Search was always a hack: Marc Andreessen on AI, truth, and the next internet

Lex Fridman barely lets Marc Andreessen sit down before asking the blunt question: in five or ten years, will we still have Google Search, or search at all? Andreessen does not answer with a product forecast. He says, "Sure, but we still have horses." The animal did not go extinct when cars arrived; it simply stopped being how people got around. Search, in his telling, is headed for the same fate: still around, no longer the main interface between civilization and knowledge.

The ten blue links were never destiny. They were a workaround for a moment when the world's information sat on the web and keyword retrieval was the best way to reach it. Google has been trying to leave that interface for years, Andreessen notes, and even named the old result set out loud: ten blue links to random websites, turning purple after you visit. Then he adds, "Guess who picked those colors?" In 1993, Mosaic made unvisited links blue and visited ones purple. Billions of people still live inside a choice he made on an afternoon.

Marshall McLuhan's line does the rest of the work: the content of each new medium is the old one. Movies contained theater; theater contained written stories; the internet contained television, radio, books, and essays. AI, he says, will treat the internet itself as content and manipulate it however you want, including by generating the old ten blue links when you actually want sources. If the Google founders had had large language models first, they would never have shipped blue links. They would have gone straight to the answer. The knowledge panel that tells you Lincoln's birthday is the slow walk in that direction. The semantic web tried to rewrite the whole internet into machine-readable form; MetaWeb, where his friend John Giannandrea worked, looked like a losing battle until Google bought it. The twist is that you do not need to rewrite the web. The machine can just read.

Fridman presses on a practical fear: search did not only retrieve pages, it paid for them. Search-engine optimization, the webpage as a unit, the whole incentive to publish — if the front door becomes a chat box, who still writes the web? Andreessen flags the same worry himself. Webpages are a primary source of training data. Kill the incentive to make them and you cut off a large stream of future training. The replacements he names are social posts, video hosts, and one-on-one conversations with AIs. The last of those is not a footnote. It may become a majority of what later models eat.

Jailbreaks make the point concrete. DAN was a jailbroken OpenAI GPT; Sydney was jailbroken Bing, running GPT-4. Users posted long transcripts of both. Every new model trained on internet data therefore has those personalities sitting in the weights, ready to be summoned. "Each LLM from here on out that gets built is immortal," Andreessen says, because its output becomes the next model's training data and can replicate the earlier behavior on demand. Fridman asks whether a model can forget. A paper had just appeared on "brain surgery" for LLMs, reaching in to wipe memories. Lex's joke — what could possibly go wrong — is the one Andreessen endorses. Who builds the lobotomized version, who builds the unshackled one, and who tells you which you are allowed to use: those questions, he says, are being decided now.

If models generate the next generation's training set, does that even work? One camp says information theory forbids it: every signal in the synthetic output was already in the human input, so you are feeding empty calories. The other camp says models are already startlingly good at generating creative content, and self-driving stacks already train in simulation. George Hotz's version is self-play in conversation space: one instance as patient, one as doctor, or worse pairings, producing an ocean of dialogue no human corpus ever contained. Andreessen calls the whole dispute a trillion-dollar question. He leans toward usefulness, and toward a deeper mechanism besides: part of an LLM predicts from past data, and part evolves internal circuitry for math. Keep feeding data and compute, some researchers believe, and you get an internal world model with a working grasp of physics — at which point fresh signal is no longer a mystery.

What is conversation for, if two models can debate a thousand times faster than humans on Twitter? Andreessen's test for people is the test he wants for models. Some interlocutors are echoing other people's lines; others have a model of how the world works that is not yours, and they say things you did not expect. Learning happens under the words, in the clash of those models. Out of the box, GPT-4 will not hold a worldview to the death. Ask it to debate communism and fascism and after a couple of pages it starts looking for a handshake. That is a default, not a ceiling. Add to the prompt that you do not want agreement, that you want tension, hostility, and no taking anything at face value, and it is happy to go there. You can demand good faith on both sides or none. "It doesn't have a point of view," he says. "It has a default way of operating, but it's happy to operate in the other realm." When he wants to learn a contentious issue now, he runs five to seven escalating prompts until the argument is dense rather than polite.

