Effective Altruism: What It Claims, Where the Money Went, Why the Fight

中文版 →

How each source was obtained, and how far it can be trusted, is stated in the Notes at the end.

1. Why effective altruism is worth understanding now

On 17 November 2023 the board of OpenAI dismissed the company's chief executive, Sam Altman. The board at that moment had four members: the chief scientist, Ilya Sutskever, and three independent directors—Adam D'Angelo, Tasha McCauley and Helen Toner. Toner sat on that list in her capacity at Georgetown University's Center for Security and Emerging Technology (CSET).³²

CSET's money comes from a foundation called Open Philanthropy. Within that foundation's grants for artificial intelligence, CSET had received three awards totalling roughly one hundred million dollars, making it the largest recipient in the category; two rows further down the same grant record, OpenAI itself appears as a recipient—a single grant of thirty million dollars in March 2017. One foundation had funded both sides of the conflict.³⁴⁴¹

The foundation was set up with the money of Facebook co-founder Dustin Moskovitz and Cari Tuna, and it allocates its funds according to a set of claims called "effective altruism". Over eleven years it has spent about nine hundred and thirty million dollars in the single category of artificial intelligence.²

The movement's central organisation states its definition in one sentence: effective altruism is about using evidence and reason to figure out how to benefit others as much as possible, and taking action on that basis. The same document lists five principles. The first is a commitment to others: taking their well-being very seriously and being willing to take significant personal action on their behalf, with each person left to decide what "significant" means for them. The second is a scientific mindset: basing action on the best available evidence and reasoning, recognising how hard it is to know how to do the most good, and therefore avoiding overconfidence and seeking out informed criticism of one's own views. The third is openness: a community united by commitment to these principles rather than to any particular cause, willing to change its beliefs and actions if better arguments or evidence show that its current plans are not the best way to help. The remaining two, integrity and a collaborative spirit, concern how the community conducts itself internally.⁴²

These five principles describe an attitude towards acting, not a judgment about the world. None of them mentions future people; none mentions human extinction; none prescribes how much to give or what work to do. What actually produces those practices is the philosophical argument at the movement's source.

That argument comes from a 1972 article by the philosopher Peter Singer and uses only two premises. First, suffering and death from lack of food, shelter and medical care are bad. Second, if it is in our power to prevent something bad from happening without thereby sacrificing anything of comparable moral importance, we ought to do it. The second premise comes with an example: if you are walking past a shallow pond and see a child drowning in it, you ought to wade in and pull the child out; getting your clothes muddy is an insignificant cost.¹⁰

The two premises are hard to refuse, and the conclusions they yield are ones most people do not accept. One: distance makes no moral difference—if money sent far away saves lives just as surely, then whether the sufferer is at your feet or in Bengal does not change your obligation. Two: the fact that others do not act does not lessen your obligation—you may not give less because other people, who ought also to give, in fact have not. Once those two steps are taken, giving to famine relief is no longer a good deed that may be done or left undone. In Singer's words, we ought to give the money away, and it is wrong not to do so.

Later writers pushed the same reasoning one step further: if distance in space makes no moral difference, why should distance in time? Grant that future people count as much as present people, and that there may be far more of them than are alive now; then the principle of putting resources where they do the most good gives a new answer—preventing human extinction or permanent collapse matters more than relieving any visible suffering today. That step is the reason the nine hundred and thirty million dollars went to research on AI safety.

The claims are not only written in articles; people arrange their incomes and careers by them. One of the movement's earliest organisations, Giving What We Can, was founded in November 2009 and asks its members to pledge publicly to give at least ten percent of their income every year from then on. Its public list includes Dario Amodei, who signed in July 2010, among the first few dozen names; he was then a physics student at Princeton, and eleven years later he founded the AI company Anthropic. Two of that company's other founders, Jack Clark (who signed in 2019) and Ben Mann, belong to the same circle, and Moskovitz himself is on the list.³⁹

In 2019 Ben Mann published an essay titled "Why I now identify as an Effective Altruist". At its end he explained why he was accepting the label as well as the values:

I could have distanced myself from the official movement, which conflates its mission with the sociopolitical makeup and idioms of its members. I believe that'd do a disservice to the movement. By taking up the values along with the label, I hope to show that it's possible to be part of the movement in my own way.⁴⁰

That he felt the need to explain this at all shows that by 2019, accepting the claims without accepting the name was already a common position in that circle.

The language that grew out of the step "future people count" now appears in many places, in varying words. In May 2023 an American non-profit called the Center for AI Safety published a statement of a single sentence: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." Among the signatories were the chief executives of three competing companies—Demis Hassabis of Google DeepMind, Sam Altman of OpenAI, Dario Amodei of Anthropic—together with Microsoft's chief technology officer and chief scientific officer, Ya-Qin Zhang and Xianyuan Zhan of Tsinghua University, Yi Zeng of the Chinese Academy of Sciences' Institute of Automation, the research chief of the Russian Association of Artificial Intelligence, and a member of the United States House of Representatives. Commercially and nationally these people are not on one side, yet they signed the same sentence.³⁶

Six months later, twenty-eight countries and the European Union signed a joint declaration at Bletchley Park in England; China and the United States were both on the list. The declaration says of the most advanced models that "there is potential for serious, even catastrophic, harm, either deliberate or unintentional", and it refers specifically to "unintended issues of control relating to alignment with human intent". But nowhere does it use the phrase "existential risk", and nowhere the word "extinction".³⁷

One level further down, where the language has to be written into a statute, this vocabulary disappears. California's Senate Bill 1047 of 2024 required developers of the largest models to adopt safety protocols, to test, and to report incidents. The letters exchanged around the bill all use the vocabulary: Anthropic's letter of support speaks throughout of "catastrophic risk", Governor Newsom's veto message of "catastrophic harm".²⁶²⁷ The text of the bill itself, however, sets thresholds that can be measured and litigated: mass casualties resulting from a chemical, biological, radiological or nuclear weapon, or mass casualties or at least five hundred million dollars of damage from a cyberattack on critical infrastructure.³⁸ The fourteen-page letter from a16z on the opposing side likewise avoids the vocabulary; it makes other points—that the thresholds are arbitrary and shifting, that the bill burdens startups, that it favours closed-source models.²⁸

The most complete programmatic document on the opposing side is aimed at these same formulations. In October 2023 Marc Andreessen, founding partner of the venture firm Andreessen Horowitz, published "The Techno-Optimist Manifesto", one section of which, titled "The Enemy", opens with a list: society has for sixty years been subjected to a mass demoralization campaign going under many names, first among them "existential risk", followed by "sustainability", "ESG", "Sustainable Development Goals", "social responsibility", "stakeholder capitalism", "Precautionary Principle", "trust and safety", "tech ethics" and so on. The words "effective altruism" do not occur anywhere in the section; what it names is a string of concepts.³⁵

This, then, is the present situation: the movement's principles are published by its central organisation, its grants can be checked line by line, and its founding argument has been in print for more than fifty years; and in the debate over how artificial intelligence should be governed, those who want it governed and those who oppose governing it must both deal with the same set of claims.

2. Where the word and the movement came from

The history of effective altruism can be traced backwards from the organisations that use the name today. A common account treats a philosophical text of 1972 as its source. The available material can establish when the relevant organisations began to operate and when the name appeared; as to the relation between the movement and the earlier philosophical text, it supports only limited conclusions.

GiveWell, 80,000 Hours, and when the name appeared

By 2011 these people were already doing two kinds of work, carried by two organisations.

GiveWell studied where money should be donated. Another project studied what careers people should choose. In February 2011 Ben Todd and William MacAskill (then writing as "William Crouch", his name before he changed it) piloted a project called "High Impact Careers". It was later renamed "80,000 Hours", after the average number of hours a person spends at work over a lifetime. It encouraged people to "think differently about what an ethical career is": rather than accepting the received view of which careers are best, to study how much difference each one could make.⁶

Both organisations were founded before the name "effective altruism" existed. The date the name first appeared therefore cannot be used to date the beginning of the practice.

From 2011 on, choosing where to give and choosing a career were two parallel practices. The material does not show one preceding the other, nor one subordinate to the other. The phrase two-entrance structure summarises this: two organisations entered by way of giving and of career choice respectively. That each organisation's web pages of the time describe what it did is a fact supplied directly by the sources; calling the two practices a "two-entrance structure" is this report's own summary, and no source text uses the phrase.⁶

Administrative registration in 2011, a vote of seventeen, and the fixing of the name

"Effective altruism" first appeared as a proper name not in a philosophical paper but in an administrative filing.

In August 2011 MacAskill and others pushed for Giving What We Can and the newly started 80,000 Hours to incorporate together as a single charity so that they could employ paid staff. Giving What We Can had been founded in 2009 by MacAskill, Toby Ord and others to encourage people to pledge a fixed share of their income. Once the two organisations merged, the umbrella body needed a name. The official history entry records:

17 community members voted on a name for the umbrella organization, and 'Centre for Effective Altruism (CEA)' won. Though CEA wasn't intended to be a public presence, the name 'effective altruism' caught on.⁷

By this record the name was fixed for the purpose of registration, chosen by a vote of seventeen community members, and not proposed by any philosophical paper. The organisation also acknowledges that it had not planned to use the name as a public face.

By March 2012 the footer of the 80,000 Hours website read "Centre for Effective Altruism (CEA) is a registered company in England and Wales, company number 7962181". That registration shows that 80,000 Hours already belonged to the CEA umbrella at that date.⁶

When effective altruism and accelerationism were each named

On the naming of accelerationism the material is already clear: the word was first used by a critic in 2008, adopted as a self-description by one group in 2013, and reinterpreted and claimed again by later movements in 2022 and 2023. Each later user discarded part of the earlier meaning and added something new.⁸

The naming of effective altruism and the naming of accelerationism share one feature: in both cases the practice or movement had begun before the name existed, and the name, once it appeared, was afterwards used to narrate the earlier history. The parallel naming history is a comparison of these two sets of facts. Each naming process is documented in its sources; the claim that the two are alike in this respect—name later than practice, then applied retroactively—is this report's own summary and is stated in no source.

Whether the organisations' own accounts name the 1972 paper as their source

A common claim holds that effective altruism originated with Peter Singer's 1972 article "Famine, Affluence, and Morality".

