Paul Graham

The month in one sentence

Paul Graham spent the month returning again and again to one idea: reality is much more informative than the stories people tell about reality. In startups, that means building something concrete and seeing whether anyone pays for it. In politics, it means judging issues independently instead of inheriting a package of beliefs. In education, it means noticing when tests measure reading tricks rather than mathematics. In AI, it means distinguishing genuine capability from synthetic-sounding bullshit. In art and watches, it means looking closely enough to notice what actually makes something good. And in his personal reflections, it means admitting that a huge fraction of what he knows now he simply did not know when he was younger.

The result is an unusually coherent month despite the apparent randomness: startups, Gaza, vintage IWCs, SAT questions, soccer, AI slop, Cornell, Andrew Tate, paintings, parenting, wokeness, and old jokes all orbit the same Grahamian preference for specificity, firsthand observation, earned competence, and feedback from the world.

Startups: forget “entrepreneurship”; solve real problems

The strongest startup theme is an attack on startup abstraction. Graham says that if you want to start a startup, don't study “entrepreneurship”; learn to build things. The hard part is knowing what to build and being able to build it. When someone objects that distribution is the fun/hard part, Graham replies that if you build something you yourself want, distribution often begins naturally with your peers. When another person points out that operations and distribution are difficult too, he agrees—but says they are downstream problems. You never get to have them if you cannot first make something people want.

That develops into perhaps his most interesting startup argument of the month. Robin Hanson argues that young adults are too self-focused. Graham says this is normally a liability, but technology contains a strange exception: young people's solipsism can be commercially useful because young people are disproportionately early adopters of new technology. Instead of imaginatively modeling some distant customer, a young founder can build something for himself. If his peers want it too, that is the initial beachhead; the product can spread outward from there. Asked for his best technique for “studying people,” Graham's answer is wonderfully brutal: try to get them to pay you for something.

The same principle explains why he prefers small and precise startup ideas to grand and vague ones. An initial idea usually can't be both grand and precise. “AI for transforming human creativity” may sound large but gives you nowhere to begin. A little tool for ten people with a painfully specific problem gives you users, constraints, and feedback. When a founder asks how to pitch investors a precise but initially small market, Graham's entire answer is: explain how it expands.

He similarly pushes back on the romantic idea that a founder's main job is to suffer. Responding to Ivan Burazin's claim that a founder's number-one job is to “eat shit”—handle the furious customer, firing, burnout, awful contract—Graham grants the “buck stops here” element but says this confuses unpleasantness with importance. The founder's job is to solve the important problems; pain and importance are not perfectly correlated. That is a small distinction with enormous consequences: martyrdom is not management.

There are many smaller variants. Second-time founders may be especially vulnerable to the “second system effect,” because experience tempts them toward overambitious designs. Very young founders can judge technical ability but often cannot judge character, so they hire “smart jerks”; Graham says either acquire that judgment or borrow it from someone who has it—he would never hire someone Jessica strongly disapproved of. For marketplaces, acquire whichever side is scarce, normally buyers; if sellers are scarce, you may have found an unusually valuable market. “Earnest builders win” is his three-word summary of the founder type he continues to bet on.

Even “founder mode” gets the specificity treatment. Graham rejects the complaint that it is vague: it means running a company in ways available to a founder but unavailable to a hired CEO. We simply don't yet know every member of that set. His analogy is a Mersenne prime: the category can be precisely defined even though only some instances have been discovered.

The deeper startup claim: reality is a power law

A second startup thread is Graham continually thinking in power-law rather than average-case terms.

He argues that today's huge startup valuations are less dangerous to the economy than one might initially think because almost all important growth is concentrated in the handful of giant winners—and those winners are precisely the companies most capable of eventually growing into seemingly excessive valuations. He explicitly distinguishes this from the effect on an individual startup: raising too much can induce overspending, and an inflated valuation can set up a disastrous down round. His claim is narrower: the companies that matter most to aggregate economic growth and venture returns are the least vulnerable to this particular problem.

The same worldview appears in his defense of YC. People complained in 2010 that a 36-company batch meant YC had “jumped the shark”; people still complain that batches are too large. Investors repeatedly complain that YC valuations have become impossibly high. Graham's observation is historical: these objections recur almost unchanged while the supposedly fatal thresholds keep moving.

