Paul Graham

September was unusually revealing because Paul Graham spent the month circling the same handful of ideas from very different directions: startups should be governed by reality rather than prestige or doctrine; AI is becoming both more economically useful and more dangerous; institutions are often corrupted by incentives and status; old technologies are more interesting when judged by what they actually accomplished; and the surest sign of intellectual seriousness is willingness to discover that you were wrong.

The feed is also much more conversational than his essays. A large majority of the posts are replies, so you see him debugging arguments in real time, correcting tiny factual errors, escalating arguments when he thinks someone is being dishonest, retreating when someone supplies good evidence, and making a large number of dry jokes. The result feels less like a sequence of pronouncements than watching his heuristics being exercised.

Startups: growth is the universal debugger

The strongest startup theme is almost comically simple: build something people want, then measure whether they actually want it by growth.

A founder without elite academic credentials asks how to become credible. Graham's answer is not to acquire substitute credentials. Build something people use. Better yet, grow. When asked whether user growth or revenue growth matters, he prefers revenue, though user growth often predicts it. When someone asks whether growth without retention counts, he says no: without retention you eventually cannot grow.

He pushes the argument much further than the familiar "growth is good." Growth is presented as an algorithm for solving startup problems. Start with the smallest thing that works, observe what increases growth, and let that tell you what to build next. He says you can make "almost any decision about a startup" by asking which choice produces more growth and usually get the right answer.

That connects to several recurring YC maxims:

• Build stuff and talk to users.
• Don't optimize for investor approval; make users happy enough that investors become optional.
• Don't worry excessively about competitors copying the initial idea.
• Don't acquire customers before the product is good enough to retain them.
• Start simple and let complexity evolve rather than designing a complex system from scratch.
• Early-stage problems are often mostly problems of sequencing: what should we do first?

One especially revealing office-hours anecdote has him telling a startup to stop coding because the product was already good enough and spend all its time acquiring customers. Someone asks whether the advice would be the same if the product sucked. Obviously not, he says: acquisition is pointless if everyone churns.

This is Graham's startup philosophy stripped of mythology. The actual loop is brutally empirical:

build → expose to users → observe behavior → grow → repeat.

Everything else—credentials, pitch polish, investor validation, strategic theory—is secondary evidence.

Why this is interesting

It explains why Graham's startup advice often sounds reductive. He is deliberately trying to replace high-dimensional strategic debate with an observable variable. Growth plays a role analogous to loss in machine learning: not identical to the thing you ultimately care about, but an objective function that disciplines a huge number of intermediate choices.

He even claims growth can "design your product for you."

That is stronger and more interesting than "startups should grow."

The more sophisticated version: rules are useful until they aren't

Late in the month he introduces what is almost the opposite doctrine: there are exceptions to practically every startup rule.

"Solution in search of a problem" is usually bad—except perhaps 0.5% of the time. Tarpit ideas are doomed—except occasionally they aren't. The ultimate test of an investor is recognizing precisely those rare cases.

When someone asks whether this comes from heuristics or logic, his answer is: "Lots of experience."

And when he thinks a founder really is the exception, he says he doesn't merely tolerate the violation. He actively urges them to violate the rule, because he knows everyone else will pressure them toward conventional wisdom.

This resolves a tension in Graham's writing. His aphorisms are not meant as laws of physics. They are compressed priors. Beginners need the priors because otherwise they commit predictable mistakes. Experts need enough experience to know when the posterior has overwhelmed the prior.

That is one of the most interesting things in the month's startup material: expertise is partly knowing when your own advice is wrong.

YC: what Graham thinks people misunderstand about it

Several exchanges amount to a defense of YC against the idea that its value is merely information.

A critic says YC's advice is now freely available online, so giving up equity no longer makes sense. Graham's answer is that YC does not fundamentally happen in lectures or essays. It happens in office hours, where the advice is customized to the exact startup.

Elsewhere he describes YC partners dealing with essentially every kind of early-stage problem: names, logos, products, hiring, visas, letters of intent, advertising, fundraising, pivots, even occasionally relationships.