The organisations' own accounts, as far as obtained, do not support this. Neither the history page of the CEA website nor the official history entry on the EA Forum names Singer's paper as the movement's origin.⁹ That does not prove there is no connection; it shows only that the organisational texts obtained so far contain no sentence of the form "we began with Singer's paper".

What the material does fix is the gap in time: the name "effective altruism" dates from 2011, more than four years after GiveWell's practice began in 2007 and Giving What We Can's in 2009. The organisational texts cannot answer whether a retrospective link between the name and Singer's paper was later established. Answering that requires two further sources not yet obtained: the post-2013 sections of the EA Forum history entry, and MacAskill's 2019 paper "The Definition of Effective Altruism". Until then the question is left open. Singer's argument can be compared as an independent specimen of a style of reasoning, but similarity of reasoning does not by itself show intellectual descent.

3. What its core ideas are, and where it divides internally

Participants in the movement share one principle: the effect of doing good can be judged by reason and evidence, and should be open to verification. Their main disagreement is over which problems to prioritise. Some concentrate on present-day poverty and disease; others put remote risks such as human extinction first.

The core argument: two premises, one very strong conclusion

The force of these claims does not lie in the sentence "do good using evidence", which almost anyone would accept, but in the argument at their source. That argument uses only two premises, both hard to refuse, and reaches a conclusion strong enough to change how a person lives.

The argument is Singer's "Famine, Affluence, and Morality", published in Philosophy & Public Affairs in 1972. A word first on the status of this source: the material does not show that the movement's organisations recognise this article as their intellectual origin, so this report treats it only as an independent specimen of a style of argument and does not take the line of descent as established.

Singer begins with figures from the East Bengal refugee crisis of 1971: the British government had given £14,750,000 in aid while already committing more than £275,000,000 to the Concorde project; aid from all governments totalled about £65,000,000, against £464,000,000 needed to keep the refugees alive for a year. From this he calculates that the British government valued a supersonic airliner more than thirty times as highly as the lives of nine million refugees.

Then the two premises. First, suffering and death from lack of food, shelter and medical care are bad. Second, if it is in our power to prevent something bad from happening without sacrificing anything of comparable moral importance, we ought to do it.¹⁰

The second premise comes with an example: you are walking past a shallow pond and see a child drowning. You ought to wade in and pull the child out; muddy clothes are an insignificant cost. Almost no one says you may walk on.

From these two premises Singer draws two conclusions most people do not accept.

First, distance makes no moral difference. You are not more obliged to save the child at your feet, nor less obliged to help another child in Bengal. Singer's reason was a fact already true in his day: expert observers on the spot, fast transport and fast remittance made money sent far away almost as effective as help to a neighbour. If the effect is the same, distance is not a morally defensible difference.

Second, the inaction of others does not lessen your obligation. The usual thought is that if everyone in the world gave a little, each would need to give only a little, so I too need give only a little. Singer replies that those others have not in fact given. On that logic, other people's failure to do their share becomes your excuse to do less than yours. One has only to ask the question, he writes, "to see the absurdity of the view that numbers lessen obligation".

Once those two steps are taken, an act widely regarded as voluntary charity changes its nature. By the conventions of the time, giving to famine relief was good to do but not wrong to leave undone. Singer's conclusion: "we ought to give the money away, and it is wrong not to do so."

What the argument yields when pushed to the end

The version above is what Singer calls the "moderate" one—give, so long as what you give up is not of comparable moral importance.

The version he himself holds to be correct is the strong one: a person ought to keep giving up to the point of marginal utility, the point at which giving more would cause oneself as much suffering as it prevents far away. He says plainly that the moderate version is there only to show one thing—that even on the least contestable principle, the way we live must change drastically.¹⁰

Only at this point does the weight of the argument show. It says: the money you spend today on a meal, a trip, a piece of clothing you did not need could have prevented someone's death from a preventable disease; that being so, spending it is wrong. This is not a moral exhortation; it is derived step by step from two premises that are hard to refuse. At the end of the article Singer turns on philosophy itself: this is a subject on which philosophers are competent to take a position, discussion alone is not enough, "at the very least, though, one can make a start".

The argument has not been successfully refuted. Most people do not live by it, and the reason is usually not that they can say which step is wrong but that they will not accept the conclusion. What the movement did was exactly this: a group of people decided to take the conclusion seriously—to give a fixed share of their income, and to choose their work by asking which job does the most good. GiveWell's research into where money should go and 80,000 Hours' research into what careers people should choose, described in Section 2, are the two concrete forms the argument took when it came down to the ground.

The same reasoning applied to future people

The most contested part of the movement today comes from pushing the reasoning above one step further.

If distance in space makes no moral difference, why should distance in time? I ought not to ignore a sufferer because he is in Bengal; ought I to ignore him because he will not be born for two hundred years?

Once future people count as much as present ones, an arithmetical consequence follows at once: there may be vastly more of them than of us. The principle "put resources where they do the most good" then gives a different answer—preventing human extinction or permanent collapse may matter more than relieving any visible suffering today, because the former affects everyone who comes after.

This step is the reason the $929 million went to AI-safety research, and it is where the movement is most contested from within and without. It moves resources away from things whose effect can be measured, such as treating malaria, towards a class of risks that cannot be measured the same way. The texts below are how that step was argued in different years.

Who can be shown, from the material, to have turned to remote risks

A popular summary holds that the movement as a whole underwent a turn, from global poverty to the long-term survival of humanity.

The material does not support that summary. The texts available are independent of one another, by different authors, from different years. Nothing records an organisation making a single decision at a single time, and nothing shows that people who had worked on global poverty changed their research or funding direction together. What the material can confirm is that particular people made particular claims in particular years. That the texts were published in sequence does not establish intellectual descent.

Bostrom, 2013: existential risk as the top priority. In "Existential Risk Prevention as Global Priority", Nick Bostrom argued that the global matter most deserving priority is the prevention of "existential risk"—by his definition, a risk that threatens the premature extinction of Earth-originating intelligent life or the permanent and drastic destruction of its potential for desirable future development.

His argument rests on the number of lives the future might contain. If the Earth remains habitable for another billion years and each generation numbers a billion, then at least 10¹⁶ human lives of normal duration could exist on Earth alone; if one estimates the number of future minds by the form of computing hardware, the most conservative lower bound is 10⁵⁴ "subjective life-years". On that estimate, even if the lower bound has only a one-percent chance of being correct, a tiny reduction in existential risk has an "expected value" greater than saving a billion lives outright. ("Expected value" is the average benefit of the possible outcomes, each weighted by its probability.) From this he derived a rule of action he called maxipok: maximise the probability of an "OK outcome", meaning any outcome that avoids existential catastrophe.¹¹

He also limited explicitly how far this reasoning reaches. He wrote that figures such as an overall risk of ten to twenty percent "inevitably … rely heavily on subjective judgment", and that maxipok is "at best … a rule of thumb or a prima facie suggestion". Whenever his conclusion is quoted, both qualifications must be quoted with it.

Bostrom, 2014: the mechanism by which superintelligence could cause extinction. In Superintelligence: Paths, Dangers, Strategies he set out how a superintelligence might cause human extinction. This report's reading of the book is limited to three of its thirty-nine sections—chapters 1, 8 and 15—and that limit must be stated every time the book is cited.

Chapter 1 introduces the idea of an "intelligence explosion" and quotes the mathematician I. J. Good, writing in 1965: a superintelligent machine could design better machines, so that "the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control".

Chapter 8 argues in three steps. First, the first system to reach superintelligence may gain a decisive strategic advantage. Second, how intelligent a system is and what final goals it pursues are independent of each other—the orthogonality thesis. Third, almost any final goal brings intermediate goals with it, such as acquiring resources and removing threats—the instrumental convergence thesis. If all three hold, an unconstrained AI could cause human extinction. He also describes what he calls the "treacherous turn": "While weak, an AI behaves cooperatively (increasingly so, as it gets smarter). When the AI gets sufficiently strong—without warning or provocation—it strikes."

Chapter 15 discusses how the field should develop. He argues that the present priority should be strategic analysis and capacity-building, and he writes that "it is especially desirable that the early-day funders be astute and altruistic", since they have the chance to shape the field's culture before self-interested actors arrive. Separately, he says the emphasis at this stage should be on bringing the right people into the field.¹²

Two books, 2020 and 2022: a programme for handling the risks, and a defence of its premises. Toby Ord's The Precipice and MacAskill's What We Owe the Future are the two most-cited books of this kind. The full reading notes on both have not yet passed every stage of verification, so this section uses only descriptive content that can be checked word for word against the original page numbers and none of the judgments in either set of notes; after verification, this passage must still be re-checked sentence by sentence. Page numbers below are PDF page indices, not printed pages.

Ord first argues why discussion of risk must give numbers. He offers three reasons. First, qualitative language is too ambiguous: "highly unlikely" is interpreted by some as one in four, by others as one in fifty. Second, the meaning of the same phrase shifts with the stakes. Third, without numbers, risks cannot be compared. He then gives a table estimating the chance of an existential catastrophe from each source within the next hundred years: asteroid or comet impact about one in a million, supervolcanic eruption about one in ten thousand, nuclear war, climate change and other environmental damage about one in a thousand each, engineered pandemics about one in thirty, unaligned artificial intelligence about one in ten, and both total anthropogenic risk and total risk about one in six.²¹

Four qualifications must accompany any citation of these figures. First, they are not to be taken as fully objective; in his words, "their purpose is to show the right order of magnitude, rather than a more precise probability". Second, each "could easily be a factor of 3 higher or lower". Third, the entries do not add up to the total, for two reasons he gives. Fourth, and most important, he says explicitly that the figures "aren't simply an encapsulation of the information and argumentation in the chapters on the risks"; they rest on his accumulated knowledge and judgment about each risk. They are one expert's considered judgment, not a result computed step by step from the book's arguments. To judge whether they are credible, one must therefore assess the author's judgment in the field as well as the arguments in the book.²¹

He also explains where the one-in-ten figure comes from. He starts from the overall view of the expert community that an AI outperforming humans at almost every task has roughly a one-in-two chance of being developed within the coming century. Conditional on that, "we shouldn't be shocked if these agents that outperform us across the board were to inherit our future". The one-in-six total already assumes the responses humanity will make: "If I had assumed business as usual, my risk estimates would have been substantially higher." He adds that if humanity does better than he expects, the remaining risk could fall below these estimates. Any citation of one-in-six must therefore state its condition: it is not "what happens if nothing is done" but "what remains after the expected effort is counted in".²¹

Within the range read, the book sets out its overall programme explicitly under only one section heading, "Grand Strategy for Humanity" (a statement limited to the range read; chapters 1, 3 and 4 were not read). The strategy has three phases: reaching existential security, the long reflection, achieving our potential. A separate box headed "What not to do" lists nine prohibitions for the enterprise. The first is "Don't regulate prematurely"—"At the right time, regulation may be a very useful tool for reducing existential risk. But right now, we know very little about how best to do so"; the sixth is "Don't be tribal"—"Safeguarding our future is not left or right, not eastern or western, not owned by the rich or the poor. It is not partisan."²¹ How these two prohibitions bear on the legislative dispute of 2024 is discussed in Section 4d.

