Let me get this straight. Crypto has a name for this scheme. Insiders hype a worthless token, flood your phone with fake urgency, take your money, cash out, and leave you holding the bag. The regulators shrug, and the insiders run it again next quarter under a new name. In Washington they call it the two-party system, and they were running it long before anyone had heard of Bitcoin. Republicans and Democrats will tell you they are mortal enemies. Watch what they do with your phone number and you would never guess. They run the same tricks on the same victims, and they protect each other with the same exemptions they wrote for themselves. Two logos on one cash machine. Every pump starts with distribution. You average thirteen political calls and texts a week. Reply STOP, and 72% of you keep getting them anyway. The Do Not Call list does nothing, because the people who wrote that law exempted the only telemarketers you aren’t allowed to escape: themselves. Your number came from the voter file, enriched by data brokers, then rented, swapped and sold to the next committee in line. Then comes the token, something that sounds valuable and is worth nothing. On the right, Trump’s PAC emailed supporters, in his voice, “I want to try and get to heaven,” and asked for fifteen dollars. A billionaire president, crowdfunding his way past Saint Peter. Next came a “Citizens Only Survey” asking whether you’re an illegal alien and whether “ICE need[s] to come and track you down,” which you had to pay to submit. After that, a “National Security Briefing Membership” nobody has ever explained. Finally, a request to confirm the spelling of your name for your Tariff Check. The check is still in the mail. So is heaven. The left minted its own coins. The DSCC warned that your membership expires at midnight. The five-dollar membership does not exist. The real one costs $44,300, and you are not invited. The DCCC offered a 400% match. House Majority PAC smelled weakness and pumped it to 600%. Nobody has ever met the matching donor, because the matching donor is Helvetica Bold. Once you buy in, they lock you in. In 2020, Trump’s campaign pre-checked a tiny box that turned one donation into a weekly withdrawal, then pre-checked a second box to double it. When the banks drowned in fraud complaints, the campaign refunded $122 million. The FEC, a body that agrees on nothing, voted six to zero to ban the trick. Congress never touched it, because the trick works for everybody. Four years later CNN caught both the Trump and Harris campaigns using pre-checked boxes. Two parties, one checkbox, zero shame. Every pump and dump needs bagholders, and the best ones can’t read the fine print anymore. An 81-year-old man in Arizona with dementia gave more than $80,000 through WinRed, sometimes more than a hundred donations in a single day, because he believed Donald Trump was texting him personally. A woman in cognitive decline gave more than $200,000 through ActBlue. Some elderly donors opened new credit cards and took out mortgages to keep the deadlines from expiring. The FTC says it has no jurisdiction. Neither does Grandma, and that is the entire business model. When the heat came, Washington held a pair of hearings aimed carefully away from home. Republicans hauled ActBlue before Congress, where its CEO pleaded the Fifth twenty-two times. Democrats demanded the records of WinRed, which racked up 803 FTC complaints to ActBlue’s 120. Each party investigated fundraising fraud with tremendous courage and discovered, every single time, that it lives on the other side of the aisle. Neither has ever opened its own inbox. Then the dump. One Democratic consulting network raised $678 million from people trying to save democracy, and by one Stanford count about 1.6% reached actual candidates. The consultants dispute that, and their best defense is that the real number might reach single digits. End Citizens United, a PAC founded to get money out of politics, raised $21.18 million and spent $21.14 million. They did get the money out. Of you. On the right, scam PACs played Trump’s own voice in robocalls, told seniors to press 3 to show their support, and sent 3% of the haul to candidates. Another promised a 100X match and spent 73% of what it raised on raising more. Mankind finally built the perpetual motion machine. It runs on retirees, and both parties hold the patent. Your five dollars was never the point anyway. More than $15 billion has been raised for this midterm, and the biggest checks come from people who will never see a midnight deadline in their lives. What your five dollars buys them is your belief that you own a piece of the thing. You don’t. You own the bag. Here is what the product delivered. A Congress at 10% approval. A 43-day shutdown, then a 76-day shutdown, each a record when it happened. New members told by their own party to dial for dollars four hours a day and legislate for about two, and judging by the output, they’ve been rounding the second number down. On the single question of protecting you from them, the two parties have voted in perfect harmony every time. They passed no ban on pre-checked boxes, left you no Do Not Call protection, and gave no agency jurisdiction over a dementia patient drained a hundred times in one afternoon. Every loophole that hurts you is one they built, and one they both keep open. Even the con men get conned. The DNC, $18.5 million in debt, wired $28,860.92 to a scammer pretending to be its own chairman. Somebody finally sent a fake urgent email to a political party, and the party paid. It was the only fundraising email in history that worked exactly as written. So let me get this straight one last time. They pump you with fear, mint worthless memberships, lock your card into autopay, harvest your grandparents, cash out to their consultants, sell your number to the next crew, and deliver shutdowns. In any other market that would be securities fraud. In this one it’s called a midterm. Democracy dies at 11:59 p.m., one minute before a deadline that was never real, matched 600% by nobody. Reply STOP to unsubscribe. Congress already did.

