Paul Graham

Paul Graham’s month of tweeting is much more coherent than the surface variety suggests. He jumps from YC office hours to vintage watches, Trump, AI-generated prose, school exams, immigration, Jessica, 17th-century painting, and teenage social-media circumvention. But the same few instincts keep reappearing: look at mechanisms instead of slogans; distinguish things that sound similar but are causally different; distrust fashionable explanations; learn by building; and prefer the person who can actually make something work over the person with the best theory about why it should work.

The most interesting thing about the month is that AI has clearly become the background condition for almost everything he thinks about. But he is neither conventionally bullish nor conventionally alarmist. He thinks AI is extraordinarily powerful, already far enough along that the old concept of a single AGI threshold is misleading, and likely to remake companies, education, software, and writing. At the same time, he repeatedly argues that people are drawing the wrong conclusions from this: software is not simply dead, human founders are not irrelevant, and delegating your thinking to an LLM may be one of the stupidest possible uses of an otherwise extraordinary technology.

Startups: almost everything comes back to the founders

The dominant subject is still startups, and Graham’s position has become almost aggressively founder-centric.

A 16-year-old is told that building yet another database is a waste of youth; Graham replies that the kid should work on whatever is genuinely interesting, even if thousands of versions already exist. Bands do the same thing. At sixteen, duplication is part of learning. When someone asks whether there is an age limit to this, his answer is essentially no: building things is a way of learning, and learning does not expire.

That principle scales all the way up to YC investing. Asked what categories interest him, he says the category is almost irrelevant: he wants “relentlessly resourceful founders.” Asked what YC looks for in biotech, he gives nearly the same answer it gives elsewhere: expertise plus an idea that the domain experts themselves find intriguing before outsiders understand it. When investors become excited about hard tech because they fear AI will destroy software, Graham thinks they are committing the classic error of investing in a category instead of people. Hard-tech founders may indeed have an advantage, but partly because deciding to build a nuclear reactor is unusually strong evidence of commitment.

His recurring startup slogan gets sharpened too. Startups usually do not die because competitors kill them. They die because the founders fail to make something people want. Someone asks what “poor execution” means: “Building something no one wants.” Someone asks why the YC slogan is make something people want rather than need: because this is business advice, and people buy what they want. Someone working part-time asks what to do: ship and get real users. Someone devastated by five incubator rejections gets the same prescription: launch; if users accept you, the incubators cease to matter.

This is one of the month’s strongest themes: external judgment is downstream of reality, and reality is users.

Investors in particular get demoted from oracles to noisy measuring instruments. Graham says the same investor will judge the same idea differently depending on whether the startup is already a hot deal and whether the founder’s manner impresses them. This does not mean those signals contain zero information; personality predicts outcomes, and investor enthusiasm itself contains some signal. The error is allowing those signals to contaminate the supposedly independent judgment of the idea. His suggested weighting for investor opinions is approximately “fortune cookie.”

There is a nice companion argument about startup valuations. If success would make a company worth $10 billion and the current cap is $25 million, the price implies roughly a 1-in-400 probability before correcting for dilution. Founders can think “we only have a 10% chance of succeeding” and feel embarrassed, when a 10% probability at that price would make the investment absurdly attractive. He says he saw one whose implied probability was around 1-in-10,000. The point is psychological as much as financial: venture returns are driven by payoff asymmetry, so a low absolute probability of success is not remotely the same thing as a bad bet.

And his favored observable metric is wonderfully concrete: across the life of a technology company, from YC batch to multibillion-dollar public company, he says the best predictor is how quickly it keeps shipping new things.

AI changes the startup landscape, but not in the simplistic way investors think

Graham has added a new YC office-hours question alongside network effects and going full-stack: can this company be made AI-proof?

His most interesting answer is that the protection often comes from building something that intelligence itself cannot simply replicate. A marketplace is not defensible because humans are too stupid to compete with it; it is defensible because its value lives in the network of participants. Therefore making the participants AIs does not eliminate the advantage. Better still, make the product useful to agents and let agents interact with one another. Patents are another example: an AI can invent the idea, but that does not necessarily confer the legal ability to implement it.

There is one giant exception: if a single model company comes to own everything, “all bets are off.” He explicitly interprets the giant valuations of model companies as bets on something like that outcome.

But he dislikes the investor question “What if the model companies do this?” because he sees it as the modern descendant of “What if Google does this?” and, before that, Microsoft. The problem is not that model companies cannot kill startups. It is that predicting which businesses they will kill is too difficult for the possibility to function as a useful screening criterion.