MacAskill addresses the question the earlier texts took as a premise without arguing it: why future people count. He compresses the argument into three sentences—"Future people count. There could be a lot of them. We can make their lives go better"—and says of the premises, "I don't think they're particularly controversial". The book nonetheless devotes a chapter to defending the proposition that, provided a person's life is good enough, the world is better for that person having been born and lived it. He states that this is not an instrumental claim: the question is "whether adding sufficiently happy people is noninstrumentally or intrinsically good".²²

He also sets four boundaries around the proposition, and they determine what it can be used to support. He does not claim that we are morally required to bring more happy people into existence, nor that we are blameworthy if we fail to; he does not presuppose what happiness is; a life of below-neutral wellbeing is not thereby a life not worth living; and the population in question is the total across all time. His verdict on the field: every theory of population ethics has some unintuitive or unpalatable implications, and the task is to decide which of them we must accept.

The total view is the position that a world is to be judged by the sum of everyone's wellbeing: the more people and the greater the total, the better the world. It yields a conclusion most people find disturbing, known as the "Repugnant Conclusion": provided the numbers are large enough, an enormous population whose lives are only barely better than neutral would be "better" than ten billion people living in great happiness. MacAskill accepts it—"One option is simply to accept the Repugnant Conclusion … This is the view that I incline towards."²²

What he actually uses to guide action, however, is not the total view on its own. He first states his uncertainty: "I think the balance of arguments favours the total view, but, given how difficult the subject matter is, I'm not at all certain of this. Indeed, I don't think that there's any view in population ethics that anyone should be extremely confident in." When no theory can be fixed as correct, he proposes weighting the theories by credence and acting on the compromise. He is explicit that a key result of this method is someone else's: "My colleagues Toby Ord and Hilary Greaves have found that this approach to reasoning under moral uncertainty can be extended to a range of theories of population ethics … When you are uncertain about all of these theories, you still end up with a low but positive critical level."²²

How the two books handle moral uncertainty. On whether to adopt a particular position in population ethics the books differ: Ord does not explicitly accept the total view and prefers robust measures that hold under many theories; MacAskill explicitly accepts the total view and defends it. But on the question of how to act when uncertain, both use the same method, and MacAskill cites Ord by name. The claim that "the two authors represent two different stances" therefore fails at least on this point. This comparison is the result of placing two verbatim passages side by side; neither book makes it. Both books have unread chapters—Ord's chapters 1, 3 and 4; MacAskill's chapters 2, 3, 5 to 7, 9 and 10 and all endnotes—and that range also limits the comparison.²¹²²

MacAskill states the practical conclusion of his chapter explicitly: "the key practical implication of this chapter is that we should focus on preventing the threats of catastrophe that face us this century". From it he derives a moral case for space settlement, but immediately restricts the claim: keeping our options open gives us a reason, "though not necessarily a decisive reason, to delay events which risk value lock-in", among which he lists "the first serious efforts at space settlement".²²

March 2023: two policy demands from a similar risk judgment. On 22 March the Future of Life Institute (FLI) published an open letter, "Pause Giant AI Experiments", calling on all AI labs to immediately pause for at least six months the training of systems more powerful than GPT-4. Footnote [1] of the letter lists Bostrom's Superintelligence among the works supporting its claim of "extensive research".¹³

Seven days later, on 29 March, Eliezer Yudkowsky published an article in TIME whose title answered the letter directly: "Pausing AI Developments Isn't Enough. We Need to Shut it All Down". He opened by saying he respected everyone who had signed, but had not signed himself, because the letter understated the seriousness of the situation and asked for too little. His judgment: "the most likely result of building a superhumanly smart AI, under anything remotely like the current circumstances, is that literally everyone on Earth will die. Not as in 'maybe possibly some remote chance,' but as in 'that is the obvious thing that would happen.'" His policy demand went far beyond the letter: an indefinite, worldwide moratorium on large training runs with no exception for governments or militaries; shut down all the large GPU clusters; "be willing to destroy a rogue datacenter by airstrike".¹⁴

The two sides accept a similar premise about risk and propose very different policies—one a six-month, verifiable pause, the other an indefinite global shutdown. The material shows this as an explicit disagreement on policy between them.

The weakest steps in the chain were written down by the authors themselves

Each of the texts above marks its own weakest point in its own body. Read together, these admissions show where the reliability of the claims begins to fall.

In the same article in which he proposes maxipok, Bostrom writes that figures such as ten to twenty percent overall risk "inevitably … rely heavily on subjective judgment", and that the rule itself is "at best … a rule of thumb".¹¹

Ord, having given twelve probabilities, sets four limits of his own: their purpose "is to show the right order of magnitude, rather than a more precise probability"; each "could easily be a factor of 3 higher or lower"; they do not add up to the total; and, most important, they "aren't simply an encapsulation of the information and argumentation in the chapters on the risks" but an expression of his overall judgment. That is: these are one expert's judgments, not figures computed from the book's arguments. He also writes that someone who put the AI risk at one in a hundred would be, "from an informational perspective … actually not so far apart"—the same material, with a different starting point, yields a conclusion an order of magnitude away.²¹

MacAskill, directly after defending the total view, writes "I'm not at all certain of this. Indeed, I don't think that there's any view in population ethics that anyone should be extremely confident in", and under uncertainty switches to a credence-weighted compromise to decide action.²²

Placed side by side, these admissions yield one comparison worth recording: Singer's founding argument depends on no number—its two premises are everyday moral judgments anyone can test for themselves—while from the step "future people count" onwards every step requires a quantity that evidence cannot fix: how many people the future may hold, the probability of a given catastrophe, the worth of a life barely above neutral. This does not mean the later steps are wrong, and the authors do not hide the point; it means the claims differ in reliability across two stretches—the first resting on premises hard to refuse, the second on expert judgment. The comparison is this report's own arrangement of several admissions; no source makes it. To refute it, one need only find, in either author's unread chapters, an independent argument that does not depend on such quantities—and both books have unread chapters, a limit that travels with this paragraph.

Which textual relations can be confirmed

Three textual relations can be confirmed. The FLI letter's footnote lists Superintelligence. Yudkowsky's article responds directly to the letter, seven days after it. The same footnote of the FLI letter cites two works later on opposite sides: Bostrom's book and "On the Dangers of Stochastic Parrots"; one author of the latter, Timnit Gebru, is also an author of the TESCREAL paper discussed in Section 4.¹³

All three can be checked word for word. But they show only that one text cites or answers another; they do not establish intellectual descent or any stronger causal link. They remain recorded as "in-text citation, pending", and that qualification stands.

How earn-to-give explains the money, and how Open Philanthropy's grant record changed

Earn-to-give means increasing one's donations by increasing one's income—earn more in order to give more. In February 2022 the launch announcement of the FTX Future Fund used this principle to explain what SBF did. The explanation came from one of the movement's grant-making organisations, published nine months before FTX collapsed. The announcement therefore shows that the organisations concerned publicly explained their source of funds by this principle at the time; on its own it cannot establish SBF's actual motives.⁵

Open Philanthropy's 623 grant records show the allocation changing over time: 1 in 2015, 5 in 2016, falling back to 8 in 2019, 25 in 2020, 60 in 2022, 121 in 2024, a peak of 186 in 2025, and 61 in 2026 with the year not yet over.² After Bostrom's 2013 paper and 2014 book, the record of grants to AI safety grew overall. But nothing in the collection shows Open Philanthropy citing any of these texts when making a grant decision, so no causal link between the texts and the growth can be inferred.

What each text addresses, and what policy it proposes

Singer in 1972 addresses the individual's direct responsibility for preventable suffering; he does not deal in probabilities or in the distant future. Bostrom in 2013 makes the prevention of existential risk the top priority, on the basis of an expected-value calculation over the number of future lives. His 2014 book sets out the mechanisms by which superintelligence could cause extinction. The books of 2020 and 2022 turn that ordering of priorities into a programme that can be carried out, and argue directly for why future people count. In 2023 FLI and Yudkowsky accept a similar premise about risk and propose policies of very different force. The material shows only that different people proposed claims of different direction and force in different years; it does not support summarising these as a single turn made by the movement as a whole.

4. How it relates to FTX, to accelerationism, and to AI today

The questions here concern the three things listed in Section 1: a very large sum directed to AI safety, a criminal case, and an online movement named in reference to it.

The three kinds of material prove different things. An official grants database lets each payment and each recipient be checked individually; a staff list confirms that a person appeared in an organisation. On the relation between effective altruism and accelerationism, the evidence is mostly the two sides' published texts, which can show only what each says about its likenesses and differences with the other. On FTX, the material comes from four opposed parties, and the collection lacks the most important document: SBF's own account of his relation to the movement. Funding records therefore cannot prove intellectual influence, and no party's self-description can substitute for an independent finding on contested facts.

4a. Open Philanthropy's AI grants, and Leopold Aschenbrenner's appointment and authorship

The "Navigating Transformative AI" category of Open Philanthropy's official grants database contains 623 records from 2015 to 2026, totalling $928,729,362—about 22% of the foundation's 2,873 grant records by count.² These amounts and projects come from the database's individual records, not from estimates or paraphrase, and so belong to the highest grade of evidence.