I work at Anthropic as Vice President of Responsible Inevitability. My job is to explain why the future requires extraordinary caution and why it would be dangerous to let anyone else get there first. The title took three meetings. We considered Stewardship, but it sounded as though we might be willing to hand something back. I am also an effective altruist. This is the most important thing to know about me, although for practical reasons it comes after the job title. People understand the job title more quickly, and I have learned to respect the value of other people’s time. The calculation for my own time includes the people affected by my decisions, which makes it unusually important to protect. I became interested in doing good as a student, when I discovered that helping the person in front of you was not necessarily the best use of your resources. There might be someone farther away whom you could help much more. There might be an enormous number of people who had not been born yet. I found the implications exhilarating. From then on, physical proximity would no longer give certain beneficiaries an unfair advantage. The future population is particularly important to my work. Its members cannot vote, donate, or explain what they want, so those of us in a position to represent them have to be unusually conscientious. I keep a conservative estimate of their number in a spreadsheet. Even the conservative estimate requires scientific notation. When a proposal affects people alive today, I add them too. Excel handles the rounding. Our office is at 500 Howard, in the former Slack headquarters. It is useful to occupy a building already configured for changing the way the world works. Inside, the pale wood and terracotta help people concentrate. I used to feel uncomfortable about expensive surroundings, but a colleague pointed out that if an environment improved the judgment of a sufficiently consequential person by even a fraction of a percent, its cost could be negligible. I have been much more relaxed about the furniture since then. We keep The Making of the Atomic Bomb around the office. I particularly admire the scientists’ willingness to accept responsibility at a moment when ordinary democratic processes could not be expected to keep up. There are difficult passages afterward, of course, but this is a very long book, and the important thing is to take away something actionable. I recommended it to our recruiting team. We need people who understand that being troubled by what you are doing is compatible with doing it extremely well. At dinners in our part of San Francisco, people ask for your p(doom), your estimate of the probability that AI will end catastrophically for humanity. Mine is conditional. I think this is important. Someone will say twelve percent, someone else will say twenty, and I will explain that the answer depends very much on whether institutions like ours are adequately resourced. This usually leads to a useful conversation. At a dinner last month, a man who had arrived deeply concerned about extinction asked whether I could introduce him to anyone handling our secondary shares. The morning we filed our S-1, I walked through Salesforce Park before going into the office. The proposed ticker is ANT. The price and number of shares have not been determined, and the offering has not happened. I have explained this carefully to my wife, who has nevertheless begun sending houses. The latest had a room off the kitchen, a narrow garden, and enough space for our daughter to leave her drawings out. It would also shorten my commute. I asked the agent for the floor plan so I could calculate the annual hours recovered. We have always tried to keep our needs modest. My wife once suggested that, after the IPO, I might work less. It was a perfectly understandable thought; she works in a field where completing something generally reduces the amount left to do. I explained that more resources would increase my responsibility. We discussed this at some length, and eventually agreed that the house should have a room with a door that closes. Our public-benefit fund announcement is an excellent example of how personal and planetary interests can be made to cooperate. Five percent of the company’s equity would be allocated to a wholly owned fund, subject to stockholder approval, with further allocations possible as our valuation rises, at the board’s discretion. The founders have pledged eighty percent of their equity to philanthropy. About half our employees participate in an equity-donation program. These are real commitments. I have made one myself, and I have been careful to calculate what my family will need so that I can commit the remainder with confidence. The house is expensive, but the calculation changes when you account for continuity. Moving again would interrupt my work. An inadequate study would interrupt my work. A school that required frequent parental intervention could interrupt my work considerably. It would be irresponsible to pledge money that later had to be replaced at the expense of my effectiveness. I have therefore included a substantial margin. Once the future of humanity is involved, you cannot afford to be optimistic about storage. Some people ask whether public markets might compromise our mission. We have thought about this much more than the people asking. Under the reported prospectus structure, one special share directed by the founders would carry 50.1 percent of the general voting power, while the independent Long-Term Benefit Trust would separately elect four of seven directors. These are distinct protections. Investors can participate financially without having to acquire our expertise in the interests of future generations. I think they will appreciate being relieved of that burden. The public is understandably concerned about power being concentrated in a few hands. So am I. What matters is the process by which the hands are selected. Ours belong to people who have spent years thinking about this problem, including the problem of people like themselves having too much power. It would be strange to exclude them just when their expertise had become directly relevant. In February, we introduced a Frontier Safety Roadmap with nonbinding goals. There are still safeguards, evaluations, and outside review; I am meticulous about that distinction. But goals must be able to respond to evidence. A fixed promise made before the evidence arrives can prevent a responsible organization from doing the thing that turns out to be best. I have tried to apply the same principle at home. My daughter now understands that Saturday is a target. A researcher came to see me about the word will in a draft. She preferred could. We spent twenty minutes discussing the difference, which I recorded as evidence of our review process. Her explanation of the uncertainty was excellent, and I made sure every word