Paul Graham
The month in one sentence
August was unusually revealing because Graham spent much of it doing YC office hours, so his feed became a running demonstration of how he actually thinks: strip away jargon, find the variable that matters, test against reality, simplify aggressively, and let small empirical gains compound into something huge.
Around that core were AI, writing, capitalism, immigration, academia, watches, art, children, and an increasingly acrimonious argument with Palmer Luckey. Oddly, they mostly illuminate the same worldview.
Startups: simplify until the real problem appears
The dominant subject was YC. Graham was in California for much of August talking to the latest batch, sometimes hundreds of companies in rapid succession.
His most characteristic observation:
A surprising amount of what I do in office hours is give founders permission to do something simpler than they were planning to.
This recurs everywhere.
Founders explain their companies in long, confusing ways; he “decrypts” the explanation and gives it back in three sentences. Demo Day pitches get reduced to five memorable “vertebrae.” A startup whose idea is failing gets permission simply to throw it away and start again. Another has three plausible routes to becoming huge; instead of predicting which grand strategy will win, Graham tells them to optimize for growth and let reality choose.
The striking example was a company ready to give up. Graham sent the founders to lunch to invent new ideas. They returned with a good one; within minutes they had also found a good domain. He emphasizes that this was freakishly easy: normally finding a genuinely promising idea requires exhausting thought.
The broader lesson is that startup ideas are not sacred objects. They evolve, and often disappear entirely. This connects with advice he later gives his son: to do great work, you have to be willing to throw things away because you never become good enough to get everything right on the first attempt.
Growth is the high bit
Graham remains almost comically reductionist about evaluating startups.
One startup thought 30% monthly growth might be inadequate. He pointed out that maintaining it would require an office 23 times larger every year. Another casually revealed that before a launch spike it had been growing 40% per week; he calculates that `1.4^52` is nearly 40 million, obviously unsustainable but overwhelming evidence that users want the product.
His “secret” for judging a startup:
What's their growth rate? That's the high bit, and in base 10.
If growth is genuinely good—not merely purchased by giving money away—then many other things must already be sufficiently good.
This also explains why founders frequently misjudge themselves. A first startup provides no reference class. Some founders fail without realizing it, but surprisingly many are doing extremely well and still think they are mediocre.
Consequently Graham spends far more time encouraging founders than the mythology of swaggering entrepreneurs would suggest. Good founders tend to take their strengths for granted and obsess over weaknesses. When they deserve confidence, he thinks simply telling them the truth about their performance is valuable, partly because investors will notice whether they project confidence.
Incremental ambition, not delusion
Someone challenged the apparent contradiction between building enormous companies and avoiding overconfidence. Graham's answer is one word:
Incrementally.
You solve one real problem, then another. Compounding does the rest. Years later you look up and discover that the thing has become huge.
He explicitly rejects the fashionable claim that extraordinary founders need to be “delusional.” His counterexample is the Collison brothers: enormously ambitious, but in his view exceptionally undelusional.
This may be the most important startup idea in the month. Graham's model is not:
Think gigantic thoughts → believe them hard enough → manifest gigantic company.
It is:
Maintain contact with reality → keep solving increasingly valuable problems → allow exponential processes to generate the gigantic outcome.
AI changes the opportunity set more than the laws of startups
Graham is clearly convinced AI is transformative, but he pushes against the idea that it has already repealed the old rules.
This is his 47th YC batch, and he says nearly all standard startup advice still applies. There is more variation between two companies in the same current batch than between pre-AI and post-AI startup advice.
The core remains:
Build what users need and get growth.
What AI has radically changed is what can be built.
He says founders worried about AI are paradoxically among the safest people in an AI upheaval because a small, fast company can change direction more easily than almost any institution. He casually mentions having met three startups in one batch with plausible plans to replace Nvidia.
AI also appears to be changing company structure. Single-founder YC companies doubled from 9% to 18% year over year, presumably because one person can now accomplish more. Graham still advises against going alone, because the important function of a cofounder is not merely labor capacity but sharing psychological load.
Does AGI already exist?
His provocative benchmark is historical rather than definitional:
If you showed current models to someone in 1980, would they say AGI had been achieved?
He thinks the answer would clearly be yes. Our uncertainty comes from standing close to the boundary ourselves.
It's an interesting reframing because it treats “AGI” partly as a moving reference point. Capabilities that would once have been regarded as decisive stop feeling decisive once they become ordinary.