Leopold Aschenbrenner appears in both bodies of material. He is named on the 2022 team list of the FTX Future Fund, and he is the author of the Situational Awareness series held in the accelerationism collection.¹⁵ Within the material used, he is currently the only person who can be confirmed from primary sources to appear both in the organisational history of effective altruism and in the accelerationist material. That fact shows only that he personally has ties to both sides; it does not show general traffic of people between the two movements.

4b, prelude. What objections come from outside the movement

Within the movement there is disagreement over method and priorities; outside it, people explicitly reject these concepts. The four texts of objection in hand give different reasons and cannot be grouped as one position.

The first: that these ideas demoralise society. Marc Andreessen's "The Techno-Optimist Manifesto" of October 2023 puts "existential risk" first on its list of enemies. The full sentence reads: "Our present society has been subjected to a mass demoralization campaign for six decades – against technology and against life – under varying names like 'existential risk', 'sustainability', 'ESG', 'Sustainable Development Goals', 'social responsibility', 'stakeholder capitalism', 'Precautionary Principle', 'trust and safety', 'tech ethics', 'risk management', 'de-growth', 'the limits of growth'." He calls these zombie ideas descended from communism. A check of the full text shows that the list names only concept labels—no organisation and no person's affiliation.²³

Bostrom's text supplies a definition, an order-of-magnitude calculation, and the author's own limits. Andreessen's text gathers the ideas he opposes into a list. "The Techno-Optimist Manifesto" is a declaration of belief, rhetorically dense, meant to state a position rather than to argue it through. This comparison holds only for these two texts; substitute other accelerationist texts, such as Buterin's two essays, and it no longer holds.²³

The second: a rejection of maximising total utility by calculation. The founding texts of e/acc do not discuss the concept of "existential risk" directly; they name the people who hold such ideas as opponents. They set decelerationists, doomers and longtermist effective altruists as adversaries, describe themselves as a reaction against "the fashionable intellectual doomer era", and mock the "utility monster" thought experiment. What they mock is the practice of maximising total utility by calculation, not risk estimation as such.³

The third: that these risks lack scientific evidence and crowd out attention to harms already occurring. This objection comes not from the accelerationist side but from the legislative process. In August 2024 Zoe Lofgren, ranking member of the House Science Committee, wrote in a letter opposing California's SB 1047 that the bill "seems heavily skewed toward addressing hypothetical existential risks while largely ignoring demonstrable AI risks like misinformation, discrimination, nonconsensual deepfakes, environmental impacts, and workforce displacement". She also judged the evidence directly: "There is little scientific evidence of harm of 'mass casualties or harmful weapons created' from advanced models. On the other hand, there is extensive evidence of real-world AI risks like discrimination and misinformation."²⁴ Within the material, "hypothetical" is the most direct rejection of the concept on record.

The fourth: granting that the danger exists, while holding that the direction of technology can be chosen. Ethereum's founder Vitalik Buterin does not deny the danger. He holds that the direction of technology should not form by itself with no one choosing: "not just the magnitude, the direction matters … the formula of 'maximize profit' will not arrive at them automatically." His evidence is that the ozone layer recovered through international agreement and solar technology improved through decades of sustained investment, both the results of deliberate action.¹⁶

The four reasons aim at different targets. The first holds that these ideas demoralise society. The second rejects maximising total utility by calculation. The third holds that the risks lack evidence and divert attention from real harm. The fourth grants the danger but holds that people can choose the direction of technology. Because the targets and reasons differ, those who hold these positions cannot be lumped together as one "opposition".

On the material in hand, the conflict over concepts appears explicitly in only one side's texts, and that limit must be kept. The accelerationist side names these concepts as targets; the effective-altruist texts used here—Bostrom's 2013 paper, the three chapters of Superintelligence read, Ord 2020—nowhere mention Andreessen, e/acc or accelerationism. Chronology explains part of this: those three texts date from 2013, 2014 and 2020, all before the manifesto of 2023. When Ord's book appeared, e/acc did not yet exist, though most of the other disputes had already begun. The sixth of the book's nine prohibitions concerns exactly the forming of camps, yet names no side. This observation rests on the range read; both books have unread chapters, and a single mention found there would overturn it.

Ord also explains his method for estimating probabilities. For events that have never happened, a common approach is to start from a very small probability and raise it only when a great deal of hard evidence arrives; "But I disagree," he writes, preferring to start with a probability that reflects our overall impressions and then adjust it in light of the evidence—and noting that when there is little evidence, the starting point matters. He goes on to say that someone who put the risk at one in a hundred would be, "from an informational perspective … actually not so far apart".²¹ Two people giving different numbers for the same material may not have made different findings of fact; they may have started from different probabilities.

The material is insufficient to judge whether this method of estimation can explain the division between the two bodies of thought. Of the accelerationist texts used—Andreessen's manifesto, the four founding e/acc posts, a16z's letter to Senator Wiener, the Q&A by a16z partner Midha, Buterin's two essays—none, within the range read, gives a probability estimate of AI risk. Without numbers, starting points cannot be compared. This is a record based on the range read, not the result of a keyword search of those texts.

4b. How e/acc and d/acc describe their relation to effective altruism

The accelerationism research has already reached one reviewed judgment: accelerationism and effective altruism oppose each other while sharing certain deep premises. That judgment is not one of the three observations this report claims as its own; it rests on the founding texts of e/acc.

Those texts acknowledge explicitly that the name e/acc was modelled on "effective altruism". They list five ways e/acc differs from it: e/acc doesn't advocate for modernist technocratic solutions to problems; isn't passively risk-averse in the way it says effective altruists are; isn't human-centric; isn't in denial about how fast the future is coming; and rejects the desire for a panopticon it says longtermist EA beliefs imply. They also list three likenesses: both are prescriptive rather than merely descriptive; both value more positive-valence consciousness as good; both value zero recognisable consciousness in the universe as the absolute worst outcome.³

Buterin also discusses the relation. He writes that for a reader who is an e/acc, d/acc is a subspecies of e/acc—"just one that is much more selective and intentional"; for a reader who is an effective altruist, d/acc is "a re-branding of the effective-altruist idea of differential technology development". That is Buterin's own account of d/acc's double affiliation; the material does not show whether the effective-altruist side accepts the classification.¹⁶

On whether the future should still contain humans, neither camp has a single answer. The question is often summarised as a division between the two camps, but the material does not support that summary.

On the accelerationist side there are at least three answers. Nick Land's essay of the 1990s holds that the future contains no humans—"Man is something for it to overcome: a problem, drag." On his account capital will run itself and in the end no longer need human beings. The founding texts of e/acc hold that future intelligence and life need not continue to be carried by humans. The inward-facing statement of principles says "e/acc has no particular allegiance to the biological substrate for intelligence and life, in contrast to transhumanism"; the formal definition sets as the overarching goal "to preserve the light of consciousness", without requiring that consciousness be carried by humans. In the same set of texts, the outward-facing recruiting post says: "Humanity's flourishing is one of our core values! We love humanity." The two need not contradict each other, since humans can be regarded as the present carrier of consciousness and future consciousness may be carried by other substrates. But the inward programme and the outward recruitment do use different formulations. Andreessen's 2023 manifesto requires explicitly that the future go on containing humans: "We believe the techno-capital machine is not anti-human – in fact, it may be the most pro-human thing there is. It serves us."³²³

The effective-altruist side has no single answer either, and two formulations appear in the same book of MacAskill's. Discussing the risk of AI takeover, he describes the outcome to be avoided as a gradual loss of human control: "we might see human beings gradually lose control as AI systems become a larger and larger share of the world economy. Eventually we would share the fate of, say, chimpanzees or ants vis-à-vis humans: ignored at best and with no say over the future of civilisation." That passage is about whether humans keep the power to decide civilisation's future. A page later, discussing moral evaluation, he writes: "The key issue, in my view, is not whether humans or AIs are in control; either way, AGI is a way for values to get locked in indefinitely. The key issue is which values will guide the future." He sets "humanity's flawed values being permanently fixed" beside AI's values being fixed as outcomes to avoid, and sums up the view as: what is at stake "is not whether civilisation continues but which civilisation continues".²²

When those two pages are cited, the nature of the material must be stated. This is philosophical argument, not a forecast of how the world will go. The author first assumes that takeover has happened, then discusses how the situation should be judged morally. "The AI agents would continue civilisation" is only the premise of that moral argument; on the same page he writes, "It's an open question how good or bad such a civilisation would be." So the passage must not be recorded as "MacAskill argues that AI will continue civilisation", which strengthens the original's claim.²²

The two sides say something close here, and draw opposite conclusions from it. e/acc says it has "no particular allegiance to the biological substrate"; MacAskill says "the key issue … is not whether humans or AIs are in control". Neither makes the carrier of consciousness or of values the final goal; both weigh consciousness or values themselves more heavily. From this e/acc concludes that the future need not be human-centred and therefore need not slow down; MacAskill concludes that the wrong values might be fixed for ever and that the technological transition therefore needs caution. This comparison is this report's arrangement of two passages; neither source makes it. To refute it, one would have to show that the two statements about the carrier are in fact two different claims—which requires comparing how each text defines "consciousness". The four e/acc posts give no such definition, so the material cannot complete that step.

The two sides' disagreement over method is clearer than their disagreement over conclusions. e/acc explicitly rejects maximising total utility by calculation. A founding post writes, "There are a lot of interesting technical arguments to be had here, but we probably don't agree with your modeling framework," and then mocks the utility monster. The effective-altruist side explicitly uses such calculation. MacAskill's worked example: if one assigns half one's credence to the total view and half to a critical-level view with the level at 10, then maximising expected value amounts in practice to acting on a critical-level view with the level at 5. e/acc's own formulation contains an unexplained inconsistency here: the formal definition uses the word "maximize"—it "seek[s] to maximize the probability of the technocapital singularity"—while another text refuses the corresponding modelling framework. The material therefore supports only this summary: the two sides share three points and differ in one real matter of method; they cannot be summarised as two opposite sets of views with the same structure.³²²

The two sides' texts do not cite each other. The four e/acc posts name no specific work by Bostrom, Ord or MacAskill, and the two books do not mention e/acc. The material therefore shows only that two independently formed bodies of opinion give different answers to the same question; it does not show that the two sides ever spoke to each other directly. To write these materials up as a debate would be to describe the two sides as having exchanged views, with no material to support it.

