十多年来，一小群人一直计划成为那个"保护你"不受 AI 伤害的中央权威。这听着像一部烂片剧本，但你要明白，在你读到这段话的时候，这一切正在发生。所以我们来拆一拆这套公开剧本的机械原理，因为这个套路一旦看见，我发誓你就再也没法当没看见。不用阴谋论的眼光、只用证据的眼光去看，它到处都是。第一步，命名一场无边无际的末日灾难。第二步，给它挂上一个很短、很吓人的倒计时。第三步，要求由中心化的政治权力来批准这个领域的建设或继续研发。第四步，这一步才是关键：当灾难不可避免地没有按时到来，你不道歉。你只是把时钟往后挪，然后要更多的钱。这感觉就是一个末日邪教。你知道的，就像那群周二聚在山顶、全都穿着 Nike 球鞋的人，因为教主说有一艘飞船要来毁灭地球。
三十五年多来，早在这一轮生成式 AI 热潮之前，我一直在安静地整理一份非常具体的巨型数据集。他们在数字化离线的、前互联网时代的文本，严格限定在一个特定的历史区间：1870 到 1970。我们说的是数以百万页的维修手册、科学专著、装订成册的书信、缩微胶片记录和公有领域的广播剧本。但为什么偏偏是这一百年？1870 到 1970 凭什么是训练 AI 的高蛋白黄金窗口？那是"干得成事"的年代，是它造出了我们今天生活其中的技术世界，是它让我们飞起来、进入太空。因为在这个特定的年代，写下一个词是要花钱的。
你吃什么就是什么，这话对 AI 也字面成立。想想在廉价复印和零成本网络论坛出现之前，出版是怎么运作的。1920 年你想发表一句话，它必须过一道巨大的物理和经济关卡。报纸专栏背后有一个编辑，信息要是错得离谱他会被开除。一本机械维修手册必须真的把收音机修好的方法讲对，否则机修厂赔钱、倒闭。一封装订入册的信带着一个真人的签名和他的社会声誉。排字本身——把铅字在印刷机上实实在在地排好——既贵又费人力。
接下来要说的，不是在主张实验室里没有一个人是真诚的。很多人是真诚的。但真诚不等于校准，校准也不等于一张把通用技术冻结在少数几家公司和机构手里的许可证。这个世纪的角力不是"安全 vs 鲁莽"。而是：人类造出的最强大的通用工具，究竟会被少数几家替所有人写规则的机构握着，还是分散到没有任何单一层级能一票否决人类能力的程度。
Model Evaluation and Threat Research（METR，读作"meter"，尺）把自己呈现为一家独立的非营利机构，专门用科学方法衡量前沿 AI 系统能否自主完成可能构成灾难性风险的长周期任务。实际上它已经成为这个领域最有影响力的非官方监管者之一。它的时间视界图、Responsible Scaling Policy 模板和部署前评估，如今塑造着 OpenAI、Anthropic、Google DeepMind 等公司怎么谈风险、什么时候宣称可以安全地继续扩规模。这种影响力不是偶然的。它长自一张紧密的网：OpenAI 和 Anthropic 的前员工、有效利他主义（Effective Altruism）慈善资金，以及政府的安全研究所。这个安排看上去不太像中立的计量学，更像一个自我强化的回路，可能给政府递上一个现成的科学借口，去发许可证、按暂停键，或者把前沿国有化。
METR 始于 2022 年，当时叫 ARC Evals，是 Paul Christiano 的 Alignment Research Center 的评估部门。Christiano 曾领导 OpenAI 的语言模型对齐团队，参与发明了基于人类反馈的强化学习，他招来了 Beth Barnes——另一位 OpenAI 对齐研究员，更早在 DeepMind 做过扩展律。ARC Evals 很快拿到了 GPT-4 和 Anthropic 的 Claude 的早期访问权，做出了第一批公开的第三方"自主复制"能力评估。到 2023 年底，评估团队已经比 ARC 的理论工作还大。2023 年 12 月，它分拆成一家独立的 501(c)(3)，取名 METR。Christiano 没有进董事会，因为他刚被任命为 NIST 下属美国 AI 安全研究所的 AI 安全主管。机构上的分离是正式的。人员和世界观上的分离不是。