At the same time he posts a terrible AI-generated map of Saracen territories in Italy and remarks on the absurdity of simultaneously worrying about AI while getting results like this. The tension is deliberate: astonishing general capability can coexist with ridiculous local incompetence.
What he would do at 17
If Graham were 17 now, he would not rush to found an AI wrapper company.
He would:
learn to build LLMs from scratch; train the strongest ones he could with available hardware; probably give them some frivolous task such as playing games; then use the resulting deep knowledge to find problems worth solving.
When someone asks why reinvent the wheel, his answer is excellent:
To learn how wheels are made.
This is Graham's education philosophy in miniature. Don't optimize prematurely for startup output. Build generative capability first.
His essay on university preparation makes the same argument: students should pursue projects of their own. He even thinks some prospective founders skipping college to start companies is already happening—microscopic among all students, but perhaps 5–10% of people who would eventually become founders—and says he does not think this is necessarily good.
AI and thinking: the danger is not fake prose but outsourced cognition
Graham has become obsessed with detecting AI writing because, as he tersely explains when asked why:
I'm a writer.
His strongest stylistic tell is “excessively colorful verbs”: AI says a proposal “drew” 100 votes where a normal writer would simply say it “got” 100 votes.
Someone defends colorful verbs by invoking writing advice. Graham notices that the quoted passage itself uses plain verbs: “have,” “choose,” “be.”
This fits his longstanding aesthetic: good writing is not fancy writing.
But his deeper worry is cognitive rather than stylistic. When someone describes sending every half-formed nighttime idea to ChatGPT for deep research, discarding 90% of the resulting reports, Graham asks a sharp question:
What if exploring some of that 90% yourself wouldn't have lead to nothing though? You could be unclogging your mind by burning your ideas.
That is probably his best AI observation of the month.
Exploring an idea yourself is not merely a costly way of obtaining an answer. The exploration changes you. If an AI cheaply jumps from question to conclusion, it can destroy the path on which your own adjacent discoveries would have occurred.
He makes exactly the same point about writing. You can discuss an idea endlessly with intelligent people and still discover something new when you write it down.
Writing is therefore not transcription of thought. Writing is a technology for generating thought.
AI may also destroy bullshit retroactively
A particularly interesting inversion concerns academic fraud.
Someone worries that future humanities fraudsters will simply use AI to manufacture scholarship, eliminating plagiarism as a detection mechanism. Graham agrees AI may help them produce new papers—but thinks it could simultaneously make their old work catastrophically vulnerable.
Bogus scholarship historically benefited from an unusual defense: deciphering it was too tedious for competent humans to bother.
AI has infinite patience.
So mountains of opaque papers that authors assumed would sit safely in archives forever can now potentially be analyzed systematically by future machines.
This is a recurring Graham pattern: technologies do not merely change future production; they can change the meaning and vulnerability of artifacts created in the past.
Capitalism: wealth creation versus extraction
Graham spends several days arguing about inequality, taxes, and capitalism.
His basic distinction is between creating wealth and extracting more of an existing surplus.
The best way to “win at capitalism,” he says, is to improve customers' lives by making something new. Squeezing customers harder might produce a 2× result; creating something dramatically better can produce 10×.
He admits plenty of businesses behave like the socialist caricature of capitalism—opportunists and bean counters really do squeeze people. His claim is that these are analogous to mediocre mathematicians doing plug-and-chug exercises: they exist, but they are not what explains the spectacular outcomes.
The “big stars” create wealth.
He illustrates this with Tesla: Jessica's new car costs about $48,000 with roughly 350 miles of range, while her 2015 Tesla cost the inflation-adjusted equivalent of $135,000 and managed around 270 miles. To Graham, that reduction in what someone must sacrifice to obtain the capability is literally an increase in wealth.
His inequality argument is narrower than many replies assume
He repeatedly distinguishes two questions:
- Did changes in tax policy cause founders to become vastly richer?
- Could tax policy prevent people from becoming or remaining that rich?
He says no to the first and obviously yes to the second.
His claim is that modern technology permits companies to be started more easily and scale far faster, producing correspondingly extreme fortunes. Larry Page and Sergey Brin did not become rich because tax rates changed; they became rich because Google could become enormous.
Whether society should subsequently tax fortunes heavily is, in his view, a separate normative question.
That distinction matters because a lot of the argument around his tweets collapses causal explanation into policy preference.
Taxes and Silicon Valley
Graham also opposes proposals that he thinks would cause wealthy founders to leave California, bringing companies with them.
