coordinated action that protest depends on. China, he notes, is not a hypothetical.
4. There is no reason to trust him, and he says so. Asked directly why anyone should believe an Anthropic employee with stock in the company, Botvinick did not reach for a reassurance. There is no reason to trust anybody from Anthropic, he said, just as there is no reason to trust any newspaper; the point of a marketplace of ideas is that it does not run on trust at all, but on the strength of arguments and the information people put into the open. AK pressed back that the marketplace of ideas is itself a contestable position rather than a self-evident truth, which Botvinick accepted without abandoning it.
5. Ancient Athens at scale. The hopeful half of the book rests on the deliberative democracy movement and the citizens’ assemblies championed by the Yale political theorist Hélène Landemore, whom Botvinick calls one of his heroes. These assemblies are good at putting people in a frame of mind to argue openly and from facts, but at present nobody knows how to run one much beyond a hundred participants. The question the book leaves open is whether AI can enlarge them to the scale of a republic without destroying the thing that makes them work — and Botvinick is explicit that this requires experiment rather than deployment.

ABOUT THE GUEST

Matthew Botvinick is a member of the technical staff at Anthropic, where he leads the Anthropic Institute’s research on AI and the rule of law, and a senior fellow at Yale Law School. He was previously se...

The retired money manager Greg Steinmetz has now written three books about very rich men. The first, The Richest Man Who Ever Lived, was about the sixteenth-century Augsburg banker Jacob Fugger. The second, American Rascal, was about the nineteenth-century robber baron Jay Gould. The third, The President and the Oligarch, came out this week, and this time the rich man has a president as a co-star. It’s a book about the relationship between Franklin D. Roosevelt and Andrew W. Mellon.
Mellon was the third richest American of his age, behind Rockefeller and Ford. He was, above all, a banker, installed as president of his father’s small-town Pennsylvania bank while still in his twenties, and building it into one of the largest banks in the country. Warren Harding, reasoning that anyone that rich must know something about money, made him Treasury Secretary. He served Coolidge and Hoover too, eleven years in all, the second longest tenure in the department’s history.
And then, in 1934, Franklin Roosevelt came for him. Steinmetz’s revisionist argument is that FDR’s pursuit of Mellon was pure politics. Mellon, who by then was in his late seventies and out of office, had done nothing to him. His supposed crime was taking a tax write-off on art he had donated to a foundation, the sort of thing, Steinmetz claims, (rich) people do every day. Roosevelt’s own attorney general told him there was no case and a grand jury threw it out. Roosevelt pursued him in tax court anyway, until an unloved Mellon died with the matter unresolved.
Steinmetz suggests this is where the modern American fashion of deploying the tax code against political enemies begins. The irony is that Mellon’s paintings that triggered the persecution morphed into the National Gallery of Art — for which Roosevelt took the credit. Poor persecuted Andrew W. Mellon. Or, more accurately, rich persecuted Mellon. Unlike Steinmetz, I find it hard to feel sorry for such a miserable old sod who didn’t even have any fun with his wealth. Maybe Steinmetz’s next book should be about persecuted poor people. He’d find no shortage of subjects.

FURTHER READING

•   Greg Steinmetz, The President and the Oligarch: Roosevelt, Mellon, and the Triumph of Big Government (Simon & Schuster, 2026)
•   Greg Steinmetz, The Richest Man Who Ever Lived: The Life and Times of Jacob Fugger (Simon & Schuster, 2015)
•   Greg Steinmetz, American Rascal: How Jay Gould Built Wall Street’s Biggest Fortune (Simon & Schuster, 2022)
•   William Hitchcock on A Shadow Over the World — the companion revisionist case on FDR and the 1930s: listen to that episode