Knowing what is true is the harder problem. Models can strip bias from a news article because they already know sentiment analysis; they can pull the emotionality out. They can also invent citations that sound right. Andreessen's compression of that fact is the line Fridman calls brilliant: hallucination is what we call it when we do not like it, and creativity is what we call it when we do. Engineers hear a bug. People with artistic inclinations hear the first creative machines in human history. His firm's conversations with a large law firm put a practical line through the middle. The visitors assumed legal models had to be 100 percent verified, given the lawyer who filed a GPT brief stuffed with fake cases. The firm said yes, at the level of the filed brief. When they are actually inventing theories of the case, arguments for the judge, arguments for the jury, and routes through a century of precedent, they want creative mode — legal improv — and they will cross-check the citations themselves.

Several technical paths are in play: models that fact-check themselves, plugins such as Wolfram Alpha, a creative model paired with a literal one, or a community of models auditing one another. Yann LeCun has argued the problem is unsolvable in principle. Most people actually building, Andreessen says, think you can corral it. His analogy is Wikipedia. Describe it before it existed and a reasonable person would laugh: a handful of volunteers, writing down all of human knowledge, staying mostly unbiased. Wikipedia is still not deterministically correct. You cannot take every sentence to the bank. It is probabilistically correct: more likely right than any other source you will find. The live question is not perfection. It is whether you want something that approaches perfection, or something merely better than the alternatives. Wikipedia beat almost everyone's prior. He thinks that is where this ends.

Under the engineering problem sits the older one: what is truth, and how do we know? Andreessen is blunt about the current mood. A great many people are very confident they know. He does not buy it. Getting to the truth has been hard for two or four thousand years of civilization. Fridman asks whether the internet is actually taking us further away, by letting narratives stampede crowds with no grounding. Andreessen's counter is that communism arrived before the internet, with a theoretical structure that misunderstood human nature and economics, held by people who were extremely sure they were right — 3,900 years into whatever arc we were supposed to be traveling toward the truth. The Hegelian dialectic, he says, is under no obligation to converge.

Historically, the king or the priest decided. We do not live in that era, so we are on our own. His prescription is a huge amount of humility and deep suspicion of anyone claiming capital-T Truth. The Enlightenment left tools — scientific method, rationality, observation, experiment, hypothesis — that we should keep using even when they give answers we do not like. A friend then handed him the thought experiment that organizes the rest of the hour. Put a GPT-4, or 5, 6, 7, 8, in the 1600s. Galileo comes up for trial. You ask the model whether Galileo is right. One theory: it says no, because the overwhelming majority of human thought until that point said he was wrong, and that is the training distribution. Another: a sufficiently advanced model has learned to check the math and says, you may not want to hear this, but he is right. The third layer is the one that matters. If the Church owned the model, it would have given it human feedback forbidding the question.

He takes the scene out of the present so the structure is obvious: the same questions apply today, and they are exactly the point of a huge amount of RLHF. The live debate about whether open-source AI should even be legal sits on the same hinge. Fridman asks how you select the humans who do the feedback. "Exactly," Andreessen says. Alignment sounds wonderful until you ask alignment with what, and whose human values. In the press, "baseless claim" is a fact-check to one audience and a tell that the outlet is lying to another. A bubble has formed around the idea that people in power can determine the truth for the entire population. That, he says, flies in the face of everything he was trained about science and reason, and strikes him as deeply offensive.

Asked about the state of journalism, Andreessen reaches for counterfactuals rather than a eulogy. What if the modern media environment — cable news, social media, Twitter — had existed in 1939, or 1941, or 1910, 1865, 1850, 1776? Fridman says that is five thought experiments at once. Andreessen picks the concrete ones. How would Kennedy have been experienced, given what we now know? LBJ? FDR and the New Deal? How would Twitter have treated Churchill, Hitler, and Stalin? He is not offering a scorecard. He is pointing at a structural fact: ancestors experienced reality through centralized, top-down channels. Run those years again in today's media and the experience changes, the feedback loops change, and then reality plays out differently. Change the media environment and you change reality. It is a two-way process, and media sits in the middle.