He similarly rejects “software is dead.” Some existing software companies may be dead. That is different. If incumbents cannot earn a return on LLM token expenditure, that is exactly what one might expect early in a technological transition: established companies misuse the new technology while new companies designed around it replace them.

His picture of AI-native companies is strikingly small. AI may let companies grow much further before crossing the roughly 10-person and 150-person thresholds where group coordination starts hurting productivity. When someone describes a Series A company with 45 mostly engineering employees and claims the winning pattern is now tiny teams amplified by agents, Graham’s immediate question is: perhaps the answer is eight people — but eight engineers?

He is also fascinated that the next computing form factor remains unresolved. Phone? Glasses? Pendant? Something else? We know there will be one, and afterward it will seem obvious. His practical search heuristic is excellent: look for the thing that is primitive but already useful, as early mobile phones were before smartphones.

His sharpest AI position is actually about writing

Graham is much more hostile to AI writing than to AI itself.

He notices that AI-generated replies tend to use a stereotyped rhetorical opposition between two things. He hates AI-written email because the sender can now impose paragraphs of plausible-sounding material on the recipient at nearly zero cost. Someone proposes countering by having AI summarize those emails. Graham’s solution is shorter: archive them.

This connects to his theory of writing. Clear writing starts with having sufficiently high standards for clarity that an unclear sentence bothers you. You ask what you actually meant and rewrite it. He says he uses essentially the same process in YC office hours: founders arrive with a muddy description of what they plan to do, and clarifying the language clarifies not merely what they should say, but what they should do.

“Compression is understanding” is probably the intellectual center of this cluster. Graham says he tries to make his prose “unsummarizable”: not obscure, and not empty, but compressed enough that removing words removes ideas. A reply attempts to compress his claim further; he points out that there is another implicit constraint — the resulting prose must not be awful. Writing should not become a puzzle either.

That explains his objection to LLM writing. He does not regard writing as a cosmetic packaging stage that follows thinking. When someone distinguishes “the writing itself” from the ideas in an academic paper, Graham says there is no such clean distinction. His fear is not merely that AI prose is stylistically bland. It is that when you outsource the writing, you may outsource part of the thinking while continuing to take credit for the resulting ideas.

He says he uses AI more than Google for research, but never lets it write or edit a sentence for him. His analogy is physical exercise: refusing AI writing will not be like refusing Google Maps; it will be like continuing to run and lift weights after machines have made transportation and mechanical lifting unnecessary. People who care about thinking will continue doing the mental exercise themselves. He even predicts that writing one’s own prose will become prestigious — and therefore something people falsely claim to do.

Education is where that AI problem becomes unavoidable

He imagines the limit case of professors and students both using AI: the student’s AI writes the assignment, the professor’s AI grades it, and the humans merely shuttle machine output back and forth. A compiler would optimize the humans away as dead code.

His proposed response is unusually binary. Schools should have some activities where AI use is expected, perhaps required, and others where it is banned. No mushy middle. For every assignment, students should know which regime applies. And where AI is banned, the ban will generally have to be enforced through changed formats — supervised writing, presentations, or similar methods — because otherwise many students will simply use it.

He thinks future AI detectors could also create retrospective academic scandals. Today’s generated prose may become obvious to much stronger future models. Academia is unusually exposed because its output is permanently published and productivity is measured through publication. He expects both scientific and humanities papers to get caught, though he gives deliberately different and rather caustic reasons for each.

His non-AI education advice is characteristically practical. A teenager whose attention has been destroyed by social media should not begin with heroic discipline and difficult canonical literature; start with books that are simply fun, build the reading habit, and move upward. For exams, once preparation is sufficient, switch to “game management”: sleep, stay calm, manage time, and eliminate stupid mistakes. For hard math-test questions, look for the trick: unlike reality, a well-designed test is not supposed to give you an impossible problem, so apparent impossibility itself contains information.

Hard tech is his favorite evidence that the future does not arrive by fashion

Graham spends a lot of the month celebrating YC’s hard-tech history: fusion, airliners, MRI machines, defense, nuclear reactors.

He gives Sam Altman significant credit. Hard tech was not fashionable among investors in 2014; Altman recruited it because he was personally obsessed with it. Before Graham persuaded him to run YC, Altman was apparently planning to start a nuclear-reactor company. That “weird obsession” gave YC years of experience before hard tech became fashionable. Graham turns this into a general rule for young founders: your apparently random interests are often less random than they look because your own needs can predict future demand. Altman was, in Graham’s formulation, living in the future and noticing what was missing.