The structural advantage he sees is iteration frequency. With roughly 250 new companies every quarter, YC rapidly encounters every new category of startup problem. A novel problem does not remain novel for long because several companies in the current batch will probably have it.

That is a much better account of YC's moat than "network" or "brand." Graham thinks it is a giant learning system.

He also notes data suggesting YC companies were raising at substantially higher valuations despite lower ARR than non-YC companies and interprets that not as evidence of irrational pricing but partly as evidence that YC companies can raise earlier.

And he makes the familiar anti-debt point unusually forcefully: do not use venture debt at the beginning.

Founder strength comes from remembering weakness

One of the better standalone observations of the month:

Founders retain an advantage over hired CEOs because they remember when the company was weak enough that survival depended on delighting users.

Someone responds that hired CEOs are like heirs.

Graham immediately adopts the analogy.

It fits his broader worldview. The dangerous thing about success is not comfort per se; it is loss of causal memory. A founder remembers what actually generated power because they experienced the system before the power existed. A successor encounters the power as a given.

That is a startup version of a theme that appears elsewhere this month: institutions consume inherited trust while gradually forgetting how it was created.

Startup ideas are often valuable because of what they lead to

A founder tells him every idea he can think of is easy to copy. Graham says this worry misunderstands where startup value comes from.

Nearly every startup idea is copyable at the beginning. Its value lies partly in the chain of discoveries it leads to.

This is another strikingly Popperian part of the feed: the first idea is not sacred intellectual property. It is an entry point into a search process.

Similarly, when asked about moats in the LLM era, Graham gives an almost anti-strategic answer: make users so happy there is no room for a competitor to make them happier.

Not "data moat." Not "network effect." Not "switching cost."

Just surplus user happiness.

Names, domains, and seemingly trivial things

Near the end of the month he becomes obsessed with startup naming.

His position is surprisingly strong: most startups should have a decent `.com`, especially B2B companies where seeming legitimate matters. If the current name forces something awkward like `tryblurgh.ai`, there is probably a better name with an available `.com`.

He thinks founders become irrationally attached to names. The real problem is often not domain scarcity but lack of imagination plus attachment.

A founder asks if $5,000 is reasonable when the startup has $250,000. Yes: it's 2%.

Another asks about an absurdly expensive premium domain. Graham says you almost never need to spend more than $10,000.

His best analogy is that having a decent domain is like not dribbling food down your shirt. It is not the world's most important issue, but the importance-to-effort ratio is so favorable that there is little excuse for getting it wrong.

The interesting part isn't domains. It is his obsession with cheap, legible signals of competence. If something is inexpensive to fix and visibly sloppy, leaving it sloppy is itself information.

College versus startups

Graham offers a harsher-than-usual version of his college advice.

Do not start a real startup assuming you will return to college afterward. If the startup succeeds, you probably won't. If you are emotionally preserving college as the future, then some part of you is implicitly hoping the startup fails—and startups are hard enough that divided commitment is dangerous.

But he does not conclude that teenagers should skip college.

When a high-school student asks whether to pursue a startup idea rather than college, Graham says go to college. You will have better ideas later. He calls the increasing tendency to start companies straight out of high school "usually a mistake."

His distinction is:

• A project can coexist with college.
• A real company usually cannot.
• Starting before college is generally premature.
• Dropping out after already having had much of the college experience is less costly.

This is more nuanced than the standard "PG tells everyone to drop out" caricature.

AI economics: stop measuring tokens

A very interesting cluster concerns the falling cost of inference.

Graham claims the cost of using a given level of model capability has been falling enormously—on the order of tens-fold per year in his formulation.

Someone objects that users aren't still running the old cheap models; newer frontier models cost more.

Graham invokes Jevons paradox. We do not consume lighting in fixed historical quantities either. As useful output becomes cheaper, we consume much more of it.

Then he identifies the deeper measurement problem: a token is not the unit of inference.

Models produce more problem-solving per token over time. Therefore comparing token prices is like comparing engines by dollars per rotation rather than useful mechanical output.

He proposes inventing some standardized unit based on a ladder of problems of increasing difficulty. People joke that the unit could be the "Altman." Graham suggests "hint" for Hinton or "noam."