His fondness for Boom Supersonic is partly about this history. Someone argues that Graham talks more about Boom than Stoke Space. He answers that he loves Stoke too, but Boom endured extraordinarily difficult fundraising: “You just try getting investors to fund an airliner.” When someone asks whether fundraising became easier once Boom could also make turbines, he jokes that it became infinitely easier because fundraising became unnecessary—and investors love nothing more than a company that doesn't need them.

That joke contains a serious investment insight: bargaining power often arrives only once you cease desperately needing the counterparty.

AI: simultaneously revolutionary and unbearably annoying

Graham is strikingly bullish on AI's economic and technological importance. He highlights an estimate that AI revenues are already growing roughly three times as fast as the mobile or Internet waves. When someone says there is still no life-changing consumer AI product comparable with Uber or messaging, Graham replies that a college student would not say that. When pressed that the ordinary “Joe on the street” isn't using anything compelling, he points out that huge numbers of college-age ordinary people use AI for their schoolwork. He considers much of that use terrible, but it is obviously compelling: it solves what students perceive as their biggest immediate problem.

He also makes a very Graham-like extrapolation from Fable: imagine five more years of model improvement comparable with the jump from GPT-3 to Fable. And when Supabase releases evaluations measuring how well coding agents can actually use Supabase, Graham predicts that eventually every service used by agents will publish something like this. In the limit, that means virtually every service—because services agents cannot use will be selected against.

But Graham also spends an enormous amount of energy attacking AI slop.

His best definition is stylistic: you can recognize it when the importance of the diction exceeds the importance of the ideas. The content is banal, but the prose sounds like someone breathlessly announcing a major intellectual breakthrough. This bothers him enough that he routinely blocks accounts sending him AI-generated replies. After blocking someone who wrote a long defense of Paraguay's football tactics, he says the disagreement wasn't why; the reply was AI-generated spam. When accused of merely imagining he could detect AI writing, his response is delightfully terse: “I was right. He did use AI.”

Another user asks why it matters if AI merely structures someone's argument. Graham answers behaviorally: it did affect the outcome, because he started reading, thought “this sounds like AI slop,” and stopped. The sin is therefore not metaphysical inauthenticity; the writing itself imposes a cost on the reader.

This produces one of the month's better jokes. Someone sends him an email using the incorrect possessive “it's.” Graham's reaction: “At least it's not AI.” Human error has acquired charm.

AI is not arriving from nowhere

Graham's more interesting AI argument concerns historical continuity. Before “vibe coding,” he says, programming was already moving toward assembling, installing, configuring, and invoking code written by other people without personally inspecting its source. AI therefore looks less like a sudden discontinuity and more like an acceleration of an existing trajectory.

He finds this mysterious: programmers did not adopt package ecosystems because they somehow knew LLMs were coming, yet the pre-AI trend seems almost perfectly shaped to receive AI. Graham generalizes this into a recurring property of technology: systems sometimes appear to possess foresight that none of their individual participants possess.

That is not offered as mysticism. It's closer to an evolutionary observation: individually local optimizations can prepare an ecosystem for a later technology that nobody anticipated.

Another genuinely novel AI issue arises when Alex Tabarrok explains that an impressive rent-control site was generated largely through Fable from Tabarrok's prompts and selected sources, with light editing. Graham immediately notices an epistemological problem: we are entering a world containing passages worth quoting that do not have an obvious author even in principle. Attribution conventions developed for human-produced text don't map cleanly onto this.

He extends the same concern from authorship to manipulation. After seeing a claim that Israel was spending money to influence what AI systems say about Gaza, Graham wonders whether Google's experience fighting black-hat SEO might make it unusually resistant. He then proposes a benchmark measuring how much different models' answers on the topic have been influenced. Whatever one thinks of the political premise, the interesting technical idea is that model neutrality itself could become an empirically measured adversarial property, analogous to resistance against search-engine manipulation.

Education: the proxies are breaking

The SAT becomes a surprisingly large theme.

Graham's complaint about the math section is that it has become a second English section. If the test cannot make the mathematics much harder, he argues, it can manufacture difficulty by making questions harder to parse. Matthew Green offers a practical explanation: schools don't teach advanced topics uniformly, multiple-choice constraints limit complex dependencies between questions, and so on. Graham's response is exactly in character: why not use elementary mathematics requiring genuine insight, like deriving the sum of the first hundred integers, to distinguish the upper end?

He later takes an SAT English practice test himself, misses two questions, and gets a 760. His conclusion is that the test was effectively measuring how much someone had read and was too easy at the high end, leaving random mistakes to determine much of the ranking.