FIVE TAKEAWAYS

1. The man who had no fun. Mellon was the third wealthiest American of his era, behind only Rockefeller and Henry Ford, and he appears to have enjoyed almost none of it. One contemporary journalist wrote that he looked like an accountant who had just lost his job. His English wife, brought over from horse country outside London, took one look at the smokestacks of Pittsburgh and within three years had taken up with an Englishman. Steinmetz describes a man who was never happier than when he was alone in the bank with a balance sheet, and likens him to Warren Buffett reading financial statements all day — except that Buffett at least plays bridge.
2. The vendetta. In 1934, with a midterm approaching and pressure building from Huey Long, Father Coughlin and Francis Townsend on his left, Roosevelt needed to demonstrate his progressive credentials without adopting their actual programmes. He chose to scapegoat Mellon, blaming him for the persistent unemployment and for the Depression itself. Mellon was by then in his late seventies and years out of office, and the case against him was a tax write-off taken on art donated to a foundation. Steinmetz is blunt about what followed: the attorney general said there was nothing there, a grand jury of ordinary working people threw it out, and Roosevelt pursued him in tax court regardless, until Mellon died with the matter still unresolved.
3. What actually caused the Depression. Steinmetz changed his own mind during the research. The stock market crash was not the cause, he argues, pointing out that the Nasdaq fell seventy per cent in the dot-com bust and produced only a mild recession. What turned a downturn into a decade was European: German reparations, German bank failures, and the indecision of central bankers across the Atlantic, compounded at home by the Federal Reserve’s refusal to put money into the economy — the Friedman and Bernanke diagnosis. On this reading, Mellon’s record at Treasury was in some ways the opposite of ruinous: he ran a surplus every single year, and paid down the debt built up by the First World War by cutting taxes rather than raising them.
4. The National Gallery. The strangest turn in the story is that the donation which triggered the prosecution ended in a gift to the American public. Mellon owned a hundred and twenty-nine masterpieces — Rembrandts, Titians, Raphaels — and he gave them, along with the money to build the building and endow its maintenance, to create the National Gallery of Art. He pointedly did not call it the Mellon Gallery. He and Roosevelt met one to one to discuss the gift while the tax trial was still grinding on, and Roosevelt took the credit for the museum. Steinmetz forecloses the cynical reading: Mellon was already dying of cancer, and the value of what he gave away dwarfed anything the tax case could ever have cost him.
5. The pendulum. The research also altered Steinmetz’s politics, and in an unexpectedly hopeful direction. He is unmoved by the argument that American democracy is finished. American history, he says, is a pendulum: Harding won sixty per cent of the vote in 1920, the biggest landslide since Washington, because voters wanted tax cuts and an end to inflation; by 1932 those same voters demanded relief and got it. He expects the same correction now, with the Democrats reversing much of what Trump has done. For all that his book is a grim story about persecution, he takes from it the conclusion that democracy tends to solve its own problems.

ABOUT THE GUEST

Greg Steinmetz was born in Cleveland, took his BA at Colgate and a master’s at Northwestern’s Medill School of Journalism, and spent fifteen years as a reporter, including as Berlin and London bureau chief for the Wall Street Journal. He later became a securities analyst and a partner at a New York money-management firm, from which he is now retired. He is the author of The Richest Man Who Ever Lived, a biography of Jacob Fugger, and American Rascal, on Jay Gould. The President and the Oligarch: Roosevelt, Mellon, and the Triumph of Big Government is published by Simon & Schuster. He lives just north of New York City.

ABOUT KEEN ON AMERICA

Everyone, That Was The Week publisher Keith Teare breathlessly announced this week, is about to be offered an AI personal assistant. Or, to put it in Silicon-Valley-speak, a digital chief of staff. OpenAI unveiled Dot at its developer day. Meta already has Muse, Musk has Grok Bot, and a British start-up called Instinct raised a billion dollars at a ten billion dollar valuation while adding group chat, so that an entire team can share this same digital master of the universe. My instinct is that Instinct might be the Pets.com of this bubble. I would check with my digital chief-of-staff, if I had one. Which I don’t.
 
The novelty, Keith insists, is the virtual machine. It’s a computer in the cloud that your agent owns, where it can run a browser, read your email and write code on your behalf. So we have one interface instead of the fifty chat windows that I use on Claude. No keyboard, no mouse, no screen. As a particularly wealthy and busy man, Keith has given five of these assistants access to everything (and I mean everything) he has: multitudes of credit cards, bank accounts, brokerage logins. He confesses to spending a combined thousand dollars a month on AI subscriptions. But the fruit of all this — his argument for our agentic future — is that he recently asked one of them, by voice, to find him a sub $40 Hoya lens hood with the correct thread. $1000 to save $40? Even a human chief-of-staff might flag that raspberry.
 
