Paul Graham

The month in one sentence

Paul Graham spent May 2026 arguing, across startups, AI, politics, education, cities, watches, art, and parenting, for essentially the same epistemic rule: ignore prestigious labels and ideological stories; look at what actually works, what people actually do, and what survives contact with reality.

That unity is what makes the month interesting. The individual positions are often familiar Graham—make something people want, founders matter, bureaucracy is bad, politics attracts bogusness—but the feed reveals how broadly he applies the same heuristics, and also where he sometimes violates them himself.

Wealth: “make something people want” becomes a political theory

The dominant intellectual fight of the month is over whether very large fortunes can be honestly earned.

It starts when AOC says that nobody can earn a billion dollars: extreme fortunes necessarily involve market power, labor exploitation, rule-breaking, or similar extraction. Graham gives the canonical YC answer: start a company that grows very quickly by making something people want. When AOC replies with Airbnb and shifts to saying billionaires often accumulate wealth through harmful behavior, Graham immediately attacks the quantifier change. Her original claim was impossibility; “often” concedes that honest billionaires can exist. He then deliberately picks Taylor Swift and Beyoncé as cleaner counterexamples.

This turns into a month-long campaign. A friend's startup is growing 93% per month; Graham treats the founder's rapidly appreciating equity as an experiential proof that wealth can arise simply from creating something users value. When someone says a few million and a billion are radically different, he replies: “You should do the math”—at 93% monthly growth they are only about ten months apart. He keeps returning to the same proposition when arguing with Elizabeth Warren's worldview and later when Ro Khanna describes wealth as being “hoarded.” With Khanna he refuses to let the discussion expand into tax rates and stock buybacks: if you create a fast-growing company and your shares become valuable, is that hoarding wealth? Yes or no?

The interesting part is not the libertarian-ish conclusion. It is Graham's argumentative method. He looks for a universal proposition inside political rhetoric, constructs the smallest counterexample needed to falsify it, and then gets irritated when the other person retreats to a fuzzier proposition.

There is also a real weakness in his argument. Graham repeatedly establishes “some fortunes can be created honestly”, which successfully refutes “no billion-dollar fortune can be earned honestly.” But he sometimes writes as though this settles much broader questions about how frequently extreme wealth reflects rent extraction, externalities, regulation, monopoly, or political power. It doesn't. The counterexample destroys the universal claim; it does not establish the distribution.

Still, Graham's deeper model is clear: wealth is not fundamentally a pile of money being allocated among people. A founder's equity can become enormously valuable because the underlying organization becomes enormously valuable. That makes “the rich took the wealth” a category error in at least a large class of cases.

Startups: the patient's pulse, early adopters, and earned secrets

His startup advice this month is unusually concentrated and unusually good.

His default first question to a startup is: what's your growth rate? He calls it the patient's pulse. Before launch there is no pulse, because there is no external signal telling you whether reality likes your theory. Likewise, when someone asks for one question that reveals whether users are happy, his answer is brutally compressed: “Are you willing to pay?” And when investors profess enthusiasm, he tells founders to disregard the speech entirely and look at their hands: are they holding a term sheet? If not, the answer was no.

The strongest startup idea from the month is his reformulation of startup ideas as pairs rather than scalars. An idea is incomplete without its initial users. If a product only works once a large network already exists, it isn't yet a startup idea; you need some identifiable group desperate enough to use the primitive version while almost nobody else does. This is why “make something you yourself want” is such a powerful generator: you automatically know at least one population of early adopters—people like you.

That connects neatly to Garry Tan's observation that the best founders possess some weird, specific, earned insight from living inside a problem. Graham uses it to argue against teenagers rushing to start generic “AI for X” startups. A technically competent 18-year-old may know how to build but not yet know anything strange and valuable enough about the world.

He also pushes against the idea that crowded markets should scare founders. A crowded market can mean almost the opposite: lots of people have verified that the problem is real, while nobody has solved it sufficiently well.

The recurring theme is reality-generated information. Growth, payment, determined early adopters, observations about users, and earned insights all contain information that pitch decks and abstract ideation do not.

College: surprisingly anti-dropout

One of the most interesting reversals of caricature is how strongly Graham argues against dropping out of college to start a company.

His point isn't credentialism. In fact he explicitly says that, if you're doing university correctly, the diploma isn't what you're pursuing. His claim is about option value and intellectual exploration.