The same idea appears when search engines look newly promising precisely because conventional search now seems mature and degraded. Once a field becomes completely unfashionable, everyone stops looking at it, and that can create room to rebuild it in a fundamentally different way. Old ideas in the “trash bin of history” can become good ideas when enabling conditions change.

That is a very Graham pattern: obsolescence can be an opportunity signal rather than a warning.

Politics: he keeps trying to turn moral arguments into causal ones

Politically, the month is combative, but the interesting part is less the particular positions than the argumentative technique.

When someone says Republicans have a “more accurate” rather than merely “lower” opinion of various institutions, Graham immediately separates the two variables: lower means lower; whether the opinion is accurate is a different proposition. When Thomas Massie says Netanyahu “convinced” Trump to start a war, Graham argues that made would be the stronger description if Trump cannot stop it. When someone credits government with creating Silicon Valley because it bought early microelectronics, Graham distinguishes procurement from subsidy: the government ordered difficult products and received chips in exchange. From this he derives an industrial-policy claim: orders for hard-to-build things may create capabilities that subsidies merely get absorbed by.

On the Iran war, he keeps asking a brutally simple counterfactual: is the deal afterward better or worse than what was available beforehand? If worse, then in the relevant sense the United States lost by going to war. He repeatedly asks people who know more than he does to answer that concrete comparison rather than discuss intentions or rhetoric.

On Trump, he is intensely hostile to personalized government, attacks what he sees as degradation of federal institutions, praises government employees who resist improper political pressure, and emphasizes due process, separation of powers, and rule of law. His response to “the Founders could never have imagined Trump” is basically the opposite: they imagined the general type of problem, which is why they built institutional checks against a bad president.

Yet he also dislikes ideological sorting itself. His ideal is the “accidental moderate”: someone whose independent conclusions cause the far left to dismiss him as right-wing and the far right to dismiss him as left-wing. That is not sufficient for correctness — random opinions could achieve it — but he calls it necessary.

Immigration becomes a long argument about selection effects and hypocrisy

One of the month’s longest political threads concerns immigration.

Someone presents figures purporting to show that immigrants from culturally similar countries produce billion-dollar founders at much higher rates. Graham’s first objection is causal: the groups differ in why they migrated. If a higher share of one group came to America specifically to found companies, naturally that group will contain more founders. Comparing raw founder rates without controlling for this selection effect does not establish the claimed cultural explanation.

His second line is intentionally provocative. People advocating a system that admits only exceptionally talented immigrants generally would not qualify under their own proposed standard. Their policy amounts to “immigration is fine as long as people like me are excluded.” He clarifies that his objection is not that countries may never exclude anyone; it is the unexamined asymmetry of imposing a standard on would-be entrants that the advocate himself could not satisfy.

The argument then becomes economic. A critic claims ordinary additional workers merely increase the labor pool while exceptional people create opportunities. Graham answers with the wonderfully compact: roofing companies are started by roofers, not physicists. Ordinary economic activity is generative too; he rejects the implied fixed-lump-of-jobs model.

Wealth, billionaires, and exponential growth

The Oxford talk running through the month is “How to Earn a Billion Dollars,” and it pulls several of his older economic intuitions into Twitter.

He attacks the idea that becoming a billionaire is basically a matter of exploitation by pointing to YC’s actual selection and advice: if exploitation were the secret, YC would have spent twenty years systematically selecting the wrong people and teaching the wrong behavior.

He also emphasizes the shift from inherited to self-made fortunes: among the 100 richest Americans, he says the number whose wealth came through inheritance fell from 60 in 1982 to 27 in 2020. For him this complicates nostalgia for the more equal mid-20th-century economy.

Underlying the billionaire discussion is a more general obsession with exponential growth. Fifteen percent monthly growth sustained for five years produces a number that surprises nearly everyone, including technical founders. His conclusion is epistemic: when dealing with exponential processes, you need extra intellectual humility, because confident declarations that some eventual number is impossible are exceptionally unsafe. The essential quantities are the rate and how long it persists.

Watches are not a distraction; they reveal how he thinks

There is an enormous amount of mechanical-watch material, and it is some of the best stuff in the feed.

He has become obsessed with the “golden age” of mechanical watches and tries to identify peaks rather than merely expensive brands. His candidate is a 35 mm IWC with caliber 853 and dauphine hands. He compares movements by engineering performance, robustness, serviceability, appearance, and historical context; argues that older watches were more modest about branding; likes dateless watches partly because they are cleaner and easier if you rotate among several; and thinks Longines is dramatically undervalued.

His broader aesthetic claim is interesting: modern makers deliberately producing “retro” watches have trouble surpassing watches produced when mechanical horology was still genuinely cutting-edge technology. The original engineers were trying to make the best instrument they could, not trying to reproduce an aesthetic associated with the past.