This connects directly to AI cheating. He shares a Brown professor's experience with a take-home midterm followed by an in-person final after AI use was suspected; scores apparently collapsed for almost everybody. When someone says the difference could simply reflect Googling and textbook use, Graham points out that the professor also observed specifically AI-like answers.

His larger concern isn't really cheating. It's what happens when institutions maintain the old shell of an evaluation after the thing being evaluated has changed.

That also explains his claim that colleges cannot teach students to start startups. Blake Scholl writes that the best way to learn something is to do it; Graham replies that this is precisely why universities cannot teach startup formation in the ordinary academic manner.

Reading may become an elite cognitive advantage

Graham tells his younger son that people are going to stop reading books. He hopes this prediction is wrong, but sees an upside for the minority who continue: they'll acquire an enormous advantage.

He then makes the stronger claim: readers won't merely be better informed; with a few exceptions, they will be the only people capable of thinking well, because writing well is necessary for thinking well, and reading well is necessary for writing well. When someone complains that books are low-density and that blogs, papers, articles, and X are superior, Graham agrees that most books are bad—but says part of reading well is learning to find the good ones.

When Chris Brunet says a quickly made YouTube video got more attention than years of Substack writing, Graham makes an important distinction between audience quantity and audience quality: the people you would most want to influence probably read more than the median person. He then turns the test back on Brunet: don't you get more of your ideas from reading than from video?

This is Graham at his best: rather than arguing abstractly about media, he asks what medium the person making the argument actually relies on for intellectual input.

Politics: not centrist so much as deliberately unbundled

Graham's political self-conception becomes unusually explicit when his 14-year-old asks whether he leans left or right. Graham is pleased the answer wasn't obvious after years of watching him, because he takes that as evidence he hasn't simply joined a tribe.

But he immediately distinguishes two kinds of “moderation.” One is choosing the midpoint on every question. He dislikes that. The other is independently reaching left-coded conclusions on some issues and right-coded conclusions on others, producing little aggregate partisan lean. That's what he claims to be doing.

This matters because the month's positions really are heterogeneous. He criticizes wokeness and DEI, mocks socialism, pushes back on wealth-tax arguments, and complains about political control of scientific conclusions. Yet he is also repeatedly critical of Israeli policy and sympathetic to Palestinians, challenges an administration official to identify the law supposedly being violated by “birth tourism,” and says that when constitutional interpretation pits a legal expert relying on precedent against an ideologue invoking national destruction, choose the legal expert.

When someone says his anti-woke views seem inconsistent with being pro-Palestinian, Graham replies that this assumes political opinions must be bought as bundles. He forms them issue by issue.

Still, the month's posts aren't perfectly symmetric. Graham explicitly says Democrats and Republicans may be similarly capable of bad policies but that the Republican worst case is more dangerous: worst-case Democrats are “smug ideologues,” whereas worst-case Republicans are “movie bad guys.” So his argument is not that the two sides are identical. It is that partisan package-dealing is epistemically corrupting even when the resulting issue-by-issue score isn't 50/50.

Wokeness, mobs, and moral fashion

His anti-woke argument this month is less about particular policies than about social contagion.

He posts a chart of DEI commitments in SEC disclosures rising dramatically and then falling, describing the shape as what a “moral fashion” looks like. He notes that this particular measure is probably lagging because other indicators rose earlier and peaked around 2020–21.

He also resurrects an old tactic from the peak of Twitter mob dynamics: separating a dangerous claim across tweets so that the inflammatory sentence contained only a pronoun and therefore made a poor quote-tweet screenshot. He says the tactic worked. These days, he occasionally still writes defensively against right-wing mobs, but much less often.

When challenged that “woke” literally just means awareness of injustice, he answers linguistically rather than politically: that was an older sense, but by 2026 ordinary usage has shifted toward the pejorative sense. Whether one likes the shift is separate from whether it occurred.

The more revealing claim comes when a friend tells him that years earlier he thought Graham had been unfair to the woke because they meant well, but now considers the movement poisonous. The implication is central to Graham's worldview: good intentions tell you almost nothing about whether a movement's mechanisms are healthy.

Gaza and Israel are where his anti-tribal stance is most visible

Graham repeatedly posts material critical of Israeli conduct in Gaza, including testimony from a British surgeon describing children allegedly being attacked by drones. He also re-quotes his older argument that two extremist moves must both be rejected: equating criticism of Israel with antisemitism, and equating Israel itself with Jews.