When someone mocks the idea that a founder would avoid becoming a billionaire merely to escape a few-percent wealth tax, Graham points out that the actual behavior at issue is moving to another jurisdiction, not refusing to become rich.
He claims California is already near a practical upper limit, comparing its roughly 50% combined income-tax rate with Sweden's roughly 52%.
Whether one accepts the empirical conclusion or not, the conceptual argument is characteristic Graham: model the actor's available moves instead of assuming the actor passively accepts the policy environment.
Merit, immigration, and the Palmer Luckey blowup
The political centerpiece of the month begins with an intentionally banal sentence:
I believe the most qualified person should get the job.
Graham's point is that the political valence of this sentence has flipped.
In 2023, he says, it was attacked from the left as a supposed microaggression. Now it provokes parts of the right because “the most qualified person” might be an immigrant.
He defines qualification operationally: possessing the qualities that predict high performance in the job. Asked how to overcome bias, he says: measure the results of your choices.
He challenges JD Vance's assertion that American corporations needing workers should hire and train Americans by asking whether Americans should likewise be allowed to work in other countries.
Then Palmer Luckey argues that American workers should receive legal preference and objects to the idea of replacing Americans whenever a foreign candidate scores marginally better.
Graham notices that Luckey's own US workforce includes noncitizens and keeps asking one deliberately narrow question:
Why did you give those jobs to foreigners rather than Americans?
This produces several days of escalating hostility.
Eventually Luckey says he actually strongly supports immigration but rejects Graham's cosmopolitan framing and calls him a “traitor” for becoming rich in America, moving abroad, and criticizing Americans from Britain.
Graham's response is basically logical debugging. If:
rich + lives abroad + criticizes Americans = traitor
then the rule captures huge numbers of expatriate Americans, including conservatives who criticize American progressives. He repeatedly tries to force Luckey either to specify a narrower rule or accept the implications of the one he stated.
By the end, the substantive immigration disagreement appears much smaller than the rhetorical war suggested. Graham explicitly says their actual positions may not be very different.
That is what makes the exchange interesting. It becomes almost a laboratory demonstration of one of Graham's recurrent obsessions: political tribes create enormous conflicts by replacing precise propositions with identity-loaded caricatures.
It also shows Graham at his least detached. The argument becomes personal, and he questions Luckey's character and judgment. Whatever the merits of either side, this is no longer the serene essayist explaining a principle from first principles.
Graham's politics are increasingly “anti-both-tribes”
Several other posts fit the same pattern.
He thinks Democrats disastrously associated themselves with pronouns and land acknowledgements, which he regards as useless symbolic politics. But he is also openly hostile to MAGA-style anti-immigration absolutism and repeatedly mocks Trump's behavior.
He describes the worst-case Republican as a “movie bad guy,” argues that Trump's Iran war may have been partly intended to distract from Epstein, and jokes that renaming Lake Ontario “Lake America” would be the next diversion:
R1: Would they really buy that? R2: Think about our base.
He also pushes back against Americans determined to believe Britain is a dystopia, arguing that a viral Oxford Union clip was being misrepresented.
His political self-conception seems increasingly explicit: the stupidity that he thought dominated one side during peak “wokeness” is now appearing in mirror-image form on the other.
After someone says John Carmack was once booed for advocating meritocratic hiring, Graham replies:
It's back, but from the other side.
Academia: truth business versus institutional ideology
Graham is particularly unforgiving of professors because he thinks they have a special obligation:
Professors are supposed to be in the truth business.
He follows the controversy around Nathan Cofnas, who said Ghent University suspended him after his role in exposing alleged academic misconduct. Graham frames it as a perverse institutional response: instead of cracking down on fraud, the university is punishing the reporter.
This connects to his older “Great Awokening” concerns. When someone compares professorial firings during recent ideological controversies with the Red Scare, Graham's immediate reaction is methodological: shouldn't the comparison be as a percentage of professors, since there were far fewer academics in the 1950s?
That tiny reply is representative of him at his best: before arguing about the ideological conclusion, fix the denominator.
Bureaucracy and jargon as linguistic smells
Graham pays close attention to phrases that reveal institutional pathology.
“At this time” in a rejection means someone has gone “full bureaucrat.” A proposed “quick call” is usually unpleasant; “quick” is being advertised because brevity is the only redeeming feature. Replacing “sales” with “GTM” strikes him as bogus jargon. Insider terminology can be necessary when genuine distinctions exist, but often signals a desire to sound knowledgeable rather than communicate.