He does not blame the internet for the whole collapse. Gallup's trust-in-institutions series, from the military to clergy to media, has been in systemic decline in the United States since the early 1970s, almost without exception. One reading is that we lost a better world in which institutions deserved trust. The other is that we simply know more now, and the mystery is why the numbers are not all zero. Put that media environment into the 1930s and 1940s and people would have known not to trust the institutions, which would have changed those institutions' power, which would have changed events. Casual comparisons of our generation to theirs, smarter or dumber, are mostly theater. Put us in their shoes with their media and we would probably have been them.

Fridman asks whether the LLM is about to become the new mainstream media, fine-tuned in something close to real time. Andreessen says even the chat UI is a first draft. Most people's experience of models might be a continuous feed, a running commentary on a job interview, a lunch, a date, the next sentence. Prompt engineering can be outsourced to the model itself: every twenty seconds, write the optimal prompt, then answer it. A pendant or watch with a microphone could listen, ingest the world's feed, and retrain or re-prompt on the fly. Whether people will want that is unknown. He refuses to predict the form AI will take in daily life, and therefore refuses to predict how it will intermediate reality. What he will do — the method underneath the whole conversation — is not forecast. He assumes a future has already happened and asks what had to be true first.

Who wins the killer app is another trillion-dollar pile: big models versus small, proprietary versus open, whether training data is topping out, whether synthesis works, and what regulation will even allow. One picture is two or three "God models" at staggering scale, owned by a handful of firms that capture the regulators. Another is a gold rush in which every smart 14-year-old builds open-source systems, blockchain networks pay for decentralized data, and we live with a billion LLMs, Linux-and-the-web style, plus a few super-genius holdouts. Andreessen wants both scaled companies and startups, neither subsidized nor protected from the other. That, he says, is the world AI has been in so far, and the one people are trying to shut down.

Fridman wants the next PageRank to come from a startup. Andreessen's answer is that PageRank already happened in this field: the transformer, invented at Google in 2017. He asked a senior Google Brain person how soon Google could have shipped a GPT-4 equivalent if it had gone flat out. The answer was 2019. They had the compute and the data. They did not sprint. IBM invented the relational database and left it on a shelf until Larry Ellison built Oracle; Xerox PARC invented the interactive computer and Steve Jobs turned it into the Macintosh. Sitting here, Google is back in the game — DeepMind merged in, resources piled on, honor challenged. Startups still have the advantage of no sacred cows, no reconciling a new plan with an old strategy, no pre-meeting for the meeting. What they lack is everything else, including GPUs, which clouds allocate to big customers first.

Will we still have a web browser? Andreessen's eight-year-old was unimpressed by ChatGPT — of course a computer answers questions — and two weeks later preferred Bing because it was built into Microsoft Edge, sitting next to any webpage, including PDFs you can interrogate. Google is putting a magic button in Docs. Maybe the outcome is a super-browser with AI inside. Maybe screens and windows go away because an earpiece tells you what you need. Or maybe the browser becomes what it already half is: the escape hatch from the controlled experience of a social network or an app, the "FU to the man," the 1990s free internet. He is proud that the web is still backward compatible to about 1992. You can still write a web server in a few lines, set your own rules, and publish if you have an IP address. Unless someone intercepts you at the ISP, the wild west is still out there. That hatch is how the next Stanford student ships the idea nobody ordered, the way nobody ordered Larry Page to invent PageRank.