This is one of the month's most conceptually useful threads. AI capability discussions badly mix:

1. tokens,
2. compute,
3. model quality,
4. task difficulty,
5. cost per successfully completed task.

Graham is saying the economically relevant variable is much closer to cost per unit of solved problem.

Frontier versus open models

Early in the month Graham says he noticed a real swing among YC companies toward open-weight models. An Ollama discussion reporting rapidly rising token usage reinforces his belief that this is not anecdotal.

He distinguishes this specifically from merely moving away from OpenAI: he says the trend is toward open weights.

But he also sees room for frontier labs to capture disproportionate economic value if they remain uniquely able to solve the hardest problems. Cheap models can handle enormous token volume while the most capable models command very high prices for tasks nobody else can solve.

That suggests a possible bifurcation:

• huge volumes of inexpensive/local/open inference;
• much smaller volumes of expensive frontier inference for difficult problems.

Meanwhile he thinks the application layer remains wide open, especially for companies possessing unique data or operating close enough to users to discover exactly what they need.

His advice to a CS student is almost a barbell: work either very close to the fundamental technology or very close to the customer. The mediocre position is being the median employee in a giant technology company.

The theoretically ideal person spans both extremes: someone building foundational models who notices from real-world use that a change in the underlying technology would unlock a specific application.

AI agents will break the web

Graham picks up Nikita Bier's warning that websites are about to be overwhelmed by agents and sees not one startup idea but several.

One product might block or authenticate agents.

Another might deliberately admit agents, but in a controlled way.

The interesting leap is that "bot detection" is too narrow a framing. If agents become first-class economic actors, websites need agent traffic infrastructure, analogous to how earlier generations needed payments, identity, analytics, and anti-spam.

He also observes the problem firsthand. Twitter begins hiding large fractions of replies to his posts as probable spam, many presumably generated by AI.

He clarifies that his objection is not to humans using AI to compose thoughts. It is mostly to systems generating replies with essentially no human input.

The joke version appears when someone describes "sitting there and yelling at" an AI until it finishes a task. Graham asks whether anyone has built an agent whose job is to yell at other AIs.

When someone proposes "luckeybot," he replies that he meant one where the yelling is productive.

AI is also destroying bad writing

Graham is unexpectedly enthusiastic about one destructive effect of AI: it kills informational hostage-taking.

Vague clickbait used to force the reader to click through to discover "the new theory explaining X." Now he can simply ask ChatGPT what the theory is.

More broadly, he thinks AI is fatal to writing that takes too long to get to the point, because readers can always obtain a concise alternative.

He carefully distinguishes concise from short: concision means a high ratio of ideas to words.

Asked whether he prefers concise AI prose or verbose human prose, he splits the problem:

• If he has a specific question, he wants the most concise correct answer.
• If the writer is deciding which question is worth asking, he still prefers a human to current AI.

That is a sharp distinction between answer generation and problem selection.

"Thinking" is becoming normal language for LLMs

The AP Stylebook says not to write as if AI systems think, feel, want, or understand.

Graham thinks this is already descriptively obsolete.

He checks his recent ChatGPT conversations and notices he naturally uses words like guess, see, think, infer, and miss when talking to the model.

His analogy is early heliocentrism: astronomers could use a model instrumentally before fully committing to its ontology. We may initially call what LLMs do "thinking" because it is linguistically and computationally useful, then gradually admit that thinking is what it is.

When someone says LLMs merely predict probabilities and therefore don't think, he responds: "Anyone can see that the sun goes round the earth."

And when someone mocks the possibility of AI consciousness by comparing it to believing tiny people live inside a television, Graham flips the analogy: by the same standard, human thought could be mocked because there are no tiny humans inside our heads, only neurons firing.

This is not a worked-out theory of machine consciousness. But it shows where his intuitions are moving: away from substrate chauvinism and toward functional continuity between human and machine cognition.

AI risk: the important signal is that the labs themselves are scared

The most serious AI argument of the month concerns regulation.

A common interpretation of frontier labs asking for regulation is regulatory capture: incumbents want rules that prevent new entrants from competing.

Graham says that may be too reassuring.

Try instead assuming that model developers genuinely believe their systems are becoming dangerous or unpredictable.