This is not merely linguistic crankiness. Graham treats language as diagnostic instrumentation. People who are thinking clearly can usually say what they mean plainly.
That same instinct drives his office hours: founders arrive with elaborate descriptions; Graham attempts to discover the simple sentence underneath.
Europe: the problem may be culture before regulation
Patrick Collison recounts a German founder having a 90-page investment contract read aloud by a notary, at a cost of €30,000, because of German legal requirements. Graham contrasts this with YC's standardized SAFE, where investors can largely trust that the boilerplate is identical and look only at names and numbers.
Yet when discussing Europe's startup deficit, Graham resists the obvious regulatory monocause.
He agrees restrictive labor laws hurt startups, but argues they may not yet be the binding constraint. Europe does not have enormous numbers of startups forming and subsequently dying because they cannot fire employees; far fewer are founded in the first place.
His deeper explanation is cultural: starting companies is simply less customary, including institutional and popular attitudes that are more hostile to entrepreneurship.
That is a more interesting hypothesis than “Europe has bad employment law” because it shifts attention from the visible obstacles encountered by existing founders to the invisible population of people who never seriously consider founding anything.
Names, domains, and concentrated ambition
Graham retains an almost YC-era obsession with company names.
A startup spends $250,000 on a domain, which sounds insane—until he learns it raised $6.5 million. Suddenly the domain was only 1/26 of the round.
His test is:
Could this be the name of a Google-sized company?
He loves Discovered Materials, partly because it echoes Applied Materials and apparently cost essentially nothing. Another startup's new idea is followed almost immediately by finding a good domain.
The implicit theory is that naming is cheap in the economic sense but consequential in the psychological one. A good name lets you imagine the company becoming enormous without the name itself becoming embarrassing or constraining.
Watches: Graham has fallen deep into the golden age
A surprising fraction of the month is horology.
He praises:
Longines cal. 291; Patek Philippe 27-460; Omega 561 and 552; IWC 85x movements, especially the 854; * Girard-Perregaux Chronometer HF.
His “golden age” is roughly 1945–1970. These watches can often still be regulated to astonishing accuracy decades later, and unfashionable steel cushion cases sometimes contain historically exceptional movements for under $1,000.
But the important explanation is why he cares:
peak golden age watch movement = the culmination of 700 years of work on mechanical clocks.
Mechanical clocks appeared around 1270, and generations of first-rate minds improved them. Quartz then made mechanical movements obsolete as timekeepers. Graham therefore sees the best mid-century movements as the terminal artifacts of a centuries-long technological optimization process.
That is a very Graham way to collect things. He is not primarily buying luxury signaling devices. He is attracted to late-stage artifacts from a technology just before its functional extinction.
He also prefers the imperfections of old watch dials to modern perfection. New dials feel too hard and flawless, like digital recreations of old metal typefaces.
Art, typography, and imperfection
The watch obsession connects to his visual taste.
He is impressed by Monet's ability to capture reflected colors that are genuinely difficult to perceive. He prefers Edward Bawden-like illustration in which “each tree is a portrait.” Asked for the most consequential painter ever, he answers simply: Giotto.
He argues that abstract modern art works unusually well painted onto racing cars because abstract art is essentially decorative art; it may look dull hanging independently on a wall but becomes effective when decorating another object.
On typography, he argues that what digital reproductions of mid-century metal type miss is not primarily physical indentation but ink spreading into paper and softening the letter edges.
Even a friend's algorithmically damaged square tiles trigger the same sensitivity: they're too close to perfect squares, trapped in an “uncanny valley of squareness.” Make them less perfect, he suggests.
Across watches, type, illustration, and tiles, Graham repeatedly prefers structured imperfection produced by a real process over immaculate synthetic approximation.
Family is not a side topic
His children and Jessica appear throughout the month, usually supplying either jokes or surprisingly good conceptual prompts.
A 14-year-old tells him:
You're the boss. I'm just the guy who's right.
The dispute was about where to have lunch.
Asked what YC's mascot should be, Graham suggests a cockroach, because great startups survive hostile environments. His son says that is not very appealing. Graham later realizes the cockroach's obvious name:
Feature.
When the same son asks what DHL stands for, Graham improvises:
Deutsche heilige Lager?
He also develops a surprisingly good analogy for adolescence: teenagers rejecting their parents' ideas is like leaving agricultural fields fallow to interrupt parasites. It is annoying, but inherited stupidity doesn't get transmitted indefinitely. Even if children later recover most of their parents' beliefs, throwing everything out and selectively rebuilding can eliminate the bad parts.