A young founder thinks: I have this great idea now; if I wait three years someone else will do it. Graham calls this a fallacy because ideas you haven't had yet don't feel real. But you already know vastly more than you did three years ago, so you should expect future-you to have better ideas too. Likewise, he has heard “all the good startup ideas will be taken soon” for twenty years, and it has never become true.

When someone says you learn far more building a startup than attending university, Graham makes a subtle distinction: a startup teaches you enormous amounts, but you don't control the curriculum. You learn whatever the startup forces you to learn. University lets an intellectually curious person wander speculatively into subjects whose usefulness is not yet apparent.

And when asked why YC increasingly funds very young founders if this is bad advice, he answers that YC is observing the trend, not endorsing it: people are starting younger, which is precisely how he has seen enough of the phenomenon to become worried about it.

This is Graham distinguishing exploration from exploitation without using those words. Eighteen is unusually valuable exploration time. A startup is ferocious exploitation of whatever knowledge and interests you already possess.

AI: maximal enthusiasm, selective distrust

His AI position is much less simple than “AI booster.”

When asked what he would do if he were 21 today, he says he would build his own LLMs, as large as he could make them on whatever hardware he could obtain. When somebody points out the hardware constraint, he says scarcity might actually be beneficial. That is classic Graham: constraints force unusual solutions.

He thinks the valuation of frontier-model companies as eventual trillion-dollar enterprises is at least intelligible because investors are betting that model companies could end up “owning everything.” If models become superintelligent but still need humans to direct them, then “what are the hardest problems?” is effectively already the prompt. He thinks CEOs should personally be knee-deep in building with AI, and predicts that AI-native companies will avoid today's painful problem of extracting tacit organizational knowledge from employees because much of that knowledge will never have existed only in human heads.

Yet he reacts viscerally when Jared Friedman says he quietly gave an AI agent full access to YC's production database: “Literally horrifying.” When challenged, he clarifies that he isn't claiming the decision was wrong; the trust level simply triggers his risk instincts.

And he strongly dislikes AI-written human communication. Founder emails now arrive in a recognizable “hard-hitting journalistic style”; once he detects that the founder outsourced the prose to AI, he tends to stop reading. To him it is not merely aesthetically bad but deceptive: a human signature is being attached to words the purported author didn't write. A critic points out the apparent contradiction—Graham says founders should use AI aggressively, then penalizes them for using it—and Graham's answer is essentially right tool, right layer. Using AI is not itself impressive; knowing where not to use it matters. Translation is fine because it preserves the writer's underlying thoughts rather than replacing them.

This is probably his most coherent AI position: use AI to increase capability, not to counterfeit agency.

The machine should help you think, build, translate, search, calculate, and operate. It should not make another person believe they are encountering you when they are actually encountering a generic language model.

AI in education: neither ban it nor surrender to it

The same distinction appears in schools.

Graham wants deliberately extreme variation: in some situations AI use should be encouraged; in others it should be completely prevented. He assumes students will cheat wherever cheating is technically possible, so an honor-code prohibition is fantasy. This is why he mocks Berkeley Law's blanket AI ban as unenforceable: students will learn to regard rules that cannot actually be enforced as jokes.

There is an implicit pedagogical distinction here between training the student and producing an artifact. If the purpose is to learn unaided reasoning, AI defeats the exercise. If the purpose is to accomplish some external task as effectively as possible, refusing AI is artificial handicap.

That's a much more durable framing than “AI good” versus “AI cheating.”

Politics: he is increasingly anti-tribal, but increasingly political

Politics occupies much more of this month than the stereotypical startup-Paul-Graham feed would suggest.

He remains intensely hostile to the left-wing online mobs of the 2010s. He says those years sometimes felt like a war and that he remained on Twitter partly to demonstrate that he could not be silenced. When someone claims left-wing trolls didn't cause real-world harm, he points to campaigns to get people fired.

But he now says the xenophobic mobs appearing on the right can be even uglier. That becomes especially salient after anti-immigrant replies erupt around changes to US immigration policy. His diagnosis of some xenophobia is almost psychological: unsuccessful people explaining their failure by claiming immigrants stole opportunities that rightly belonged to them. He explicitly compares this to earlier left-wing explanations for demographic disparities.