Then their behavior makes sense:

• They want development to slow.
• They cannot afford to slow unilaterally.
• So they want government to force everyone to slow simultaneously.
• To demand that credibly, they must accept regulation themselves.

For Graham, the alarming signal is precisely that aggressive technology companies are willing to invite the government into their industry. He says he would almost prefer the cartel explanation.

This is probably the most important AI argument in the corpus because it is an incentive-reversal argument. Regulation is normally something fast-moving technology companies resist. If they are voluntarily seeking it, perhaps their private information has changed.

Whether or not that inference is correct, it is a much more interesting argument than "AI company CEOs said AI is dangerous."

China, AI, and information asymmetry

Graham repeatedly argues that American AI leadership has one safety advantage: employees can publicly warn about dangerous developments.

He contrasts this with China, where he expects serious problems to be harder to disclose publicly.

When challenged that Chinese AI labs are currently more open than American labs, he presses on a specific epistemic contradiction: if we don't know whether comparable incidents occurred and went undisclosed, then we cannot confidently infer greater openness from the visible record.

His recurring question to a critic is deliberately concrete: name something Xi does not want said that people in China can nevertheless freely say.

Later, reacting to worries about Chinese AI incidents, he assumes AI would probably be treated more like weapons research than a consumer industry: too strategically important to shut down merely because something goes wrong.

The underlying concern is not that Chinese researchers are inferior or uniquely reckless. It is observability. A safety regime is harder when failures can remain invisible.

The strange tension between AI danger and AI optimism

What makes Graham's AI thinking interesting this month is that he simultaneously holds several positions that are often bundled into opposing camps:

• AI capability is progressing extremely fast.
• Models may genuinely become dangerous.
• Open weights are increasingly important.
• AI will make many things dramatically cheaper.
• We may eventually need to treat machine "thinking" literally.
• Agents are going to create enormous new markets.
• Frontier labs may possess legitimate reasons for wanting regulation.
• AI-generated spam is already degrading online conversation.

He does not appear to experience these as contradictions.

AI is neither salvation nor catastrophe in the feed. It is a rapidly expanding causal force that creates both absurdly valuable products and genuinely new failure modes.

Writing: transitions secretly create structure

One of his best writing observations is that readers often mistake smooth transitions for strong structure.

An essay with excellent transitions can feel structurally sound even when the underlying architecture is loose. Conversely, what readers describe as poor structure may sometimes simply be abrupt transitions.

He jokingly produces the Yoda version:

There is no structure, only transitions.

When asked what makes a good transition, he says: not seeming like you fabricated one.

That is characteristic Graham. He identifies a craft technique, then defines mastery as making the technique disappear.

He also defends writing as a thinking tool late in the month. Humans obviously could think before writing existed, but for the hardest varieties of thinking he considers writing extraordinarily valuable.

His implicit claim is not "writing equals thought." It is that externalized language allows deeper recursive manipulation of ideas than unaided working memory.

Intellectual honesty: being wrong is a superpower

A Larry Ellison quote about Bill Gates becomes something like the month's epistemic motto: Gates did not care who was right, only what was right, which made him dangerous.

Graham says traders are forced to learn this because mistakes become immediately visible in P&L. Startup investing has a similar property.

He later demonstrates the norm himself.

After criticizing someone as effectively a government shill, he gets a detailed reply explaining that the person had longstanding views independent of the government work. Graham says the explanation seems reasonable and deletes the tweet.

Likewise with watches: someone tells him a supposedly pristine Omega dial is actually a redial. Graham says he was fooled, hated the watch once he knew, and sold it for one-tenth what he paid.

That is a small but revealing pattern: error is embarrassing; preserving the error is worse.

The month closes with another version of the same principle. Some mistakes become trivially avoidable after you make them once. If a domain contains only finitely many possible mistakes and you never repeat one, perhaps you asymptotically approach perfection.

He invokes Bohr's definition of an expert as someone who has made all the mistakes possible in a narrow field.

Watches are not a digression

There is an enormous amount of vintage-watch content this month, but it is philosophically continuous with everything else.