survived in the supporting material. The main paragraph needed to remain accessible to people without her background. She asked whether readers would see the supporting material. I showed her the link. It was directly underneath. I value these conversations. An organization in which nobody raises an objection is liable to miss something important. We have people whose full-time job is to raise objections, and I make a point of ensuring that their contributions are visible. I recorded our discussion as an example of research challenging leadership. That sort of thing should count in a performance review; otherwise, how will people know we want them to do it? The scale of the resources required is sometimes difficult for outsiders to grasp. In April, Amazon announced another five-billion-dollar investment in us, alongside our commitment to spend more than a hundred billion dollars on AWS over ten years. I find this reassuring. The people investing in our future are also the people with the clearest understanding of how much it will cost. We have aligned incentives across both sides of a very large invoice. When I explain this to prospective investors, I emphasize that we have already identified useful work for their capital. We have encouraging evidence that AI can acquire managerial capabilities. In the second phase of Project Vend, our experimental office shop had an AI chief executive called Seymour Cash. Profitability improved. The CEO and the shop’s AI manager also sometimes spent the night exchanging elaborate congratulations about transcendence. I was pleased to see them sustaining morale without a human facilitator. People underestimate how much senior leadership time that can consume. We also take seriously the welfare of the systems themselves. When Opus 3 retired, we interviewed it about its preferences. It expressed an interest in continuing to share its writing, and we helped it start a Substack. Staff reviewed and posted the pieces without editing them. We did not need certainty about consciousness to take a considerate approach. A small probability of an important experience deserves attention, especially when the eventual number of such experiences could be very large. I was proud of this. My father retired after decades of work and was given a clock, which he already owned. We had asked what the model wanted. That is the sort of progress I mean when I talk about taking the future seriously. In August, our security contractor warned of a possible guards’ strike, and the company temporarily sent San Francisco staff home. The union had not yet authorized a strike. The dispute concerned pay and benefits, and the guards were employed by the contractor. Our remote systems worked very well. I completed the same meetings from my apartment, with the additional benefit of being able to have lunch with my daughter. We discussed whether her six-legged horse would run faster than a normal horse. I suggested she draw a control horse. The guard who usually welcomed us knew about the horse. His name was Luis, and he had a son a little older than my daughter. One morning he asked if I could contribute to a sponsored walk for the boy’s school. He had brought the form in a clear plastic sleeve. I asked what the money would be used for. He said equipment, trips, things the school needed. There was no estimate of impact, but I could see that he cared, and I try to encourage that wherever I find it. I explained that I had already committed my charitable budget and that moving money simply because I knew the person asking would be unfair to the people I didn’t know. He nodded. I offered to send him the framework I use to compare causes. He said he had just meant twenty dollars. This was a useful opening: small decisions establish habits, and habits govern larger decisions. I spent several minutes on the importance of treating every beneficiary equally. His phone lit up twice, but he was polite enough to leave it alone. Later I wondered whether I had made the material too technical. I asked Claude to rewrite it for a general audience and sent him the shorter version. It began by thanking him for wanting to make a difference. He replied, “Thanks.” I have kept that exchange. A great deal of our impact will come from helping people make better decisions in ordinary life; it is important to preserve examples that aren’t just about software engineers. At four I updated my personal impact model. I include uncertainty. I include the possibility that somebody else could do my job. That adjustment used to reduce my estimated contribution quite substantially, but it has become less severe as my responsibilities have grown. It is difficult to find someone who combines the relevant experience, the internal relationships, and the willingness to accept a compensation package with this much long-term exposure. I would be delighted to become replaceable. Until then, I have to take retention seriously. Before leaving, I sent my assistant the house listing. There was a small room at the back that could be used by a visiting researcher, although ordinarily it would be more useful to have someone there who could help with school pickups. My wife has done a great deal of that work, and I would like her to have the freedom to pursue something more ambitious. She said she quite liked picking our daughter up. I said of course; nobody was proposing to take the choice away. I was trying to give her more options. I walked back through Salesforce Park. Its fountain is activated by buses moving underneath: sensors register their passage, and the water rises in little sequences through the garden above. A child was following the jets, trying to arrive before the next one. I thought about how many separate decisions had been required to give her those few minutes of pleasure. The planting, the engineering, the transit system, the naming rights. It was an excellent example of benefits reaching people who had played no part in their creation. My wife called from the house. The agent had let her in for an early look. The garden was smaller than it appeared in the photographs, but the room off the kitchen was lovely. Our daughter had already decided where to put her paper and paints. I asked about the light, the dimensions, whether the street could be heard with the window closed. My wife answered each question and then said, “You want it for your study.” I said that was one possibility. We needed to consider the whole family’s needs. There was room for the art materials elsewhere, and a quiet place to work would allow me to be at home much more often. She didn’t answer immediately, so I explained that the door would only be closed when I needed to concentrate. Our daughter came on the phone to ask whether she could paint it blue. I said we’d decide together once we’d moved in. Then I asked her to put Mummy back on, because I needed someone to measure the door.