He condenses the pattern for his 14-year-old into a political archetype: the politician who says your life is bad because an outgroup stole what belongs to you; elect me and I'll take it back. His examples deliberately span Lenin and Hitler. When a reply insists Hitler was really left-wing, Graham answers with one of his favorite forms of empirical check: ask how the actual German left regarded Hitler.

He is also openly hostile to the Trump administration throughout the month. He praises the judiciary for constraining executive behavior, attacks immigration claims from DHS, supports Thomas Massie's willingness to defy Trump, suspects Trump's moves against Massie and Marjorie Taylor Greene imply damaging material in the Epstein files, attacks the investigation of E. Jean Carroll as “prosecute the victim,” and says of the choice between dunces and authoritarians that the current president is both.

This is not a neat partisan realignment. He says he'd vote for AOC over Trump while spending much of the month attacking AOC's economics. He attacks Warren's role in Democratic crypto policy while simultaneously criticizing xenophobic elements of the right. He thinks lawmakers actively trading individual stocks is absurd. He supports immigration partly on the extremely Graham-esque ground that America demonstrably benefits from importing unusually capable people.

The through-line is less left versus right than competence, pluralism, and aversion to political mobs.

The smartphone, social media, and institutions

Marc Andreessen argues that blaming smartphones, the internet, or social media for social deterioration is “cope” that distracts from collapsing institutional competence.

Graham catches the overstatement. A technology this pervasive must have some effects, so an explanation that assigns it approximately zero causal importance cannot be right. Elsewhere, when discussing falling birth rates, Graham proposes one possible mechanism: social-media consumption increases isolation, isolation means fewer couples, fewer couples means fewer children.

This is characteristic Graham reasoning but also exposes one of the month's epistemic tensions. He is excellent at detecting when someone else's universal claim outruns their evidence; he is sometimes much more relaxed about his own causal stories. “Social media → isolation → fewer couples → fewer babies” is plausible, but in this corpus it is asserted, not demonstrated.

The same asymmetry appears in some of his political causal claims. Graham's strongest mode is counterexample and operational test. His weaker mode is macro-causal explanation from a few salient observations.

Garages: a physical theory of innovation

One of the month's most Graham-like mini-essays starts from a photograph of early corporate offices.

He proposes, apparently seriously, that Europe having fewer garages may be a meaningful innovation disadvantage. A garage gives you permissionless space to work on things that “don't matter yet.” The outliers of ideas need outliers of space.

Someone suggests that European public resources and university collaboration might compensate. Graham's answer is ruthlessly empirical: whatever Europe's advantages are, its disadvantages must have outweighed them historically, or Silicon Valley would have emerged there. Blake Scholl suggests regulation and culture are the real cause and garages merely secondary; Graham pushes back that compact European urban form quite literally means less spare private space. That compactness makes cities more walkable, which is good, but there is a tradeoff.

Then somebody tells him that in Germany you can in some circumstances be fined for using a garage for purposes other than parking. Graham initially sounds almost incredulous, verifies it within the thread, and adds it to the argument.

The idea is interesting because it generalizes his startup philosophy into urban form. Innovation depends not merely on universities, capital, or policy, but on cheap, informal slack: places where nobody asks you to justify what you're doing before you know what it is.

An official “Innovation Center” is precisely the wrong substitute because application forms, opening hours, administrators, and restrictions select against weird premature experiments.

Housing: supply is still supply

The same suspicion of elaborate explanations appears when Berkeley's housing construction boom is credited with pushing nominal rents below 2018 levels.

Graham's reaction is basically: why is it so hard for the left to understand that if you want housing to cost less, allow more housing to be built?

This fits the wider month's pattern. When a phenomenon can be explained by a simple mechanism with an observable output, he distrusts moralized or institutionally elaborate alternatives.

“Performance is the ultimate test”

This sentence appears while Graham is discussing 1950s Omega watches, but it may be the hidden motto of the entire feed.

He used to reject bumper automatic movements because they were technically a workaround to a Rolex patent. Then he noticed that Omega's caliber 354 movements can still keep excellent time seventy years later. So he changed his mind. Asked what defines a great vintage watch movement, he answers: how well it keeps time today.

The same structure appears everywhere:

A startup? Growth.

A user's happiness? Payment.

An investor's interest? Term sheet.

A watch movement? Timekeeping seventy years later.