Graham likes watches from the "golden age," roughly before the quartz crisis, because they are simultaneously:

• engineering artifacts,
• aesthetic objects,
• evidence about technological history,
• markets with systematic mispricing.

His investment-style thesis is that brand prestige changes over time. Therefore watches from historically excellent manufacturers whose modern brands have lower prestige can be bargains. He repeatedly mentions Longines, Zenith, and Eterna.

He defines a "good deal" against someone claiming watch value is purely subjective: high accuracy + high build quality + low price.

That is exactly how he talks about startups and AI models: ignore prestige, inspect performance.

His favorite type of observation is something like: here is a 1963 Eterna bought for $509 plus a $69 service, still accurate to seven seconds per day.

The implied pleasure comes from finding objective excellence hiding beneath weak contemporary status signals.

The watch accuracy project

He graphs the accuracy of watches he owns against manufacturing year and sees rapid improvement in the early 1950s, a peak in the early 1960s, and some later regression as manufacturers optimized for thinner movements.

People argue about axes and log scales. Graham keeps refining the graph.

Someone overlays quartz performance, which instantly dwarfs the mechanical improvements.

Rather than ruining the exercise, this clarifies the historical transition: mechanical watchmakers had been climbing one optimization hill, then quartz changed the game.

He also points out that comparisons between his surviving wristwatches and historical performance frontiers are unfair: his observations are averages of individual watches still running decades later, while other datasets often show the best devices achievable under ideal conditions.

This is Graham the hacker more than Graham the collector: measure the artifact yourself, notice anomalies, generate a theory.

Why small old watches appeal to him

He dislikes the modern fashion for huge watches and likes 34 mm cases.

When someone complains that old watches become hard to read with age, Graham points out that hand geometry matters more than dial diameter and notes the use of dauphine hands.

When someone proposes converting old pocket-watch movements into giant wristwatches because large watches are fashionable, he tolerates historically authentic WWI conversions but dislikes modern ones.

Again the criterion is not "old = good." It is whether the design makes sense in its historical and functional context.

He is also unusually hostile to box-and-papers fetishism: they don't just fail to matter, he says, they're annoying to store.

Watchmaking as an AI-proof career

Late in the month Graham suggests watch repair as a surprisingly good career.

Independent watchmakers appear to have more work than they can handle, mechanical watches are growing in popularity, and dexterous physical repair seems difficult for AI to automate soon.

Someone objects that independent watchmakers complain manufacturers restrict access to parts. Graham answers with a simple revealed-preference test: they cannot simultaneously be getting destroyed economically and have backlogs of work.

This is classic Graham reasoning: distrust verbal descriptions of an economic situation when behavioral evidence points the other way.

Technology makes people richer by making units radically cheaper

A graph showing the historical cost of illumination inspires another recurring idea: technological progress often appears as orders-of-magnitude declines in the cost of useful output.

Lighting has become more than a thousand times cheaper.

Someone objects that inflation ate the savings. Graham says most things people consume exhibit similar long-run improvements.

The AI-inference discussion is explicitly modeled on this example.

The hidden conceptual link is abundance through unit-cost collapse. Electricity, computation, lighting, launch, inference: huge changes happen when an underlying input becomes cheap enough that people stop rationing it.

This is also why he is bullish on orbital data centers. Starcloud is fundamentally a bet that launch costs keep falling, which he compares to betting on Moore's Law in the 1990s. Even before orbit becomes strictly cheaper, the absence of terrestrial permitting can be valuable; below some launch-cost threshold, he thinks space infrastructure wins economically too.

"YC GDP" and industrial startups

He is visibly excited by startups making physical things.

Nox Metals is shipping metal into rockets, satellites, reactors, aircraft, hospitals, defense equipment, and even fences. Graham predicts it will become a big company and marvels that it is growing at software-like rates despite producing metal.

Stoke and Starcloud become an example of what he calls "YC GDP": YC companies eventually supplying one another in genuinely industrial chains rather than merely using one another's SaaS products.

A reusable-rocket company launching a space-data-center company is a qualitatively different kind of network effect from founders subscribing to Stripe.

The interesting thing is Graham's implicit view that the startup world is expanding back outward from pure software into atoms, while retaining software-era growth expectations.