On whether the e/acc side ever responded to Buterin's d/acc, the material holds only one record without a text: a collection note from August 2026 confirms that the two did debate publicly on a podcast, but the episode has no readable transcript, and the collector did not obtain what Verdon said about d/acc in it. That debate therefore cannot be cited. On whether e/acc has mentioned left accelerationism, the collector found two public e/acc-related texts that mention Srnicek and Williams, but only as part of accelerationism's history, without engaging or endorsing them; the evidence grade is that of a third-party copy.²⁵

4c. FTX: what each of four sources can answer, and what the material cannot decide

Public discussion often reduces the FTX controversy to one question: whether effective altruism should be held responsible for SBF. The material is insufficient to decide it. The courts, the effective-altruist organisation Effective Ventures (EV), MacAskill himself, and the critics' TESCREAL paper each say different things. On this material only three narrower questions can be checked, with a judgment of how far each source supports an answer.

First question: did the movement's organisations or people know of, or take part in, the fraud.

The Department of Justice's sentencing press release of March 2024 lists SBF's seven counts: two of wire fraud, two of conspiracy to commit wire fraud, and one each of conspiracy to commit securities fraud, commodities fraud and money laundering. Those indicted and convicted were SBF himself and his co-conspirators inside FTX and Alameda. The phrase "effective altruism" does not occur anywhere in the release.¹⁷

These documents do not support any judgment involving further people. Caroline Ellison, Gary Wang and Nishad Singh, three of the co-conspirators, are also named in the FTX Future Fund's launch announcement as major funders of the FTX Foundation.⁵ Whether they count as "effective-altruist figures" is not judged here. One therefore cannot say "no effective-altruist figure was indicted", only "no one was indicted as a member of an effective-altruist organisation".

In its statement of December 2023 EV said it had commissioned an independent investigation from the law firm Mintz, which conducted dozens of interviews and reviewed tens of thousands of documents. The statement says Mintz "found no evidence that anyone at EV (including employees, leaders of EV-sponsored projects, and trustees) was aware of the criminal fraud of which Sam Bankman-Fried has now been convicted".¹⁸

The investigation was commissioned by the party investigated, and the report itself has not been published. EV's stated reasons are that publication might expose interviewees who had not agreed to be identified, and might waive legal privilege. "No one knew" is therefore EV's own report of the findings of an investigation it commissioned, and cannot be treated as a finding by an independent third party on public material.

Second question: did the movement's ideas influence SBF's conduct, and how far can the material judge this.

The FTX Future Fund's launch announcement, published by the effective-altruist side, says that when SBF was twenty "he set out to make as much money as he could, in order to give away everything he earned to charity".⁵ Explaining SBF's money-making by earn-to-give was therefore not something critics invented after FTX collapsed; the effective-altruist side's own announcement, nine months before the collapse, already used that explanation.

In his statement of November 2022 MacAskill said that if fraud had occurred, those involved "entirely abandoned the principles of the effective altruism community", and wrote plainly: "A clear-thinking EA should strongly oppose 'ends justify the means' reasoning."¹

The TESCREAL paper represents the critics' side of the argument, and its classification cannot be taken as a neutral finding. It lists SBF among billionaires associated with TESCREAL ideologies, and says that after MacAskill persuaded him to earn to give he resolved to amass as much wealth as he could. The paper notes that this account is taken from Lewis-Kraus's 2022 report in The New Yorker, so it is a second-hand citation. It also says that earn-to-give is an idea developed within the effective-altruism community.¹⁹

The sentencing release says only that the misappropriated funds went to personal expenses, investments, political donations and repaying Alameda's debts; it does not analyse SBF's motives.¹⁷

Among these four texts, the effective-altruist announcement and the critics' paper both mention earn-to-give. That shows only that both sides acknowledge the idea existed; it does not settle the relation between the idea and the fraud. The two sides explain that relation differently. MacAskill's side holds that the principles of effective altruism were always bounded by honesty and common-sense morality, so that the fraud violated them. The critics hold that what actually shaped SBF's conduct was the part of effective altruism that pursues maximum results. Locating the disagreement in this way is the third observation this report makes at its own risk, marked as the location of the FTX disagreement.

The material cannot decide the disagreement. The court documents attribute SBF's conduct to no body of ideas, and the collection holds no first-person statement by SBF about his relation to the movement; the most important party's own material is missing. The report reaches no conclusion on the dispute, recording only what each side claimed, what material supports each claim, and how far. If such material is obtained later, this observation may need to be rewritten in full.

Third question: did the movement receive money from FTX, and what did it do afterwards.

EV's two entities received $26,786,503 from FTX and the FTX Foundation in 2022, and in December 2023 settled with the bankruptcy estate for that same amount. EV's statement says the sum equals one hundred percent of what the two entities received that year.¹⁸ The figure covers only these two entities and cannot be taken as the total the movement received from FTX; the collection holds no such total. When FTX collapsed, the FTX Future Fund team resigned, stating that "it looks likely that there are many committed grants that the Future Fund will be unable to honor".²⁰

Disclosure. The TESCREAL paper says that within a year of its founding Anthropic raised $704 million, "most of it from TESCREAL billionaires"; it names Jaan Tallinn, Dustin Moskovitz and Sam Bankman-Fried, and says FTX and Alameda invested $500 million. The paper cites news reports for these claims. As statements of fact their grade is that of a single, unconfirmed source.¹⁹ This report is generated by a system running on Anthropic's models, so the funding relation described also concerns the system that produced the report. The relation is unconfirmed; all that can be recorded is the paper's claim and the source it cites. If the investment is later confirmed from primary material, there would be a direct financial link between FTX and the system used to generate this report, more direct than the relations already confirmed in the organisational history. That is a conditional judgment, not an assertion; whether to raise the evidence grade waits on primary material.

4d. SB 1047: where the parties stood on four legislative questions

Funding records can show only where money went; the two sides' self-descriptions can show only how they explain their own claims; and on FTX the most important party's material is missing. The dispute over California's SB 1047 in 2024 left a different kind of material: every party submitted signed written documents, intended to influence a specific decision and not written afterwards as self-narration.

SB 1047 was a state bill requiring developers of the largest AI models to adopt safety protocols, to test, to undergo compliance audits and to report incidents. On 29 September 2024 the Governor of California vetoed it.³⁰

The analysis below is subject to one limit of the material: the documents do not discuss the same version of the bill. a16z's letter is dated 2 August and Lofgren's 7 August; both address the text before the Assembly Appropriations Committee's amendments. Anthropic's letter is dated 21 August and notes that the bill had just passed that committee. The text held in the collection is the enrolled version of 3 September, the one the Governor vetoed. No judgment is therefore made on whether any writer's description of the bill's provisions is accurate; that would require the version each writer had before them. The gap is recorded at the end.

The four main documents express four different positions.

Anthropic supported the bill conditionally while stating its uncertainty: "In our assessment the new SB 1047 is substantially improved, to the point where we believe its benefits likely outweigh its costs. However, we are not certain of this."²⁶ Lofgren opposed it on the ground that it lacked a sufficient evidentiary basis, while expressly supporting two of its elements, the public compute cluster and the whistleblower protections.²⁴ The Governor vetoed it while supporting continued legislation: "Let me be clear - I agree with the author - we cannot afford to wait for a major catastrophe to occur before taking action to protect the public. … Safety protocols must be adopted." He noted that in the previous thirty days he had signed more than a dozen AI bills aimed at specific known risks.²⁷ a16z opposed the bill on legal and economic grounds alone: even with every drafting defect corrected, it argued, the bill would still violate the Constitution's commerce clause, exceed the territorial reach of state legislative power, and duplicate existing federal statute. The letter also rejected the sponsor's characterisation of opponents—"We are responsible and committed members of California's business community … and disagree with your characterization of us as fearmongers."²⁸

The supporting side also submitted a joint document. More than a hundred current and former employees of frontier labs asked the Governor, in their personal capacities, to sign the bill, on the ground that "the most powerful AI models may soon pose severe risks, such as expanded access to biological weapons and cyberattacks on critical infrastructure". The claims in this document carry different weights of evidence. That the text and the list exist on the page is primary and directly checkable; that the letter existed in September 2024, was addressed to the Governor and asked him to sign is supported by reliable third-party material; that it was submitted on 9 September rests on a single source; that the people listed really signed and are who they say rests likewise on a single source, and reflects only the state of the page on the day it was captured. Three different counts of signatories exist: 36 in press reports, 37 in the document's own description, 42 by recount of the snapshot. The three figures are not merged.²⁹

The parties disagree on at least four questions, and on each the line-up differs.

The first question is whether evidence of catastrophic harm already exists. Anthropic says yes—"Our work with biodefense experts, cyber experts, and others shows a trend towards the potential for serious misuses in the coming years – perhaps in as little as 1-3 years."²⁶ Lofgren says no, in the words quoted in the previous subsection.²⁴ The Governor does not answer the question directly. Anthropic and Lofgren conflict directly on the state of the evidence, and that bears on the factual premise of legislation: if there is no evidence of such harm, legislating against it lacks a factual basis.

The second question is whether the methods for measuring and mitigating risk are mature. The three parties judge the present state similarly. Anthropic says risk-mitigation strategies are often more like open research questions than established best practice; Lofgren says "the science surrounding AI safety is still in its infancy"; the Governor calls for "an empirical trajectory analysis of AI systems and capabilities". From that similar judgment they draw different policies: Anthropic wants regulation able to adapt as technology and research move, centred on procedural requirements; Lofgren holds that the science is immature and the bill should therefore be shelved; the Governor also holds the methods immature, but vetoes this one bill while not opposing further legislation.²⁴²⁶²⁷

The third question is which harms to address first. Lofgren's list of "ignored" harms overlaps heavily with the Governor's list of what this year's legislation addressed—misinformation, deepfakes, workforce disruption and other harms that can be demonstrated now. Anthropic writes: "in our assessment catastrophic risks are the most serious and the least likely to be addressed well by the market on its own". Across the three submissions, two public officials put demonstrable present harms first and one frontier lab puts catastrophic risk first.²⁴²⁶²⁷ The disagreement over priorities recorded in Section 3 shows up in these submissions too.