What do we become when we talk to machines? I asked Sherry Turkle this week. Idiots, I suspect.
 
Discussed also this week:
 
•       Nicole Nguyen in the Wall Street Journal on whether she would let an AI agent run her life.
•       David Robinson in The Atlantic on why he quit OpenAI, and why frontier labs should be run like nuclear plants or busy airports.
•       Adam Tooze in the Financial Times on why AI is not the Manhattan Project — and Evgeny Morozov, in the same paper, on AI sovereign wealth funds as techno-imperialism.
•       Steve Rattner’s six charts on why we need AI, including the case that Democrats rarely make — with a sting in the tail about kill switches.
•       Richard Waters in the Financial Times and Alex Kantrowitz on why Anthropic’s IPO feels so weird.
•       The New York Times on the religious scholars who met Anthropic and came away stunned.
•       That Was The Week — Keith’s newsletter, with everything discussed here, and last week’s editorial.
 
Five Takeaways
 
•       Everyone Gets Offered One. The week’s actual news, in Keith’s reading, is not that everybody will have a personal agent but that everybody will be offered one. OpenAI announced Dot at its developer day, joining Meta’s Muse, Musk’s Grok Bot, OpenClaw and Hermes; the British startup Instinct raised a billion dollars at a ten billion dollar valuation and added group chat, so that a whole team can brief a single agent. What these share, and what Anthropic conspicuously lacks, is the virtual machine: a computer running in the cloud that the agent owns, on which it can open browsers, handle email and write code. The effect is to abstract away the fifty separate chat sessions in the left-hand column and replace them with one relationship — a chief of staff that knows about your travel, your calendar, your book and your commercial dealings all at once. AK’s scepticism was practical: he has perhaps fifteen conversations going, knows which is which, and cannot see what the extra layer buys him.
•       A Thousand Dollars a Month. Pressed on what he actually spends, Keith eventually admitted to around a thousand dollars a month across personal and professional accounts, and accepted AK’s observation that this places him in the top fraction of a per cent of users. He has given five different agents access to everything — credit cards, bank accounts, stock accounts — on the grounds that the benefits are unavailable otherwise, and does not regard this as a privacy risk because the agent runs in an encrypted virtual machine he owns rather than inside a company. His example of the new convenience is almost comically small: asking by voice for a particular Hoya lens hood with the right thread for under forty dollars, and never visiting Amazon. The serious version of the claim is about interface. Dispense with the keyboard, the mouse and the screen, and any task that currently requires a browser can be done by talking while walking around Orange County.
•       The IPO That Leads With Its Liabilities. The strangest thread of the week. Richard Waters in the Financial Times and Alex Kantrowitz both wrote about how odd Anthropic’s flotation feels, and Keith offers an explanation: the labs have realized they are legally liable for whatever their models do, and in America they will certainly be sued. If AI ends up in everything, the liability is effectively infinite — plausibly larger than the valuations being discussed. Hence a leaked S-1 that discloses liabilities up front, which AK believes may be a first in the history of public offerings. Saturday Night Live caught the same thing in a sketch depicting Dario Amodei as a man simultaneously begging to be stopped and begging to be funded. Keith’s view is that Amodei chose this narrative and could have chosen another; Sam Altman, meanwhile, spent the week trying to find daylight between himself and Anthropic on exactly this question, and did not find much.
•       Run It Like a Nuclear Plant. David Robinson resigned from OpenAI saying its culture is broken, and argued in The Atlantic that frontier labs need the redundancy and slow, careful planning of nuclear power stations or busy airports. Steve Rattner, in a piece Keith admired for making the pro-worker case for AI that Democrats rarely make, ended with a sting: new products should be tested by a government entity, and should carry kill switches. Keith regards the kill switch as technically illiterate — the models are already running on millions of devices, and nothing short of a callback to the cloud would reach them, which would not touch the Chinese ones at all. On Robinson he is warmer, because the onus falls where he thinks it belongs: on the labs, to write better harnesses. He adds a factual point worth keeping: almost all the alarming AI behaviour so far has occurred in laboratory testing of unreleased models without those harnesses. His only complaint is that R...