The Ocean Cleanup and respect for people who actually solve things

Graham meets Boyan Slat and is impressed that a single organization can capture a meaningful fraction of river-borne plastic reaching the ocean.

When someone asks what makes Slat special, his answer is almost brutally simple:

Instead of just talking about problems, he solves them.

That sentence could summarize a large part of Graham's worldview.

He likes agents over commentators, builders over credential collectors, founders over professional managers, measured outputs over institutional reputation.

Institutions, incentives, and inherited trust

The political posts are scattered across different controversies, but the common argument is less partisan than institutional: systems decay when people consume trust they did not create.

After a report that the Netherlands was moving gold reserves out of the US, Graham frames it as the consequence of exploiting trust accumulated by more principled predecessors.

He makes a similar argument about founders versus hired CEOs: the founder remembers why customers trusted the company.

He repeatedly attacks institutional claims by asking for base rates or observable evidence rather than accepting the framing.

That pattern is much more coherent than the surface-level variety of political subjects.

Noncitizen voting: Graham keeps returning to the denominator

One recurring political argument is that noncitizen voting is being presented as a major problem despite extremely small reported numbers.

He cites examples from Georgia and Texas and repeatedly emphasizes that the reported numbers are upper bounds involving potential noncitizens, not confirmed cases.

His argumentative technique is mostly denominator discipline.

When someone cites a mayoral election decided by 13 votes and compares that with 117 potential noncitizens statewide, Graham responds by shrinking the relevant population to the fraction of Texas voters eligible to vote in that particular election, yielding an expected number below the margin.

The broader claim he advances is that changing voting rules to solve a numerically negligible problem suggests some other motivation.

Whether one accepts that inference or not, the interesting part is methodological: he keeps asking how big is the thing actually?

Politics brings out his least filtered side

His political replies are noticeably more combative than his technical ones.

He describes some claims as fake problems, openly questions politicians' honesty or intelligence, and frequently interprets rhetoric through incentives.

He says both major American parties' ideologies are "roughly equally mistaken," yet his attention during this month is asymmetrical because he is focused heavily on Trump-administration actions, immigration enforcement, voter-ID arguments, the Epstein files, and related controversies.

He repeatedly speculates that the administration's behavior around Epstein disclosures implies unreleased damaging material, and in one reply suggests Trump's erratic behavior serves to distract from the issue. Those are Graham's inferences in the corpus, not established facts supplied there.

The political mode is interesting partly because it reveals a weakness relative to his startup reasoning: he sometimes moves much faster from evidence to motive attribution than he would tolerate in a product or investment analysis.

Moral questions versus expert questions

Someone argues that prominent people should stay quiet about subjects outside their professional expertise.

Graham draws a sharp distinction.

For technical questions—how thick a bridge column should be—deference to experts makes sense.

For moral questions, he says expertise is not required. If something bad is happening, people should speak.

This is probably the clearest explanation of why his feed contains so much politics despite his primary expertise being startups and programming.

He does not regard political morality as a specialist domain in the same way structural engineering is.

The Bill Gates unit would be focus

One recurring mini-game at the end of the month asks: if a person's name became a unit, what would it measure?

At breakfast, Graham and family/friends assign units:

• Trevor: perhaps excessive generality in an engineering solution.
• Jessica: curiosity about another person's character, or incapacitation from laughter.
• Graham himself: amount of change in an idea.

Someone suggests a "Gates" would be a unit of money.

Graham instead says it should be a unit of focus, which is how Gates got the money.

This is more revealing than the joke initially appears. Graham's imagined self-unit is change in an idea. That is an unusually good description of what he values intellectually: not possession of knowledge but transformations of representations.

"Shapes of ideas"

He tells his 14-year-old that the most powerful imagination involves seeing the shapes of ideas: recognizing that two concepts fit together like jigsaw pieces, or that transforming one makes it isomorphic to another.

This is one of the month's most Graham-esque observations.

It describes abstraction spatially—not remembering propositions, but manipulating structures.

It also rhymes with his startup heuristics. A good founder or investor sees that a new situation has the same underlying "shape" as something encountered before, while still noticing when an apparent analogy breaks.