That is, again, basically his epistemology: inheritance should be subjected to destructive testing.
Work and family
The family material is not merely comic.
With founders of mature YC companies, Graham says one of the central questions is how to combine work and family. His strongest advice is simply not to ignore the tradeoff:
You can't get those years back.
This stands out because the stereotypical startup doctrine is to optimize nearly everything for the company. Graham is explicitly saying that the objective function has another irrecoverable variable.
The month's best jokes and throwaway lines
A few deserve preservation:
Blake Scholl jokes that immigrants having visa trouble should switch to Mastercard. Graham: that's already the administration's plan—Trump has a “card.” A YC cancer startup helped achieve Sid Sijbrandij's remission. Graham addresses the disease itself: “It was a big mistake attacking Sid, cancer. You created a very dangerous enemy.” Someone opens an SF bar called Liquidity where you enter through the Exit. Graham calls startup dad jokes common in his house, “or maybe not popular, but at least common.” “Dial shows significant age-related patina.” Nothing else; the euphemism is the joke. Asked about a bizarre founder photo: “It looks like one of us is trying to imitate the other's facial expression.” Someone suggests Bryan Johnson as YC's mascot: “He's already the mascot for Bryan Johnson.” Jessica describes herself as “the squarest eccentric I know.” A guest brings a ten-pound block of aluminum as a house gift. A founder's leadership book is described as getting competitors only “to third base.” On a prediction cryptographically committed for revelation in 2028, Graham's immediate question is simply: “Is it about AI?” On excessively elaborate self-tracking: “Did I safely skip a whole cycle?” On the YC veteran claiming Graham spent 12 hours with founders: “7, but it felt like 12.”
The deeper pattern: Graham keeps searching for the high bit
What makes the month interesting is that subjects that appear completely unrelated are processed with almost the same algorithm.
For a startup: growth rate.
For a pitch: the five things investors will remember.
For a company description: three clear sentences.
For hiring: qualities that predict performance.
For bias: measure the outcomes.
For choosing among startup directions: let growth decide.
For capitalism: distinguish creating wealth from extracting it.
For inequality: separate what caused fortunes from whether taxes can eliminate them.
For writing: ideas first, plain language second, decoration far behind.
For AI education: understand the mechanism, don't merely consume its outputs.
For artistic taste: look at the process that generated the artifact, not superficial perfection.
For political arguments: turn slogans back into precise propositions and force them to survive their implications.
This is why the office-hours material is especially useful. You can see what “Paul Graham advice” looks like before it has been polished into an essay. It is not primarily a library of startup doctrines. It is a habit of repeatedly asking:
What actually matters here?
A second pattern: variation plus selection
Another thread connects surprisingly many posts.
Graham wants rejected YC companies to succeed because a successful reject reveals a mistake in YC's model. He wants YC partners to disagree somewhat because “evolution requires variation.” He tells companies with multiple directions not to choose by grand theory but to let measured growth select among them. He tells teenagers, metaphorically, to throw inherited beliefs out and reconstruct them. He tells hiring managers to evaluate whether their supposedly unbiased selections actually produce good performers.
Even his writing process is evolutionary: write material, then throw much of it away.
The implicit method is:
- generate alternatives;
- expose them to reality;
- preserve what survives;
- remain willing to discover that your previous filter was wrong.
That may be a more fundamental description of Graham's worldview than “startup thinking.”
A third pattern: AI makes fundamentals more valuable, not less
Despite all the excitement, Graham's strongest AI-related advice is surprisingly conservative.
AI makes implementation cheaper, so judgment becomes more valuable. It makes writing easier, so doing your own thinking becomes more valuable. It enables individuals to accomplish more, but doesn't eliminate the emotional value of a cofounder. It generates endless information, making first-hand understanding more important. It can accelerate design, shifting the bottleneck to physical construction and eventually regulation.
And if everyone can ask an AI for answers, the scarce club may consist of the remaining people willing to think and write at length.
He predicts that in ten years there may be very few people who can or want to read and write anything longer than a couple pages.
His reaction is not despair:
There will be at least a few of us, and we'll be a powerful club.
That line probably captures the month best. Graham is bullish on AI precisely because he believes the things it commoditizes will reveal more clearly what was scarce all along: taste, judgment, curiosity, courage to discard bad ideas, and sustained independent thought.