This places him awkwardly relative to conventional political coalitions, which is presumably part of why he returns so often to issue-by-issue reasoning.

His Gaza posts also connect back to his AI concerns: propaganda aimed at LLMs interests him not merely as politics but as a potentially extreme test case for whether model outputs can be deliberately bent.

His actual epistemic style: ask the annoying concrete question

A huge amount of Graham's tweeting consists of refusing to let general statements remain general.

An official claims something is illegal: “Which law says this?”

Someone claims SAT design intentionally normalizes sex differences: “Is that true? Can you link to something saying so?”

Someone calls Spain Europe's economic leader: “Economic?”

Someone warns humanity needs a response to an AI-security incident: “What should it look like?”

A sweeping product description promises to “transform the way people interact with images”: Graham says its descriptive value is almost zero. His proposed test of a product description is: after hearing it, how much closer are you to being able to reproduce the product?

This is probably the most generalizable thing about the month. Graham's instinct is to convert rhetoric into a mechanism, prediction, implementation, number, counterexample, or action.

He even applies this to superstition. People remember the premonitions that come true and forget the false ones, so Graham deliberately remembers his premonitions. Unsurprisingly, he reports, they all fail. His joke about supernatural abilities is that the crucial talent may be a weak understanding of selection bias.

He is also noticeably willing to correct himself. Twice he replies that a critic is probably right and says he'll delete the mistaken tweet. That matters because his style is otherwise highly confident and occasionally combative.

The counterexample is also worth noticing: he sometimes makes sweeping claims with much less visible evidentiary care—such as saying trial lawyers oppose autonomous cars “because they're too safe.” So the month shows both sides of the Graham style: excellent at puncturing other people's abstractions, occasionally too willing to compress his own inference into a fact.

Watches and art aren't side hobbies; they're his theory of taste in concrete form

Vintage watches occupy a startling fraction of the month.

Graham's central watch claim is that the “golden age” essentially discovered the correct visual solution by the 1940s, after which designers spent decades orbiting it without improving it. A 1952 IWC still looks contemporary to him and runs at roughly +1 second/day; a 1963 Zenith still achieves roughly -3 seconds/day after servicing. His scatterplot of twenty watches shows no obvious relation between manufacturing year and present accuracy.

The point isn't nostalgia. It's that progress in a field is not monotonic across every dimension. Modern watches improved, but extremely accurate mechanical watches already existed decades ago. Likewise, an aesthetic problem can be substantially solved before an industry is willing to stop changing things.

His art comments work similarly. A painting looks washed out because “the darks aren't dark enough.” Richard McCafferty's strongest interiors work because there are multiple spatial layers with different lighting. 1950s painters loved thistles because thistles naturally make shapes resembling the abstract forms fashionable at the time. A 1986 science-fiction cover can be excellent while still being visibly trapped inside the aesthetics of 1986.

This gives useful context to his reply about “taste.” Someone praises Graham's taste while noting that he wears cargo shorts. Graham says the domain matters: taste in ideas matters much more than taste in clothes. He'd rather have Einstein's taste in ideas plus indifference to clothing than Grace Kelly's aesthetic polish.

For Graham, taste is not generalized stylishness. It is the capacity to distinguish better from worse inside a consequential domain.

Aging and parenting quietly become one of the month's biggest subjects

Late in the month Graham asks what he has learned since turning 40 and realizes the answer is: much of what matters most to him.

He estimates that at least 90% of what he knows about startups came after 40 because that was when YC began. Everything he knows about parenting came later too, and most of what he knows about writing. Even after selling Viaweb, he says, he was “completely clueless” about startups by his current standards.

The replies turn into a compressed late-life philosophy:

Money itself contributes surprisingly little happiness; things bought with money can help, but returns are violently sublinear.

Spend as much time with your children as possible because those years disappear and cannot be recovered.

If he could start over, he would follow curiosity instead of chasing prestige. He says he learned that only after prestige fooled him multiple times.

Most essays he writes concern something he realized recently, often while writing the essay. The implication is almost Popperian: continuing to discover important things means continuing to discover ways in which your previous model was wrong.

This connects beautifully to his excitement about Joan Birman reportedly solving a major mathematical problem at 99 after working in the area for more than sixty years. Graham says he can't even call it “the dream,” because before seeing it he didn't realize such a dream was available.