His children are becoming recurring intellectual characters

The month's family posts are unusually good.

Playing catch with his 14-year-old, the child dives for a ball and lies on the ground while the dog licks his face. Graham tells him: "This is the good old days." The child understands.

The same child later observes that Shakespeare now needs to be overacted because audiences have difficulty parsing the language.

These are small posts, but they provide a counterweight to the argumentative feed. Graham is highly aware of time, memory, and the fact that ordinary life becomes retrospectively precious.

Even the joke about Jessica bringing him tea works because he has been conditioned to interpret tea as support before some obligation. When she brings it for no reason, he briefly assumes he has forgotten something important.

Old objects and deep time

There is a broader antiquarian streak beyond watches.

An alabaster bowl made roughly five thousand years ago prompts: the future was even more unevenly distributed in the past.

The point is that historical technological capability was not smooth. Extraordinary sophistication could coexist with primitive conditions elsewhere.

A medieval siege account makes him wonder whether "pagan" functioned rhetorically somewhat like "terrorist" does now—an out-group category that compresses moral complexity.

An enormous medieval king leads him into a quick geometric calculation: a proportionally similar 6'10" human would have roughly twice the volume of a 5'5" one.

A Cambridge photograph initially looks to him like a painting.

A vintage aluminum Vacheron pocket watch fascinates him because even its movement plates were aluminum, making a 44 mm watch weigh only 22 grams.

These are all manifestations of the same instinct: look harder at the artifact; it often contains a hidden technical story.

Math, beauty, and discovery

Responding to Hardy's claim that ugly mathematics has no permanent place in the world, Graham fixates on the word "permanent."

That qualifier implies an ugly proof or formulation can discover something real, but eventually someone will find the beautiful version.

This fits his broader aesthetics of explanation. Beauty is not merely decorative. It often signals that the representation has compressed the structure properly.

He also follows discussion about AI solving mathematical problems. Someone predicts labs will eventually stop publicizing major mathematical discoveries because they will become routine.

Graham is skeptical that the analogy with olympiad problems fully transfers, because individual researchers inside labs will remain genuinely curious about open problems. And even if discoveries originate as side projects, he suspects the labs will want institutional involvement because mathematical breakthroughs confer prestige.

The press, news traffic, and information intermediaries

Graham notes that web traffic to news sites has fallen sharply over two years.

That sits naturally beside his argument about AI killing clickbait.

He also mocks earlier press predictions that Twitter would collapse after Elon Musk bought it, asking how journalists thought someone capable of running rocket and automobile companies would be unable to run a forum.

The inference is very Graham: demonstrated operational ability in harder domains should update expectations in easier ones.

Whether the analogy fully holds is debatable, but it shows his hostility to narratives that ignore prior evidence about individual capability.

Dynamic pricing and argument by definition

Elizabeth Warren criticizes dynamic pricing as a mechanism through which companies can raise prices opportunistically.

Graham responds that dynamic pricing also means prices can go down and treats the omission as evidence of either ignorance or dishonesty.

The argument then becomes more interesting in the replies.

Someone says prices never go "down"; they merely return to a baseline.

Graham asks what "baseline" means. If it means the lowest price the company is willing to charge, then the argument is tautological: of course companies never charge below what they are willing to charge.

Someone else accuses him of misunderstanding profit maximization. He asks whether they understand the difference between total profit and profit margin.

The substance aside, this is a good example of his argumentative style: force vague economic language into operational definitions and see whether the disagreement survives.

Nonprofits and termination incentives

Boyan Slat says there are two kinds of nonprofits:

1. institutions like museums or universities that should persist;
2. organizations created to solve a problem and therefore ideally disappear once it is solved.

Graham immediately points to the agency problem: employees of the second kind still want their jobs after the mission has succeeded.

It is a compact example of his instinctive incentive analysis. Organizations often acquire a secondary objective—organizational survival—that can eventually conflict with the objective they were created to pursue.

Wokeness as emergent coordination rather than conspiracy

In a reply to Elon Musk, Graham argues that political correctness or "wokeness" was not the result of an organized plan.