He also learns publicly. Near the end of the month he realizes that an Omega he had recommended has a repainted dial. He does not delete the old recommendation. He leaves the mistake up as a warning, explains that he had been suspicious, says the auction house reassured him, and concludes that both he and they could have discovered the truth simply by looking at reference images. That small episode is one of the more revealing things in the whole month: his preferred response to being fooled is to preserve the evidence and extract the error-correcting rule.

The most gloriously nerdy watch episode is a 45.5 mm Heuer Calculator whose bezel is literally a circular slide rule. He excitedly explains how it is simultaneously representing every multiplication by 23 and walks through recovering 23 × 17 = 391 from the scale. The month ends with him teaching his fourteen-year-old the same underlying idea: multiplication as addition of logarithms.

The jokes are mostly miniature acts of compression

A lot of Graham’s humor works the same way as his serious writing: a long setup collapses into one precise sentence.

Asked whether scaling a company to unicorn status or getting him into a suit is harder, he says unicorns certainly happen more often.

Asked about YC’s official club tie, he defines it mathematically as the degenerate case of a tie consisting of a single point.

His two actual ties were bought by Jessica. One looks like a primary-school uniform; the other looks like something Trump might wear.

When someone explicitly asks him for $200,000 with no strings attached, proving his point that requests should be clear, he replies that it worked beautifully: it took him no effort to understand the request and say no.

Someone asks why a heavily pregnant friend needs nine pillows. Graham: because of the condition that requires nine pillows.

Someone proposes using AI to summarize AI-bloated email. Graham: archive.

A photograph of galaxy NGC 2217 becomes funny because he initially reads it as NCC-1701, the Enterprise’s registry number.

Someone asks why his chair is not symmetrical; he points out that putting another seat on top would push his head down.

When Jessica says “Dinner’s ready” while he is cooking it, he translates the actual information content: Jessica is ready to eat.

And after teaching a visiting friend Rummikub, Jessica announces that she feels “like a spider who’s caught an insect in its web.”

Jessica is effectively a recurring character. She remembers ancient misdeeds while remaining perfectly polite; worries about an expanding radar screen of children, friends, and acquaintances; goes to bed at 8:27 after noting that nobody criticizes 4:30 a.m. people for sleeping at 8:30; becomes expressionless at cards after Graham points out she has tells; and is probably the constraint preventing him from testing exactly how informally one can address the Oxford Union.

What is actually interesting about the month

The tweets reveal a fairly consistent method of thought beneath the eclecticism.

First, Graham constantly decomposes claims. “Lower” is not “more accurate.” Government purchasing is not subsidy. Want is not need. Investors disliking an idea is not evidence that users will. A low probability of startup success is not a bad investment if the payoff is asymmetric. AI being able to perform intellectual work does not imply that humans should stop exercising the capacities involved.

Second, he likes variables you can actually observe. Growth rate. Return on GPUs. Shipping frequency. User retention. Whether users complain because they care. Whether the postwar deal is better than the prewar deal. Whether someone proposing an immigration threshold would personally clear it. He repeatedly tries to replace qualitative atmospherics with something that could prove a claim wrong.

Third, he treats criticism as useful when it contains information. Angry users may be your best users if their complaints are correct. An auction mistake becomes a preserved warning. He deletes a political cheap shot after a critic persuades him the target is no longer doing the thing that made the criticism relevant. He likes the fact that traders inhabit a culture where changing one’s mind in response to new information is rewarded.

Fourth, he is deeply suspicious of proxies becoming goals. Investors substitute founder manner and deal heat for idea quality. Universities risk becoming prestige systems in which machines do the actual intellectual work. Politics substitutes ideological loyalty for causal reasoning. AI writing substitutes plausible prose for thought. Startup founders watch competitors instead of users. Search engines insert monetization into results until the underlying search product deteriorates.

And fifth, he remains unusually optimistic about individual agency. Teenagers should build things. Founders should launch instead of seeking permission. People can learn technical skills rather than defining themselves as “non-technical.” Old discarded ideas can become new opportunities. AI may radically change the environment, but the response is still to understand the new environment better than everyone else and make something people want.

That is why the watches, paintings, school advice, startup arguments, and political fights do not feel entirely unrelated. The common fascination is with how things actually work — a mechanical movement, an exponential curve, an investor’s judgment, a sentence, an institution, a market, a teenager learning math. And the recurring enemy is anything that prevents you from seeing the mechanism clearly: fashion, ideology, sloppy language, herd judgment, prestige, or AI-generated sludge.