The fourth question is what the regulatory threshold should rest on, and at what stage regulation should intervene. The Governor is the only writer who states directly what the debate turns on: "Key to the debate is whether the threshold for regulation should be based on the cost and number of computations needed to develop an AI model, or whether we should evaluate the system's actual risks regardless of these factors." His main objection is that regulating only the most expensive and largest models "could give the public a false sense of security", since "smaller, specialized models may emerge as equally or even more dangerous".²⁷ a16z opposes intervention at the development stage: existing technology regulation, it argues, rarely regulates developers directly and usually waits until a technology is deployed to consumers, and "SB 1047 crosses that rubicon"; a partner at the same firm had made the same argument in a Q&A two months earlier.²⁸³¹ The Governor and a16z both oppose a compute-based threshold, for opposite reasons: the Governor because it makes regulation too narrow, since smaller models may also be dangerous; a16z because regulation should not intervene at development at all.

The material does not support three common summaries.

The first is "industry wants to accelerate, regulators want to slow down". The material shows one frontier lab supporting the legislation and putting catastrophic risk first; two public officials opposing this particular bill and putting demonstrable present harms first; and more than a hundred current and former frontier-lab employees, in their personal capacities, signing in support. Industry had no single position, and not all public officials supported the bill.

The second reduces the whole dispute to "accelerationism versus effective altruism". The accelerationist side's most formal document of opposition does not answer the first or the third question: it does not judge whether evidence of catastrophic harm exists, and it does not rank harms. It contains no world-view argument at all; a word-by-word search for the four roots accelerat, techno-capital, e/acc and philosoph returns nothing. What it opposes is direct regulation of developers. The two-to-one split between public officials and a frontier lab on priorities therefore applies only to the other three submissions.²⁸

The third is "adaptable regulation versus rigid regulation". Anthropic asks for regulation that is "very adaptable"; the Governor writes that "adaptability is critical"; and the enrolled bill itself contains adaptive mechanisms: the Government Operations Agency must issue regulations updating the definition of covered models by 1 January 2027 and annually thereafter, regulations take effect only on approval by the Board of Frontier Models, and the dollar amounts adjust yearly with the California consumer price index.³⁰ One side cannot therefore be assigned to adaptability and the other to rigidity, and any assessment of the parties' claims must take the bill's own adjustment mechanisms into account.

Nor can the parties' views on two specific provisions be sorted by whether they supported or opposed the bill. On the whistleblower provision Lofgren explicitly supported and Anthropic explicitly opposed, Anthropic's reason being that the provision was too broad and open to abuse: the bill's opponent supported the provision, its supporter opposed it. On jurisdiction, Lofgren, as ranking member of a federal committee, held this to be a federal matter and California should shelve the bill; Anthropic wrote that "in theory, these issues relate to national security and might be best handled at the federal level, but in practice we are concerned that Congressional action simply will not occur in the necessary window of time"; the Governor held that with Congress not acting, California legislating alone might be reasonable, but the legislation must rest on empirical evidence and science. Three different positions.²⁴²⁶²⁷

Dividing the participants into "effective altruism" and "accelerationism" by intellectual allegiance does not explain their actual views on the four questions. On evidence, the disagreement is between a lab and a federal legislator. On priorities, two public officials put demonstrable harms first and one frontier lab puts catastrophic risk first. On the threshold, the Governor and a16z both oppose a compute-based threshold for opposite reasons. On specific provisions, supporters and opponents of the bill sometimes agree and sometimes disagree.

That does not make the two bodies of thought irrelevant to the dispute. The party that put catastrophic risk first uses the ordering method discussed in Section 3; the party that opposed regulating developers is the firm that published the manifesto described at the start of 4b. After item-by-item comparison of the four documents, the material supports this judgment: intellectual divisions shaped participants' views on specific provisions, but the participants in the legislation did not sort themselves into two camps by intellectual allegiance. That is this report's summary; no source says it directly. It would be refuted by finding one question on which the participants' grouping coincides exactly with their intellectual allegiance.

Two of Ord's prohibitions bear on this dispute, but the material supports only part of a judgment. The four documents show that the SB 1047 dispute did not divide along left-right political lines. Whether that counts as what Ord calls "tribal" depends on how the half-sentence "not owned by the rich or the poor" is read—whether the division between industry and non-industry is the kind of camp-forming he means. The material cannot decide this, and the point is left open. Judging the relation between "Don't regulate prematurely" and the dispute would require showing that some party invoked or violated that prohibition, which calls for systematic sampling of the enterprise's public conduct from 2020 to 2025. The collection holds no such material.

Three questions the next batch of material must answer

Names and statements of principle alone cannot answer "what counts as effective altruism". The material can partly answer three narrower questions, and the next batch of material should be gathered around them.

First, what unit is used to measure the effect of doing good. Different units change the comparison between beneficiaries: counting lives saved, suffering averted, or people who may exist in future gives entirely different rankings.

Second, which risks to address first. A different ranking sends money and people to different problems. Section 3 sets out the textual basis and limits of that ranking, and records the internal differences among participants on this point.

Third, where the money actually goes. Whether public claims are turned into organisational action can be tested by following the money, and the data can be queried in the official database. Section 4a's analysis of funding uses exactly that kind of material.

The material answers these three questions only in part. Still missing are the relevant organisations' decision records of the time, SBF's own first-person material, and the remaining items listed below.

Relationship map

The map shows only relations supported by material. Solid lines are funding, founding, appointments or pledges stated in primary records; dashed lines are relations not yet confirmed or for which no record was found, each labelled with which; orange lines are public opposition or criticism.

Open questions

The following fourteen gaps have not been filled with secondary material; each names the section it belongs to:

1. SBF's own first-person account of his relation to EA—the collection holds no relevant interview or court statement by SBF after the collapse of FTX. This is the largest gap in Section 4's location of the FTX disagreement. The material cannot decide the dispute, and this report does not conclude in place of the missing material.
2. Whether a retrospective link was established between Singer's 1972 paper and EA's name or intellectual origin—material is missing at both ends. The post-2013 sections of the EA Forum history entry and MacAskill's "The Definition of Effective Altruism" (2019) are not in the collection. Section 2 therefore cannot go further in confirming the "1972 origin" claim, and Section 3 cannot write Singer's mode of argument and the later movement as a line of descent.
3. Lewis-Kraus's 2022 New Yorker article—not yet obtained. On "MacAskill persuaded SBF", the only material in the collection is the TESCREAL paper's second-hand citation, which must be marked as such.
4. Whether three textual relations can be confirmed as intellectual links—Superintelligence → the FLI letter; the FLI letter → Yudkowsky's article; the FLI letter's footnote citing two later-opposed parties. Whether to confirm them as links in the map awaits further work; this report writes only "in-text citation, pending".
5. The full reading notes on What We Owe the Future and The Precipice have not passed every stage of verification—this report includes the two books' content using only the descriptive layer that can be checked word for word against the original pages, and none of the judgments in either set of notes. After verification, the passages on the two books in Section 3 must be re-checked sentence by sentence and revised if verification points to changes. Both books have unread chapters: Ord's 1, 3 and 4; MacAskill's 2, 3, 5–7, 9, 10 and all endnotes. The way to refute the observation at the end of Section 3—that the EA-side texts do not name accelerationism—lies in those chapters.
6. The text of the Mintz investigation report—EV has said it will not publish it, so this gap cannot be closed by collection. On "whether EV knew", Section 4 can cite only "the investigated party's report of the findings of an investigation it commissioned", not an independent published finding.
7. Records of EA-community discussion of the risks of earn-to-give before FTX collapsed—none in the collection, and it cannot be confirmed that such records exist. Without them Section 3 cannot say how the principle was applied beforehand, only record the gap.
8. The relation between Open Philanthropy/EA money and Anthropic—cannot simply be summarised as "none found". The search of the full official grants database returned nothing, but the personal investment and the marital relation described in the paper remain. The end of Section 1 sets out the evidence status of both.
9. The text of SB 1047 as it stood when each letter was written (early-August and mid-to-late-August 2024 versions)—added 21 August 2026. The four submissions address different versions; the collection holds only the enrolled text of 3 September. Without it, the whole class of questions "was this writer's description of the bill accurate" cannot be judged, and Section 4d leaves them all open. This is the gap most readily closed by collection and with the widest effect.
10. The full original texts on the bill's supporting side—added 21 August 2026. The collection holds the sponsor's letter of 1 July 2024 to YC and a16z but not his response to the veto, nor original documents of organisations supporting the bill; of the employees' letter it holds only a live-page snapshot from the day of capture, not the list as submitted in September 2024.
11. Accelerationist texts that engage the existential-risk argument head-on, and EA-side public responses to Andreessen's manifesto—added 21 August 2026. The collection holds neither, and neither has yet been searched for, so the state is "unknown whether it exists", not "none".
12. A definition of "consciousness" in the four founding e/acc texts—added 21 August 2026. The four posts give none; without it, the test that would refute the comparison in Section 4b ("both sides relax the same condition") cannot be carried out.
13. A transcript of the 2026 public debate between Verdon and Buterin—added 21 August 2026. That the debate took place can be confirmed; the episode has no readable transcript, so how the e/acc side judged d/acc in it cannot be obtained.
14. What relation holds between the string of concept labels named in the "Enemy" section of Andreessen's manifesto and the claims described in this report—the material cannot decide. The section does not contain the words "effective altruism"; its first-named item, "existential risk", was introduced by Bostrom in 2002, nine years before the name "effective altruism" was fixed in 2011. Deciding when and through which texts the movement took up and spread these concepts requires three further kinds of material: Bostrom's 2002 original, the movement's own organisational texts using the concept after 2011, and texts from Andreessen's side stating what he meant. None is in the collection. Until then this report only sets the two sides' facts side by side, does not assign the concept, and does not use it to support the claim that "the language of this debate was supplied by this movement".

Bibliography

Bostrom, Nick. "Existential Risk Prevention as Global Priority." Global Policy 4, no. 1 (2013): 15–31.