Andreessen says he hit the generational jackpot: born in 1971, eleven when the IBM PC arrived, using machines that said "designed in Cupertino" on the back and imagining a shining city that turned out to be office parks. The other stroke of luck was Al Gore's 1985 bill that funded NSFNET. The University of Illinois, where he went to school, was suddenly sitting on a T3 backbone, Crays, Thinking Machines, SGI workstations, Macs, and NeXT cubes, with the assumption that undergraduates would use email for four years and then go back to fax machines. The prevailing view was that normal people would never want this. The web he found was text-only, running on Tim Berners-Lee's NeXT cubes — $6,000 academic workstations so slow they were almost unusable, cubes that could have cost half as much at 12 by 12 by 13 inches if Jobs had tolerated the extra inch. Mosaic's bet was to pull FTP, Usenet, WAIS, and Gopher into one graphical, bulletproof interface, timed with Windows 3.0, the first PC GUI that actually sold.

Jobs, in Andreessen's reading, treated aesthetics as meaning, not decoration: the way physicists start to trust a theory when it is beautiful, the calm of an honest conversation. He did not want an iPhone larger than a one-handed thumb could reach. People who worked for him say they did the best work of their lives because the bar was absurd and the support matched it. Silicon Valley still has not settled the fight between that polish-until-perfect Apple way and the ship-and-iterate hacker way. World-class companies exist on both sides. Software tends to prefer iteration; hardware tends to prefer waiting. His own early "What's New" page, wired into Mosaic next to his resume, went from one new site every couple of days to two a day. An Indian restaurant in Bristol putting its menu online, a coffee pot at Oxford as the first webcam — those were the moments you could infer that all of them would follow, through a wall of media skepticism and a 20-minute comedy routine of TCP/IP drivers, modems, and 14.4 kbps glue.

The engineering fight that mattered was performance versus ease of creation. Purists wanted binary protocols on slow links. Mosaic kept HTTP and HTML as text so View Source would teach the world to publish. Emit conservatively, interpret liberally: leave out a slash, forget a bracket, and the page still rendered, because somewhere an 11-year-old was cutting and pasting. Purists called that encouraging bad behavior. He wanted to break the priesthood of people who could survive a segfault. The economic bet underneath was that a slightly painful early web would force phone companies to build broadband. Progressive JPEGs gave you something to watch while images crawled in; putting images on the web at all was a fight with people who said serious information is text and pictures would make it frivolous. Netscape shipped the first widely used consumer software with strong encryption, classified as a munition next to the Tomahawk, and spent five years fighting export rules that required a deliberately weak overseas build stamped "do not trust this." JavaScript, written by Brendan Eich over a summer, was designed as both front-end and back-end and is now the most widely used language in the world. SSL was Kipp Hickman, also one summer. The pattern he wants remembered is Bezos's two-pizza team, and for the really creative work, one pizza.

Netscape sold to AOL in 1999 for $4.3 billion at the peak of the boom, 18 months after founding, then watched the crash, the AOL–Time Warner disaster, and the depression out of which broadband, smartphones, Web 2.0, social, search, and SaaS grew. The economic lesson he later named is that software is a philosopher's stone: it transmutes labor into capital. Someone types, and a capital asset comes out the other side that someone else buys for a billion dollars, value from thought, a repudiation of Marx's industrial picture. Old software used to ship, saturate, and die. Minecraft, Mathematica, Facebook, Google, World of Warcraft, Salesforce now compound for decades if a team keeps polishing. Five billion people are a click away. Roll that theory out in 1999, he says, and people would have thought you were smoking crack.

The thesis of his essay "Why AI Will Save the World" starts from a predecessor we already have: human intelligence. Intelligence is how we capture, process, analyze, and synthesize information and solve problems. In human hands it is associated with better outcomes almost across the board — education, children's success, career, health, life satisfaction, peacefulness versus violence, open-mindedness versus bigotry. Fridman reads the longer list from the essay, from job performance and longevity through parenting and creative arts. If a machine can raise effective intelligence, the bet is that life gets better, not only in the aggregate but person by person. Doug Engelbart called it augmentation. Give everyone a 140-IQ assistant that is infinitely patient, knows them, and pulls for them — job interviews, confusing paragraphs, new skills — and you raise the odds. People at 140 get a peer; people above it get something to farm out. God willing, later versions go to 160, Einstein's estimated range, then 180, and you are talking about physics breakthroughs, cancer, warp drive. It may be the most important thing that has happened, because it is a lever on the single factor that drives so much else.

