Vitalik's sharp critique of Anthropic: safety rhetoric's link to power expansion exceeds 90%
Vitalik donated billions for AI safety, yet now says AI labs preach safety while their actions align over 90% with expanding their own power.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Anthropic's succession is designated by founders and kept from the public
Anthropic's prospectus uses "safety" 248 times—"AI safety" 36 times—and its seven co-founders will control 50.1% of voting rights on ordinary matters through special shares; a long-term trust can elect a board majority. More critically: founders can pre-designate successors for voting power after losing office, requiring internal approval but with names kept secret—the public doesn't know who might decide Anthropic's future.
— YiwenEffective altruism shifted from global health to AI as its core
Afra splits EA's decade-plus history into four phases: the mosquito-net era of quantifiable evidence for global health; around 2015, the long-termist shift toward future generations and existential risk, when SBF and other billionaires poured in money; post-FTX crisis, amid ChatGPT's emergence, an era of trust loss; then today's AI safety as core agenda—making EA a phrase ordinary people now know.
— AfraAnthropic's auditors have financial stakes in the company's success
METR, which audits Anthropic independently, is funded mainly by Coefficient Giving, whose founder Dustin Moskovitz is an early Anthropic investor holding major shares. Afra sees no conspiracy—but the real issue is: when a thought movement gets vast funding, it doesn't just find important problems; it decides which problems count as important. That's the AI policy world today: young researchers are being hired away in high-paying AI-safety roles.
— AfraDonating later enabled early ruthless profit-chasing in the name of effective altruism
MacAskill told Oxford students that a doctor saves ~140 lives, but earning on Wall Street or in consulting, then donating wisely—buying more nets—can save ten times as many. Afra argues this logic later became moral cover for quant traders (including SBF) to justify ruthless profit-chasing: just claim you'll donate it, and greed becomes charity.
— AfraShiba Inu accidentally created billions in AI safety funding Vitalik couldn't control
Vitalik recalls Shiba Inu's team putting half the token supply into his public wallet. He assumed it would collapse fast, so distributed it to a few organizations—figuring they might extract $20+ million max. Three months later, no crash. Combined cashing out: ~$470 million for India's CryptoRelief, $400–600 million for FLI—far beyond his expectation or control.
— VitalikAI safety no longer meets the criteria for effective altruist funding
Vitalik's core test for whether to fund a field: neglectedness—is this problem important but understaffed? Ten years ago, AI safety and longevity research were lonely. Today AI safety has vast funding, roles, and attention. The issue isn't scarcity of resources; it's that current methods may be wrong. And that type of problem has no simple fix.
— VitalikAnthropic's safety arguments correlate over 90% with expanding its own power
Vitalik says he understands Anthropic's every argument—accept the premises, the conclusions follow—but watching what it actually does, he sees correlation with expanding power "likely over 90%." His question: why should a company decree "we're safe, so we win, others shouldn't have power"—while attacking open AI as dangerous, yet still pushing the AI race forward?
— VitalikDecentralized technologies offer better AI defense than concentrated corporate control
Vitalik proposes d/acc: technology advance is worthwhile, but invest first in tools protecting regular people through decentralization, not concentrating power in a few models or firms. Example: some EA organizations spend 80% of resources on "making AI refuse to help make pathogens"—centralized defense, useless if rivals have their own models. Real protection: detection, filtering, faster vaccine R&D—decentralized means.
— VitalikIn their own words · checked verbatim
That means founders can pre-designate who replaces their voting power if they lose capacity—those picks need internal approval but the list is secret. In other words, the public doesn't know today who might have a say in Anthropic decisions tomorrow.
那就是创始人可以提前指定 在自己失去履职能力后 由谁接替自己的投票权 这些人选需要经过内部批准 但名单是保密的 也就是说公众目前并不知道 未来谁是有可能参与 Anthropic决策的人
Yiwen1:01
They don't want AI to become uncontrollable and cause consequences humans can't bear. That's their agenda. There's nothing hidden about it.
就是他不想让AI变得不可控 同时造成让人类无法承受的后果 这就是他们的议程 其实没有什么隐藏的议程
Afra19:10
Your future self will always betray your present self, because in 10, 20 years your beliefs change, your actions change, your friends and circles change, what you want to do changes.
你未来的自己 永远会背叛你现在的自己 因为你10年 20年之后 你的想法会变 你的行为会变 你身边的朋友 你的圈子会变 你想做的事情会变
Afra27:15
Each 2000 yuan saves a human life—yet most people would spend 2000 yuan on, say, a business-class ticket.
你每次付2000块 你就可以救一个人的命 其实大部分的人 他们会花2000块做什么 有可能是一个公务舱的机票
Vitalik41:22
I myself in 2016 spent $25,000 buying some Dogecoin—my money turned into maybe $10+ million.
其实我自己在2016年的时候 我花了25000美元买一些狗狗币 我的钱都变成了 可能1000多万
Vitalik1:00:37
Man versus nature is actually simpler. Man versus man is very complex, because the challenge you're fighting is as smart as you are, so the effects of what you do can be very intricate.
Man versus nature其实更简单 Man versus man是特别复杂的 因为你在对抗的那个挑战 是跟你一样聪明的 所以你做的事情能有的效果 有可能是会非常复杂的
Vitalik1:17:54
There are roughly three points. First, acceleration—technological progress matters. Second, defense—we need to invest more ability now in technologies that protect us.
大概有三个点 第一个点是acceleration 就是技术的发展是重要的 第二个点就是defensive 我们现在需要花更多的能力 做一些可以保护我们的技术
Vitalik1:38:10
Figures
| Times "safety" appears in Anthropic prospectus | 248 (AI safety: 36) | 0:01 |
| Voting power controlled by Anthropic founders | 50.1% (ordinary matters) | 1:01 |
| Coefficient Giving cumulative grants since 2014 | Over $7 billion | 37:21 |
| Vitalik's estimated cryptocurrency donations in 2021 | Approximately $1.5 billion | 38:21 |
| Final cash value of Vitalik's tokens donated to FLI | Approximately $500 million | 6:03 |
| Vitalik's estimated P(doom) | 0.1 | 55:31 |
Glossary
- Effective altruism
- Using evidence and rational analysis to allocate limited money and time to where they bring the most good.
- e/acc
- Advocates accelerating AI and technology development; opposes slowing it for safety reasons.
- d/acc
- Proposed by Vitalik: accelerate technology while prioritizing defenses that protect ordinary people over concentrated control.
- Neglectedness
- An EA standard for assessing donations: whether a field still lacks sufficient attention and resources.
- P(doom)
- Subjective estimate of the probability that AI or other risks cause human extinction.
- Orthogonality thesis
- AI's intelligence level and whether its goals align with human interests are two independent dimensions.
How to listen
Entrepreneurs and investors tracking AI policy and philanthropic money flows; those seeking to understand how the EA/rationalist community shaped Silicon Valley's AI narrative.
The first half has Afra tracing EA history. If you're already familiar with MacAskill and Yudkowsky, skip to the dialogue.