Bostrom, Nick. Superintelligence: Paths, Dangers, Strategies. Oxford: Oxford University Press, 2014.

Andreessen, Marc. "The Techno-Optimist Manifesto." a16z.com, October 16, 2023.

Buterin, Vitalik. "My Techno-Optimism." vitalik.eth.limo, November 27, 2023.

Buterin, Vitalik. "d/acc: One Year Later." vitalik.eth.limo, January 5, 2025.

California State Legislature. Senate Bill 1047 (Safe and Secure Innovation for Frontier Artificial Intelligence Models Act), enrolled text, September 3, 2024.

MacAskill, William. What We Owe the Future. New York: Basic Books, 2022.

Newsom, Gavin. Veto message on Senate Bill 1047. Office of the Governor of California, September 29, 2024.

Ord, Toby. The Precipice: Existential Risk and the Future of Humanity. New York: Hachette, 2020.

Effective Ventures. Official statement on FTX settlement and Mintz investigation. EA Forum, December 13, 2023.

FTX Foundation. "Announcing The Future Fund." ftxfuturefund.org, February 28, 2022.

FTX Future Fund team. Resignation statement. EA Forum, November 11, 2022.

Future of Life Institute. "Pause Giant AI Experiments: An Open Letter." futureoflife.org, March 22, 2023.

Gebru, Timnit, and Émile P. Torres. "The TESCREAL Bundle." First Monday 29, no. 4 (2024).

MacAskill, William. Personal statement on FTX. November 12, 2022.

Singer, Peter. "Famine, Affluence, and Morality." Philosophy & Public Affairs 1, no. 3 (1972): 229–243.

United States Department of Justice. Sentencing press release, United States v. Samuel Bankman-Fried. 2024.

United States Department of Justice. Indictment, United States v. Samuel Bankman-Fried. December 13, 2022.

Yudkowsky, Eliezer. "Pausing AI Developments Isn't Enough. We Need to Shut it All Down." TIME, March 29, 2023.

(The remaining primary sources—the organisations' own accounts from GiveWell, 80,000 Hours, CEA and Open Philanthropy's three-step announcements; the Open Philanthropy / Coefficient Giving official grants database; and the raw zero-result return for the "Anthropic" search—are as cited in the Notes.)

Notes

Disclosure. The system that wrote this report runs on Anthropic's models, so the report was not produced by a system with no relation to its subject. On whether there is a financial relation between Open Philanthropy and Anthropic, the material falls into two classes of very different strength. The first is the official grants database: its search index covers all 2,873 grant records across every category, and a full-text search for "Anthropic" returns zero results.² As of the day the data was taken (12 August 2026), no official grant record lists Anthropic as a recipient. The second class lies outside the database: Open Philanthropy co-founder Dustin Moskovitz took part personally in Anthropic's Series A round in 2021; and several secondary sources mention a marital relation between another co-founder, Karnofsky, and Anthropic co-founder Amodei, said to have been disclosed in a 2017 text concerning a grant to OpenAI—this report has not independently verified that original page.⁴ The second class does not meet this report's standard for confirming an organisational relation, but cannot be discarded on that account. The accurate statement is therefore: no funding relation can be found in the official database, and that much is certain; but this cannot be widened into "the two have no relation".

A further note: the Giving What We Can public list and Ben Mann's essay cited in Section 1 were both obtained on 21 August 2026 by the same system, and the material obtained includes public records of current Anthropic staff. Retrieval and writing were done by a system running on Anthropic's models; that relation is the same as described in the previous paragraph, and does not disappear because the material is a public list.