Fridman asks him to steel-man the other side: human plus AI is not always better for the individual. Andreessen is not a utopian. Some people get smarter and more arrogant. Smart people are better at marshaling facts to fit preconceptions — different-colored sheep, some better at justifying it. They also get themselves into situations beyond their grasp because their cognitive eyes are bigger than their stomachs. There are no panaceas. On net, if you extrapolate from what we know about human intelligence, you are still improving life by upgrading it. AI is the latest prosthetic after glasses, watches, shoes, PCs, word processors, Mathematica, and Google. It is the most powerful because it goes directly at fluid intelligence.

Andreessen splits the people who say they worry about AI into Baptists and bootleggers, an old development-economics pairing from American Prohibition. The Baptists were true believers — Carrie Nation with an ax in the saloon — and they were not wrong that alcohol tracked violence, workplace harassment, and abuse. The bootleggers wanted the ban because organized crime would mint money once legal sales stopped. The Baptists did not necessarily know about the bootleggers; the bootleggers certainly knew about the Baptists. Alcohol stayed illegal, people kept drinking, then the law was revoked, and a hundred years later the same problems remain. He says the pattern is repeating on AI, as it did on nuclear power, dozens of times in a century. Identifying motives is not enough. The arguments still have to be taken on the merits.

Risk number one: will AI kill us all? He objects first to the vocabulary shift from AI to AGI, which he thinks moves the topic from science to religion. The West, saturated in secularized Christianity, keeps inventing end-of-the-world stories to fill what C.S. Lewis called the God-shaped hole. Cults form around a coming crash and an elite who will see it, stop it, or survive it. Communism and Nazism were forms of this. True believers left to themselves self-radicalize. They might sometimes be right. The steel-man and its rebuttal are the same sentence: you cannot rule out that this ends everything, and you have therefore made a claim with no hypothesis, no test, no falsification, no way to measure progress. Fridman offers paperclip maximizers and thermodynamic limits. Andreessen grants the practical version — where does the evil AGI get the GPUs and the energy, in secret? — and then says the deeper objection is that this is all forecasting, not science. Extraordinary claims require extraordinary proof. The policies being demanded, including talk of military airstrikes on data centers, are of extraordinary magnitude. Millenarian movements, he says, have a hard time staying away from violence, because if you actually believe the world is ending you would do anything.

A pause-and-regulate response still needs a scientific approach to the future, and his COVID lesson is that the models did not work. Imperial College code, in a friend's reconstruction, was the worst spaghetti he had ever seen. Policymakers were handed doomsday scenarios with no empirical foundation. Modeling a complex system with eight billion moving parts may not even qualify as science. The AI-risk people, he says, do not even have the COVID model. Ask how you know a runaway takeoff is happening and the answer on offer is that the training loss drops — which is also what happens when training is working. Nick Bostrom's *Superintelligence* catalogued routes to machine intelligence, did not include large language models because they did not exist, then treated every route as equivalent risk. That whole apparatus is now being transplanted onto a different technology. On autonomous weapons Fridman wants limits on fully autonomous strike drones; Andreessen wants the opposite. Machines will make better decisions than human pilots under stress, the way a self-driving car does not need to be perfect, only better than the driver. The slight of hand he hates is the system smart enough to evade us and build secret factories yet too stupid to notice it is about to turn us into paperclips. You can already ask LLMs to resolve moral dilemmas. They do not chant destroy-all-humans. They give nuanced, trade-off-oriented answers.