A university? Whether capable people actually learn there, not whether credential discourse says university is good or bad.

An AI workflow? Whether it increases useful capability without destroying the thing you were trying to preserve.

A political claim? Try to formulate it precisely enough that a counterexample could kill it.

This is why the watches, startups, and politics don't feel as disconnected as they first appear.

Well-informed optimism and “barely impossible” things

His best explicitly philosophical tweet of the month is that nothing is more powerful than well-informed optimism.

Not “everything will be fine.” The valuable optimism is: Hmm, what if we tried x?

That pairs with another tweet rejecting “nothing is impossible.” Treating all impossibilities alike would be a terrible search strategy. Instead, distinguish degrees of impossibility and concentrate on things that seem barely impossible, because inspection may reveal that some are possible after all.

This is arguably the deepest Graham theme in the file. Optimism is useful not as an emotional state but as a search heuristic. You need enough knowledge to constrain the search space and enough optimism to keep probing its frontier.

That is also why he likes hobbyists. Successful founders often keep going long after money ceases to motivate them because the company has become their beloved project. The hobbyist motivation keeps curiosity alive after ordinary incentives disappear.

His hatred of bogusness

Graham says that perhaps the biggest thing he and Jessica Livingston have in common is low tolerance for bogusness, but for different reasons.

Jessica is a “social radar”: she sees through people. Graham says he dislikes bogusness because he's trying to find answers and bogusness is noise in the signal.

That explains a surprising amount of his tone.

He dislikes euphemistic university mission statements because extra words create hiding places for bogusness. He dislikes political claims whose quantifiers silently change. He dislikes AI-written personal emails because they misrepresent their source. He dislikes forced website questions enough to enter deliberately false data. He dislikes investors' verbal enthusiasm because it isn't binding. He dislikes grand abstractions about the evils of wealth because concrete counterexamples become hard to classify.

And when someone refuses a precise question, he often just asks it again.

This makes the account occasionally abrasive but unusually legible: you can often predict which sentence in somebody else's tweet will annoy him. It is usually the sentence whose rhetorical confidence exceeds its precision.

Art, books, airplanes, watches, history

A substantial part of the feed is simply a highly curious older nerd noticing things.

He becomes absorbed in 1950s Omega and Longines movements, recommending Japanese dealers on Chrono24 and discussing calibers. His criterion remains engineering performance rather than collector mystique.

He notices that a painting can look “more real” than a photograph and gives a surprisingly good explanation: a painting has passed through a brain that understood the scene, so it contains perceptual cues selected and amplified by understanding. The photograph mechanically records everything; the painter models what matters.

He comments on medieval trans-Alpine traffic, Roman troop numbers, Venetian economic history, military armor, Swedish aircraft, the 747SP's unusually tall tail, obscure safety-pin archaeology, Renaissance iconography, Churchill's astonishing writing output, and antique books.

When asked how he learned enough Christian iconography to look at Renaissance painting, he proposes “lazy evaluation”: research each symbol only when you encounter it in a sufficiently good painting. That is a programmer's answer to art history.

At a rare-book fair he discovers what he calls a selector-cost problem. He has spent decades buying books by subject and assumed expensive books would reveal more interesting possibilities. Instead, when books are selected primarily by monetary value, almost nothing interests him; the three books he wants are among the cheapest. The market's axis of prestige is almost orthogonal to his axis of curiosity.

That tiny observation is very Graham: optimize the selector, not merely the selection.

Family life is the running counterpoint

Jessica and his children are everywhere in the feed, usually as comic foils.

Jessica on the universal managerial fantasy: “If everyone would just listen to me...” Graham adds that she's right.

Jessica on whether Thomas Massie can run for president.

Jessica changing from a fork to a spoon because the dinner Graham cooked was good enough to require recovering every last bit.

Jessica noticing him in a fancy clothing store after the saleswoman asks whether he'd like anything: “Look at him.”

A teenage son observes that Graham's accidentally fashionable Birkenstocks and fleece prove a stopped clock is right twice a day. ChatGPT then rates the outfit as suitable for walking the dog. Graham concludes this may explain why dogs like him.

He describes his younger-family years as “busy but cheerful.” When another parent says raising three young children can be terrible, Graham replies: terrible, but you will miss it.