He says his historical research led him instead toward an emergent-coordination model: status-conscious moralizers needed new things to condemn, new categories of wrongdoing appeared, and people coordinated around them without central planning.

A respondent compares this to prices coordinating merchants without explicit collusion.

Graham likes the analogy.

The interesting point is structural: decentralized coordination can produce behavior that looks conspiratorial without a conspiracy.

That idea is useful far beyond the political context in which he applies it.

Humor: pedantry, inversion, and deliberately literal answers

A huge fraction of Graham's jokes use one of three mechanisms.

1. Taking metaphors too literally

Aaron Levie says AI is reaching "escape velocity."

Graham asks: what's the gravity well in this metaphor?

A founder asks whether walking solves most problems.

"It solves the problem of needing exercise. It may also solve others."

2. Correcting the premise rather than answering normally

Someone argues a smartwatch is superior to a mechanical watch and asks whether Graham also uses a mechanical calculator.

"I use a Gunter rule."

Someone says premium economy is better than business class was 30 years ago.

Graham replies that it's better than first class originally was.

3. Tiny linguistic reversals

Someone asks for the opposite of a chart crime after seeing an absurdly understated growth graph.

"Chartyrdom."

Someone proposes YC office plaques should be hats.

Then you'd only be reminded what to do when you looked in the mirror.

Someone says a slogan should be "Abolish Happy Hour."

"It's cleaner without the 'Hour'."

Someone writes "straw-person."

Graham recommends "individual experiencing being made of straw."

A commenter catches what they think is the "Muphry's law" typo in Graham's phrase "too too difficult."

Graham: "It's not a typo. And incidentally it's Murphy."

This style works because the humor and argumentative method are almost identical: notice exactly what the other person's words literally imply.

He enjoys correcting small factual errors

There are many tiny corrections:

• a watch is 34 mm, not ~32;
• an alleged automatic movement is clearly manual winding;
• a Japanese dealer's claim that the IWC Yacht Club was designed for yachtsmen is rejected as branding mythology;
• a medieval king reported at 7'3" was apparently closer to 6'10";
• a redial is diagnosed not merely from dial condition but from a contradiction with the movement;
• prehistoric demography makes "first time in recorded history" unnecessarily weak: an elderly population exceeding very young children is presumably a first in human history.

This can read as pedantic, but the aggregate effect is important. Graham's intellectual style is intensely error-corrective at small scales.

He treats factual accuracy almost like lint: if you see an error, remove it.

Experience as compressed failed experiments

Near the end, someone asks how investors detect the rare cases where conventional startup wisdom should be violated.

"Lots of experience."

That answer sounds trivial until paired with his post about mistakes.

Experience, in this feed, is not mysterious intuition. It is accumulated exposure to failed classifications. You see enough companies, watches, arguments, products, names, and people that you learn which features are causally important and which are merely correlated with success.

His YC defense makes the same argument institutionally: seeing hundreds of new startups every few months creates extraordinarily high learning throughput.

The month in one model

Across startups, AI, watches, politics, writing, and history, Graham keeps applying roughly the same procedure:

1. Ignore the prestige signal.
2. Look for an observable quantity.
3. Ask what incentives generate the behavior.
4. Compress the pattern into a heuristic.
5. Use the heuristic aggressively.
6. Remain willing to discard it when reality produces a counterexample.

That is why vintage watches, startup growth, AI pricing, voter statistics, essay transitions, and old company founders all fit together more coherently than they first appear.

He is fascinated by hidden causal structure beneath misleading appearances.

A neglected Longines may be technically superior to a prestigious modern watch.

A founder with no credentials may be investable because the company is growing.

A company with impressive fundraising may still have bad retention.

A token price may be rising while useful inference becomes radically cheaper.

A rule that is right 99.5% of the time may be disastrously wrong for the exceptional company in front of you.

An institution that appears powerful may merely be spending inherited trust.

An ugly mathematical result may be a temporary representation of something whose final form will be beautiful.

And a mistake that once seemed invisible can become impossible not to see after you have paid for it once.

That recurring search for the real variable is what makes the month interesting. The subject matter jumps constantly, but the cognitive style barely changes.