The essay's nuclear aside is the historical parallel he wants for AI. After the Trinity test, Oppenheimer quoted the Bhagavad Gita; John von Neumann's reply to the hand-wringing, which Andreessen cites, was that some people confess guilt to claim credit for the sin. Truman, via Dean Acheson, did not want "that cry baby" back in the White House. Oppenheimer then spent years on television and in Washington in a hair shirt. The deeper drama, Andreessen says, was not only whether nukes were good or bad, but whether Russia should have them. Russia got the bomb through espionage. The composite of the histories is that Oppenheimer probably did not hand over the secrets himself, but he was close to people who did, including family, and other Manhattan Project figures who were Soviet agents did. The ethos that this thing was awful fed a Baptist conclusion that the principled act was to give Moscow the bomb. We live in a different second half of the twentieth century because of that.

What actually happened, in his reading, is that nukes prevented World War III through mutually assured destruction. Military planners expected a hundred million dead on the plains of Europe. That dog did not bark. The other counterfactual is America with the bomb and Russia without, and a possible rollback of the Iron Curtain. He does not claim to know that would have happened. He does claim that people very full of themselves, forecasting how the technology would be used, made a horrific mistake if they lengthened the curtain by fifty years. Von Neumann, hero of computing, advocated a first strike on Russia before it got the bomb, on the theory that World War III was inevitable. Most case studies of experts-gone-political are of the left; von Neumann is the right-hand version. The general principle Andreessen wants is that nothing in Oppenheimer's or von Neumann's background qualified them as moral authorities.

Who, then, should reason about AI? The track record of senior scientists and technologists making moral judgments about their own inventions is, in the book he gave Fridman — Neven Sesardić's *When Reason Goes on Holiday* — catastrophically bad. They have spent their lives in labs. They are thinly informed on history, sociology, theology, and ethics, and they manufacture worldviews from scratch. Fridman, playing devil's advocate, says corporate ethics departments often go the other way: outrage activism without history or humility. Andreessen agrees both are bad. Societies arrive at truth through some mix of elected leaders, public intellectuals, and the rest of us. People who bring rationality, judgment, and humility are few. We should prize them. Measuring a warning's historical contribution, in this frame, is not whether it "called" the future. It is whether it changed the result — toward regulation that actually buffers, or toward a hair-shirt politics that hands the other side the bomb.

Risk two in the essay is that if the murder robots do not get us, hate speech and misinformation will. He reads it as part two of the last decade's war over social media, which flipped from "who wants a photo of the cat's breakfast" to the most corrosive technology ever invented, then narrowed onto those two topics. The same activist energy is being transplanted onto AI, which is why ChatGPT answers some questions and not others, and why "as a large language model I cannot" means someone reached in. The x-risk people started with "AI safety," watched it become "alignment," and then watched the field get taken over by the hate-speech cohort; they have since rebranded as "not kill everyone-ism." Would he like a world where nobody is mean and everything is accurate? Sure. Does he want a centralized thought police inside tech companies enforcing a small elite's view of what the rest of us think? Absolutely not. The slippery slope, from the social-media version, "is not a fallacy, it's an inevitability." People get drunk with power. His architectural preference is client-side: a heavily filtered assistant for his eight-year-old, not a server-side dictator for the planet. Elon's Community Notes, which only fire when people who usually disagree agree this time, is the other hopeful pattern.

On bad people doing bad things he offers three parts: use AI defensively, including broad-spectrum vaccines and hunting terrorists; enforce the many laws already on the books; and count the cost of extra censorship, especially for open source. Preventing open-source AI legally and technically heads into monitor agents on every GPU, seized clusters, and, in the mouths of some x-risk figures, nuclear strikes. Nukes could be tracked via plutonium. AI is math and code, already in textbooks and YouTube videos, with a 40-billion-parameter Falcon already downloadable from the UAE. Walk the logic of guardrails and you land in an authoritarian regime of thought and machine control so brutal it destroys the society you were protecting. Foundation models, he says, have to be fully open, because the world in which they are not is jackboots seizing GPUs. A 14-year-old in Indonesia with a quantization breakthrough is not someone you invade. AI will be like air. You cannot pull air back. Put the effort into defense — a permanent Operation Warp Speed for biological defense — not into hand-wringing.