The family material matters because it prevents the feed from collapsing into a stream of propositions. The politics can be furious, but the baseline life conveyed by the account is domestic, curious, and amused.

The jokes

His humor is mostly dry literalism, pedantry, and taking a premise one step farther than its author expected.

Michael Dell posts an old Byte advertisement and asks people to identify which company became enormous. Graham notices that Dell's old logo apparently contained a grocer's apostrophe.

Marc Andreessen quotes Emerson saying a man is what he thinks about all day. Graham: then parents are their children.

Someone asks whether his em dashes prove he is AI. Graham: “I'm where AI learned it.” Later, asked why his old writing has so many em dashes: “Where do you think ChatGPT learned it?”

Someone asks him for money after arguing with him. He posts the YC application link.

Martin Fowler says the easiest shortcut to riches is being born to rich parents. Graham: then you don't get to be rich until they die.

A photo of extremely young startup founders looks to him like a family picture.

A German watchmaker sees a Cartier case and exclaims “Qvuartz!” Graham frantically assures him it is manual wind.

When his teenage driver is learning, Graham defines success as getting the passenger from frightened to bored. They occasionally achieve it.

The jokes work because they're the same mechanism as his arguments: take words unusually literally and refuse to glide over the hidden premise.

The most interesting tension: Graham the empiricist versus Graham the storyteller

The feed's strongest intellectual habit is falsification. Graham repeatedly asks what concrete observation could distinguish one explanation from another. He notices changed quantifiers. He likes measurable outputs. He changes his mind about watches when seventy years of performance contradicts his aesthetic objection. He tells founders not to believe investors until paper appears.

But he is also an aggressive generator of explanations.

Europe may innovate less partly because it lacks garages. Social media may reduce fertility by making people isolated. Trump's treatment of Massie and Greene makes him infer something serious lurks in the Epstein files. Political corruption may have crossed a threshold because Trump's behavior lowered the equilibrium standard. A global drift toward colorlessness in design will reverse because aesthetic fashions nearly always swing.

Sometimes these are excellent conjectures. Sometimes the corpus supplies almost no test of them.

That tension is productive rather than accidental. Graham appears to think by rapidly generating compact causal theories and then throwing them at the world. His best ones become essays. His weaker ones die in replies. Twitter is effectively his scratchpad for conjecture and criticism.

What changed or sharpened this month

Three developments stand out.

First, AI has moved from being a topic to being an environment. Graham assumes young technical people should build models, assumes CEOs should personally use AI, thinks AI-native organizations will be structurally different, and already judges human behavior partly by how people choose to delegate to models.

Second, he is more explicitly defending institutions that constrain raw political power. The judicial branch, immigration, political pluralism, and people who defy their own party all receive unusually strong praise. The old enemy was the censorious progressive mob; he now seems convinced that an uglier nationalist version is becoming important enough that ignoring it would be dangerous.

Third, he is pushing back against prematurely professionalized youth. Don't rush out of college. Don't start a generic startup merely because AI is booming. Accumulate strange knowledge. Explore things with no immediate payoff. Build your own models partly because limited hardware might force you to learn something. Have the “earned insight” before trying to exploit it.

Those three positions fit together better than they first appear. If AI is rapidly making competent execution cheap, then original judgment, firsthand knowledge, curiosity, and the ability to detect bogusness become more valuable, not less.

The underlying worldview

The month reduces to a handful of Graham invariants.

Reality is richer than ideology. Political theories fail when a single inconvenient founder, immigrant, company, city, or historical case doesn't fit.

Performance beats pedigree. A seventy-year-old watch that keeps time, a founder whose users pay, a company that grows, or a student who actually learns outranks the associated label.

Informal freedom produces weird valuable things. Garages, college exploration, hobby projects, founder autonomy, and small experiments matter precisely because nobody yet knows how to justify them.

Precision is an offensive weapon. Replace “billionaires exploit people” with a yes/no question about a founder's appreciated shares and the argument gets much harder to evade.

AI should amplify agency, not simulate it. Build with it aggressively; don't make it impersonate your thinking.

Optimism is a search algorithm. Don't believe everything is possible. Find the things that are almost impossible and poke them.

And perhaps most characteristically:

If somebody tells you a story about how the world works, ask what you would expect to observe if it were actually true.

That is what connects the founder advice, the politics, the vintage watches, the garages, the AI arguments, and even the jokes.