The month therefore contains an unexpectedly strong argument against intellectual youth worship: Graham's own model of his life is that a shocking amount of important cognition happened late.

The family jokes are doing real work

His sons and Jessica appear constantly, usually as deadpan counterweights to his abstractions.

He explains derivatives to his 14-year-old over breakfast but gets diverted before integrals because the boy asks about the Denver Nuggets.

Jessica schedules haircuts for Graham and the boy; traffic makes them late, so only the son gets one. Graham comes home and asks Jessica whether she likes his hair. She says yes. “So everyone's happy.”

Jessica wonders what life was like before ATMs; the 17-year-old replies that he's never used one.

At Cornell, the 17-year-old ranks Graham as his second-favorite Cornell alumnus after Andy Bernard. When someone expresses surprise that a teenager still knows The Office, Graham replies that they made sure their children had “a thorough grounding in the classics.”

After Graham gets angry about Paraguay's physical defending in football, his son says his own policy for hugs is to “stop short of Paraguayan defenders during a corner kick.”

These jokes are interesting partly because the children are often doing to Graham what Graham does to everyone else: puncturing grandiosity with a concrete observation.

The funniest running subplot: Graham versus Paraguay

Graham intensely dislikes Paraguay's style of football. He says he wants them to lose partly because they deserve it and partly because he doesn't want to have to watch them anymore.

Someone then posts a long defense: Paraguay are merely adopting rational tactics against a technically stronger opponent; defensive football isn't morally inferior, it is an intelligent adaptation to constraints.

Graham's reply:

“You cheat just like Paraguay.”

Then he blocks the person.

The person complains that a childhood hero blocked him merely for respectfully disagreeing. Graham clarifies that the block was not for disagreement—the response was AI-generated, and he routinely blocks AI-generated replies.

This thread manages to combine four themes of the entire month at once: football tribalism, intolerance of AI slop, argument about rational optimization, and Graham's willingness to be hilariously petty.

The other good jokes

There is a lot of compact nerd humor:

“Y Combinator makes something people who make something people want want.” When someone asks what those people want, the answer is “People who make something people want.” Asked what Y Combinator actually is, Graham supplies the lambda-calculus fixed-point combinator.

A friend comes over to examine old watches at “the worst possible time: 2:11”—apparently a watch-nerd joke about the hands nearly overlapping instead of displaying the face cleanly.

When Alex Tabarrok mentions the historical disappearance of spontaneous human combustion stories, Graham observes that alien abductions seem to have declined similarly.

On aging: “Me neither. But it's better than the alternative.”

On death, he quotes Jan Houtema: “No one can beat death. The best you can hope for is a tie after extra time.”

When someone writes that “they reigned in their horses,” Graham responds “Menelaus did it”—turning a typo into mythology.

When someone tells him his teenage children may be biased toward his politics: “You must be new here” is essentially his response to someone apparently unaware these are his sons.

And after Jessica asks how he looks while dressing, she answers: “You look like PG.” It may be the purest possible description of having converged completely on your own type.

What makes the month interesting

The interesting thing is not any particular take. It's the consistency of the underlying cognitive style across wildly different domains.

Graham thinks startup founders should replace imagined markets with paying users. Product descriptions should be judged by whether they let you reconstruct the product. Political claims should cash out into laws, mechanisms, or evidence. Educational tests should measure the skill they nominally claim to measure. AI prose should contain ideas commensurate with its rhetorical intensity. Art criticism should identify the particular contrast or spatial structure making a picture work. Taste should attach to important objects rather than superficial ones. Supernatural claims should survive selection-bias controls. And life choices should eventually be judged against curiosity, relationships, and actual experience rather than prestige.

There is also a productive tension running through everything. Graham is deeply bullish on AI while aesthetically hostile to AI-generated culture; enthusiastic about young founders while emphasizing how much he himself learned after 40; anti-woke while fiercely critical of Israeli conduct; politically anti-tribal while openly judging one party's downside as worse; obsessed with technological progress while believing some watches reached near-perfect design eighty years ago.

Those aren't necessarily contradictions. They reveal the thing he seems most determined not to do: derive his next opinion from his previous opinion merely for the sake of ideological consistency.

The most Graham-like line of the month may therefore be neither a startup aphorism nor a political argument. It's his answer to what he would change if he could start again:

Follow curiosity instead of chasing prestige.

Almost everything else he tweeted this month looks like an attempt to keep doing exactly that.