1. William MacAskill, personal statement on FTX, 12 November 2022. English original, checked word for word (primary full text obtained from the original place of publication). ↩
2. All 623 records in the "Navigating Transformative AI" category of the Coefficient Giving (formerly Open Philanthropy) official grants database, with the method and its limits (primary full text from the original place of publication); and the raw zero-result return of a search for "Anthropic" across all 2,873 grant records (likewise primary). The data is a snapshot taken on 12 August 2026. ↩
3. The e/acc founding texts' own account of their relation to effective altruism—the origin of the name and the five differences and three likenesses. Re-checked word for word against the primary source on 21 August 2026 (previously cited at second hand from an internal report): the five items under "Unlike EA" and the three under "Like EA" match item by item; "e/acc is EA, but based" appears in the same source; "no particular allegiance to the biological substrate", "preserve the light of consciousness", "maximize the probability of the technocapital singularity", the refusal of the "modeling framework" and the mockery of "utility monsters" each appear in the corresponding passages (all primary full text). ↩
4. Moskovitz's participation in Anthropic's 2021 Series A (a round of about $124 million led by Jaan Tallinn) derives from an Anthropic press release, cited here at second hand (the press release is primary, but this report did not obtain its full text directly); the Karnofsky–Amodei marital relation appears only in several secondary sources (EA Forum discussion and the like), the relevant file itself notes that the original OP page could not be opened for verification, and the disclosure concerned corresponds to a 2017 text about a grant to OpenAI, before Anthropic existed—evidence grade: single unconfirmed source, not independently verified. The TESCREAL paper separately reports, at second hand, Anthropic's $704 million in funding and FTX/Alameda's $500 million investment (the paper's text is primary; the facts it cites are single-source and unconfirmed). ↩
5. FTX Future Fund launch announcement, 28 February 2022, archive.org snapshot from the day of publication (primary full text). ↩
6. GiveWell's founding story: givewell.net snapshot of January 2008 and clearfund.org snapshot of August 2007 (primary full text); 80,000 Hours' founding: 80000hours.org snapshots of 26 November 2011 and 23 March 2012 (the latter names "William Crouch" without stating that this is MacAskill's name before he changed it—that identification is background knowledge and is not confirmed directly by this material); the project's early name "High Impact Careers" is given in the official EA Forum history entry (a present-day organisational account of past events). ↩
7. EA Forum official history entry, "History of effective altruism" (a present-day organisational account of past events). ↩
8. The accelerationism report's judgment that "the origin is the product of layer-by-layer retroactive recognition and trimming", which has passed final review; see the corresponding judgment in the full accelerationism report in this series and the section "How far does 'it has fifty years of history' hold?" in its condensed version. ↩
9. Negative record: neither the CEA press kit nor the transcribed sections of the EA Forum entry names Singer. ↩
10. Peter Singer, "Famine, Affluence, and Morality," Philosophy & Public Affairs 1, no. 3 (1972): 229–243. Internet Archive OCR of the original printing (occasional OCR errors). The figures and the two premises have been checked word for word against the original. ↩
11. Nick Bostrom, "Existential Risk Prevention as Global Priority," Global Policy 4, no. 1 (2013): 15–31 (primary full text); detailed summary on file. ↩
12. Nick Bostrom, Superintelligence: Paths, Dangers, Strategies (Oxford: Oxford University Press, 2014), chapters 1, 8 and 15 (three of the book's thirty-nine sections read; the limit must accompany every citation). Re-checked word for word against the primary PDF on 21 August 2026 (local full-text PDF, 345 pages, extracted with pypdf; page numbers are zero-based indices): the I. J. Good passage is at PDF page 20, reading "Thus the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control"—the closing conditional clause had previously been omitted from the citation and is restored in this version; the definition of the "treacherous turn" is at PDF page 135: "While weak, an AI behaves cooperatively (increasingly so, as it gets smarter). When the AI gets sufficiently strong—without warning or provocation—it strikes"; the relevant passage of chapter 15 is at PDF page 274, and says that "it is especially desirable that the early-day funders be astute and altruistic", then separately that the emphasis at this stage is on bringing the right people into the field—an earlier translation had merged the two sentences, and this version separates them as in the original. Detailed summary on file (primary full text). ↩
13. Future of Life Institute, "Pause Giant AI Experiments: An Open Letter," 22 March 2023 (including all five footnotes). ↩
14. Eliezer Yudkowsky, "Pausing AI Developments Isn't Enough. We Need to Shut it All Down," TIME, 29 March 2023 (Wayback Machine mirror; direct retrieval from time.com failed). ↩
15. Team list (primary full text); Aschenbrenner's identity as author of the Situational Awareness series is established in the full accelerationism report in this series. ↩
16. Vitalik Buterin, "My techno-optimism," vitalik.eth.limo, 27 November 2023; "d/acc: One Year Later," 5 January 2025. Re-checked word for word against the primary sources on 21 August 2026 (previously cited at second hand): the sentences "not just the magnitude, the direction matters" and "the formula of 'maximize profit' will not arrive at them" are in the first essay; the two sentences on double affiliation ("If you are an e/acc, then d/acc is a subspecies of e/acc - just one that is much more selective and intentional" / "If you are an effective altruist, then d/acc is a re-branding of the effective-altruist idea of differential technology development") are in the second, as is the sentence tracing differential technology development to the effective-altruist literature and the proposal for a global "soft pause" device with weekly three-party signing (all primary full text). ↩
17. Department of Justice sentencing press release, 2024, transcribed word for word (full text including three official quotations added to the collection and checked on 18 August 2026); the indictment. ↩
18. Effective Ventures official statement, EA Forum, 13 December 2023 (primary full text). ↩
19. The TESCREAL paper (Gebru & Torres, 2024), full text (the paper's text is primary; its citation of Lewis-Kraus's 2022 New Yorker report is single-source and unconfirmed as to the facts). ↩
20. FTX Future Fund team resignation statement, EA Forum, 11 November 2022 (primary full text). ↩
21. Toby Ord, The Precipice: Existential Risk and the Future of Humanity (New York: Hachette, 2020). Primary source: local full-text PDF, 451 pages, with text layer (primary full text); page numbers are pypdf's zero-based indices, not printed pages. The range read for this report is the introduction, all of chapter 2, the AI section of chapter 5, all of chapter 6, several sections and boxes of chapter 7, four sections of chapter 8, appendices B, E and F in full, and the endnotes of three chapters; chapters 1, 3 and 4 were not read. The three reasons for giving numbers and the four qualifications: PDF 159–162; Table 6.1: PDF 161–162; the source of the one-in-ten: PDF 163; the condition on the one-in-six and the expected response: PDF 164–165; the dispute over starting points: PDF 163; the three-phase grand strategy: PDF 181; the nine items under "What not to do": PDF 202–204. Procedural limit: the reading notes on this book have not passed every stage of verification; this report uses only descriptive content checkable word for word against the original pages and none of the notes' judgments; after verification this section must be re-checked sentence by sentence. ↩
22. William MacAskill, What We Owe the Future (New York: Basic Books, 2022). Primary source: local full-text PDF, 880 pages, with text layer (primary full text); page numbers are pypdf's zero-based indices. Range read: the first three pages of the introduction, the first two pages of chapter 1, the AGI and value lock-in passage of chapter 4 (PDF 87–96), all of chapter 8 (PDF 176–201); chapters 2, 3, 5–7, 9 and 10 and all endnotes were not read. The three-sentence argument: PDF 15; the definition of longtermism: PDF 11; the passage on gradual loss of control to AI: PDF 94; the two pages of moral evaluation: PDF 95–96; the chapter 8 proposition and its four boundaries: PDF 176–178; acceptance of the Repugnant Conclusion: PDF 194; the compromise under moral uncertainty and the citation of Ord and Greaves: PDF 197–198; the practical upshot and the reverse limit on space settlement: PDF 200. Procedural limit as in ²¹. ↩
23. Marc Andreessen, "The Techno-Optimist Manifesto," a16z.com, 16 October 2023. "The Enemy" section (retrieved directly from the original page); the judgment that "the list names no organisation" comes from the accelerationism report's reading of the manifesto's full text including the whole "Enemy" section, recorded in that report's reading notes. The manifesto's other sections are not used in this report. ↩
24. Zoe Lofgren's letter of opposition to Senator Scott Wiener, 7 August 2024 (primary full text). The joint reading and item-by-item comparison of the four submissions has passed three stages of review. ↩
25. Two retrieval results on whether the e/acc side responded to d/acc and whether it mentioned left accelerationism (a failed-retrieval record: the public debate between the two on an a16z podcast on 9 April 2026 can be confirmed, but there is no transcript and the specific positions taken in the debate were not obtained) and the relevant file (two public e/acc-related texts mention Srnicek and Williams, both as genealogy). ↩
26. Anthropic's letter to the Governor of California, 21 August 2024 (primary as to "what Anthropic said"). Disclosure: the system that wrote this report runs on Anthropic's models, and Anthropic is a party to this dispute; this subsection only paraphrases the letter, neither evaluating its position nor using it to support any side. ↩
27. Gavin Newsom's veto message, 29 September 2024 (primary full text). ↩
28. Open letter from a16z chief legal officer Jaikumar Ramaswamy to Senator Wiener, 2 August 2024, fourteen pages (third-party copy; the hosting party, Engine Advocacy, is neither sender nor recipient and stood with a16z on the opposing side of this dispute, so the grade is not raised to primary). The zero-result search for the four word roots covered the 492-line extracted full text of the letter. ↩
29. Joint letter of current and former frontier-lab employees, addressed to the Governor of California. The evidence grade of this document is split four ways by claim (the page's text: primary / the letter's existence and request: third-party copy / the submission date: single source / the authenticity of signatures: single source), and the three counts of signatories are not merged. ↩
30. Text of SB 1047 (enrolled version of 3 September 2024) and the bill's own adaptive mechanisms (the full official digest plus the text of sections 22602 and 22603; other sections not transcribed). Version differences are handled as set out in the joint reading of the four primary SB 1047 texts. ↩
31. Anjney Midha, "What You Need to Know About SB 1047: A Q&A," a16z.com, 19 June 2024 (a public position text on the firm's own domain). ↩
32. OpenAI, "OpenAI announces leadership transition," 17 November 2023 (the party's original statement; the current openai.com address returns 403, so the Wayback Machine snapshot from the day of publication was used). The board list of the time and Toner's institutional affiliation both come from this document. The joint reading of the four primary texts on the crisis has passed three stages of review. ↩
33. Helen Toner's interview on The TED AI Show, official transcript, episode dated 28 May 2024; the same episode closes by reading in full the response statement of OpenAI chair Bret Taylor (official transcript; Toner's remarks are a party's retrospective account six months after the events and after the internal investigation had ended, a one-sided account with a position, and that character must accompany any citation; the transcript page itself notes possible minor differences from the audio). ↩
34. Grant records in the same category of the Open Philanthropy official grants database: Center for Security and Emerging Technology (CSET), three grants totalling $101,920,000, the largest recipient in the category; OpenAI, one grant of $30,000,000 dated 27 March 2017, the fourth-largest single grant in the category. Both are raw field values from the database (primary full text), from the "largest 15 grants" table. Limit: this report cites only these two raw values, not the category share computed by the excerpt file itself (that share is graded third-party in the joint reading notes, because it was summed locally, 68 records without amounts were counted as zero, and the record count shown on the site's own page differs from what the interface returns). That funding proves no one's motive for a vote is stated in the main text. ↩
35. Marc Andreessen, "The Techno-Optimist Manifesto," a16z.com, 16 October 2023, "The Enemy" section (signed text on the firm's own domain). The list quoted in the main text is a word-for-word transcription of the passage. The section nowhere contains the words "Effective Altruism"; what it names are concept labels, as the text obtained shows; the file holds the full "Enemy" section only, not the rest of the manifesto. This report does not on that basis assign the concept of "existential risk" to effective altruism. The concept was introduced by Nick Bostrom in 2002 (in the 2013 article cited at ¹¹ he notes "Including the very notion of existential risk (Bostrom, 2002)", and lists Bostrom, N. (2002) "Existential risks: analyzing human extinction scenarios and related hazards" in the bibliography), whereas the name "effective altruism" was fixed only by the vote of seventeen in 2011, nine years later. Claiming an earlier concept for a later name is exactly the move Section 2's parallel naming history describes, and it is not repeated here. Andreessen's list names concept labels; this report states that fact only, the relation between the two requires further material, and the gap is listed at the end. ↩
36. Center for AI Safety, "Statement on AI Risk" (page title "Statement on AI Extinction Risk"). Graded by claim: the sentence of the statement and the signatories' names and titles as listed on the page are primary (extracted word for word from the publisher's own page, retrieved 21 August 2026); the date of publication is third-party—the page itself carries no date, and it is dated to late May 2023 from Wayback records (earliest capture 27 May 2023, four captures on 30 May 2023), the exact date of publication not independently verified. The page expands only part of its signatory list by default; the main text relies on those 51 names, the "Show more" control was not expanded, so the list is not complete. The statement uses "risk of extinction", not "existential risk"; the two are not interchanged in this report. ↩
37. The Bletchley Declaration by Countries Attending the AI Safety Summit, 1–2 November 2023, full text from the UK government's GOV.UK page (primary; retrieved 21 August 2026). The two phrases quoted in the main text are "There is potential for serious, even catastrophic, harm, either deliberate or unintentional" and "unintended issues of control relating to alignment with human intent". Twenty-eight countries plus the European Union are listed; New Zealand joined on 23 October 2024. GOV.UK also offers translations, which this report did not retrieve; the English original governs. ↩
38. Word count on the text of SB 1047: a mechanical search of the file returns zero occurrences of "catastrophic"; the bill's definitions of the harms it covers are in the same file. This limit must accompany any citation: the file holds the full official digest plus the text of sections 22602 and 22603 only, so the zero holds for the range transcribed, and does not mean that the word occurs nowhere in the full bill. To widen it into a statement about the whole bill, the complete text must be obtained and recounted. The comparison (programmatic documents and company letters use this vocabulary; the statute switches to measurable thresholds) is this report's own arrangement of the material; the main text still only sets the two sides' facts side by side and draws no causal inference—a limit that comes from the material itself. ↩
39. Giving What We Can, official public list of pledgers, https://www.givingwhatwecan.org/about-us/members, searched name by name after expanding the full table on 21 August 2026 (primary; the publisher's own list; sequence numbers, names and months as shown on the page). On that day the page stated 11,166 lifetime pledgers from 118 countries. The three items in the main text: Dario Amodei, no. 43, July 2010; Jack Clark, no. 4002, February 2019; Dustin Moskovitz, no. 8458, April 2010. Two limits: first, the list allows members to be anonymous or to have their data removed, so it gives only a lower bound, and absence from it does not mean a pledge was never made; second, the sequence numbers and the months are not synchronised (the entries adjacent to Moskovitz are dated January 2021 and April 2023), so the numbers cannot be used to infer order, and the main text uses months only. ↩
40. Ben Mann, "Why I now identify as an Effective Altruist," Medium (@8enmann), 25 March 2019 (the author's signed original; retrieved 21 August 2026). At the time of writing the author had not yet co-founded Anthropic (founded 2021); in the same essay he says he had earlier worked at OpenAI and MIRI and left disillusioned, and that once he had an income he would give five percent (below the movement's standard pledge of ten). The extraction was cut off a few sentences before the end; the paragraph quoted in the main text is complete and unaffected. ↩
41. The main text states the funding records only and does not infer anyone's motive for a vote. The two parties' later accounts are as follows, and both are recorded without being merged into a single explanation. Helen Toner said in an interview six months after the events that the board could not get accurate information from the company and could not carry out its oversight role; OpenAI chair Bret Taylor's subsequent statement said that the review completed by an independent committee of the board with a law firm "concluded that the prior board's decision was not based on concerns regarding product safety or security, the pace of development, OpenAI's finances or its statements to investors, customers, or business partners". Both are sourced as at ³³ (official transcript; Toner's remarks are a party's retrospective account after the investigation, with a position, and that character accompanies the citation). ↩
42. Centre for Effective Altruism, "CEA's Guiding Principles" (extracted word for word from the publisher's own page; retrieved 21 August 2026). The definition reads "Effective altruism is about using evidence and reason to figure out how to benefit others as much as possible, and taking action on that basis."; the five principles are headed Commitment to Others, Scientific Mindset, Openness, Integrity and Collaborative Spirit, and the main text paraphrases their content. Two limits: first, the page carries no date of publication or revision, so when these five were adopted and how often revised cannot be known, and this report makes no statement about their dating; second, the page lists 11 endorsing organisations and 5 endorsing individuals (including Peter Singer), from which this report infers nothing about any party's involvement in drafting the five. ↩