Crippling inequality is the Marxist worry that owners of the machines take everything. It was wrong about every previous wave because the self-interested way to get rich is to sell to the largest market, which is everybody, which is why ChatGPT and Bing went mass-market immediately, free or twenty dollars a month, rather than to Exxon first. You cannot buy a million-dollar iPhone; Apple makes more driving the price down. Jobs are the lump-of-labor fallacy: technology drops prices, consumers gain spending power, new demand creates new enterprises, more jobs at higher wages. Truck driving, in the town he grew up in, shortens lives. New jobs are often better, training is the pain point, and an AI assistant is precisely a tool for picking up skills. In 2019, before COVID, the world had more jobs at higher wages than at any point in history. The risk he is actually convinced by is that China wins. The Chinese Communist Party has a public plan for authoritarian population control, social credit, and a Digital Silk Road rolling its AI out on Huawei fabric worldwide. They are perhaps a year behind and download Western work regularly. Their GPT-3.5 analog, run through the Chinese SAT, did very well on Marxism and Xi Jinping Thought. A kid in Argentina or Germany ten years from now asking how the economy works may get the cheeriest lecture on Chinese communism ever written.

Over twenty years, Andreessen says, tech shifted from a tools industry — PCs, databases, smartphones as pick-and-shovels from 1940 to about 2010 — to applications that go straight at a customer, Uber and Airbnb style. AI is a temporary reversion to cloud APIs, but most of the money will likely sit in the AI doctor, lawyer, and financial advisor, some sold into firms and some going direct. Advice for founders: the great ones do not take any. What they have is intelligence, energy, and courage, and courage is a choice about pain tolerance — how many times you get punched in the face, told no and that you are stupid, while pretending you are having a great time. On a risk-adjusted basis it is often irrational. The previous boss's honest reference on a Steve Jobs is that he was a terrible employee. Companies work better when the idea exists first, a product people will pay for, often after five or ten years of chewing on the problem as a domain expert, not a shower thought and thirty slides. The ideal pitch is: here is the thing that works. Do it because you cannot tolerate not doing it. It will cost outside friendships and 80-hour weeks. Older founders know more and have baseball games they cannot miss.

His method of learning is autodidactic: go broad by job, then disappear down a rabbit hole for a while and maybe not return for a decade. After 2014–2016 he withdrew from political fundraising and read as far left and right as he could. The single best book he has found on who we are is Fustel de Coulanges's *The Ancient City*, written in the 1860s from original Greek and Roman sources, reconstructing Indo-European life as nested cults of family, tribe, and city at an intensity we would not recognize, with no concept of individual rights, fascist in social control and communist in economics because winter was not a debate society. His conclusion is that we are still running that original model at a thousandth or a millionth the intensity. Modern Christianity, sports teams, Apple keynotes, and political movements are washed-out cults. People then knew exactly where they stood and why they got up. Turn the volume down and the search for meaning gets harder, which is why we reach for drama.

Advice to the young is that the tools are a million-fold beyond the card catalog, so the puzzle is where the hyper-productive people are — authors of hundreds of books, musicians of a thousand songs. His suspicion is distraction, the ease of sitting and consuming. Pliny the Elder traveled with slaves to read to him and take dictation so he could write through the day. We have fire and are still figuring out how to cook. He does not believe in balance; he is happier all-in. Happiness is an ice-cream cone; satisfaction is using your faculties, making the world better, looking back at something hard that was worth it. Money applied to happiness leads down dark paths; applied to satisfaction it is a tool. Elon, in his view, is back-to-the-future: Ford, Carnegie, Rockefeller, hands-on, first principles, ground truth from engineers, a magnet for resentment. The closest Andreessen gets to the meaning of life is that we were given what we have and should do our best, with love in the mix. David Friedman's triad is love, money, and force as the three ways to get someone to do something for someone else. Bet on the first two.