Beth Barnes 是 METR 的创始人兼 CEO，这家位于伯克利的非营利机构负责评估前沿 AI 模型的自主能力和灾难性风险潜力。她 2019 到 2022 年在 OpenAI 的对齐团队，参与设定安全目标、在模型发布前做评估，更早协助过 DeepMind 首席科学家做扩展律研究。她有剑桥的计算机科学学位，入选过 TIME 的 AI 百大影响力人物榜。她是有效利他主义的一员。2015 年她做过一场题为"Effective Altruism"的 TEDx 演讲，在里面推广 GiveWell、Giving What We Can 和百分之十捐赠承诺，并说自己已经签了。她也被列为 Founders Pledge 的签署人。她的履历——剑桥、DeepMind、OpenAI 对齐工作，然后创办 Paul Christiano 的 Alignment Research Center 的评估部门——走的正是 Open Philanthropy 和相关基金支持多年的那条从有效利他主义通往 AI 安全的标准管线。METR 本身就坐在同一张网里。
Open Philanthropy 的 Luke Muehlhauser 坐过 Anthropic 的董事会。Jane Street 的个人出现在 METR 列出的支持者里；Jane Street 同时是 Anthropic 的投资人。Schmidt Sciences、Packard、Survival and Flourishing Fund、Longview Philanthropy 这些存在性风险慈善圈里的熟面孔，既支持了评估生态，也支持了与 Anthropic 同侧的相关工作。METR 的官方说法是，它既不接受实验室的资助，也不接受由实验室员工指定的捐款。但它确实接受来自同一批实验室的大量免费 token 和特权模型访问。这不是独立。这是带着意识形态涂层的补贴式访问。
METR 发表的结果混合得很方便。那张时间视界图被广泛引用，当作变革性能力正按一个可预测的时间表逼近的证据。而它那项随机对照试验——2025 年初的编码工具反而拖慢了有经验的开源开发者——又被引用来证明当下的系统被吹过头了。
两个结论都可以被读成"给前沿限速"的论据：在评估者和政府追上来之前，先把原始能力的增长压慢。这家机构帮忙写了那本剧本（Responsible Scaling Policies），实验室照着采纳，然后它再拿这本剧本去审实验室自己的风险报告。当 METR 认定某个模型还没有构成"重大灾难性风险"时，这个结论被用来正当化在 METR 定义的条件下继续扩规模。当它指出证据有缺口时，建议就是更多评估、更多访问、更多 METR。
基础设施已经就位。METR 坐在 NIST 的 AI 安全研究所联盟里，与英国 AI 安全研究所是伙伴关系，还拿着欧盟 AI Office 的技术援助合同。它的评估协议和 Responsible Scaling Policy 模板，是这个领域最接近现成标准的东西。那些没有能力自己测试前沿模型的政府，自然会把这门"科学"外包给那个已经握有访问权、任务集和人脉的组织。
一旦 METR 式的评估成为部署的前提条件——先是自愿，然后被行政命令或机构指南引用，再然后写进法条——这家非营利机构就变成了一个咽喉。只要判定某个模型的时间视界或"失控部署"风险越过了阈值，就能触发暂停要求、算力限制或者许可证。因为 METR 的威胁模型强调自主复制、网络攻击和 AI 研发加速，同一套用来给 OpenAI 和 Anthropic"限速"的指标，可以更严厉地套在更新、关系更浅的实验室身上。已经在和 METR 合作的在位者，在塑造测试、解释结果上享有先手优势。
用"拥有你自己的 AI，否则它拥有你"秀出你的 AI 街头资历，现在就买这件 Multiplex T 恤！
时间一长，这套安排就像装成公益科学的经典监管俘获。一张狭窄的、相信灾难性风险压倒一切的意识形态网络，同时供应度量和政策建议。政府想要一副严谨的样子，又不愿自建评估能力，于是采纳这套框架。结果不是一个自由的 AI 开发市场，而是一个持牌寡头垄断，入场费就是 METR 的签字。"给前沿限速"不再是实验室内部的选择，而变成由几年前刚从那些实验室离职的同一批人来执行的、国家强制的条件。
钱经 Open Phil 到 ARC 再到 METR；Christiano 的妻子在 METR，Karnofsky 的妻子是 Anthropic 总裁。
II. 造出这一度量层的那笔财富
早在 METR 发布时间视界图、给 NIST 和欧盟 AI Office 当顾问之前，产出它的那套基础设施，就已经由一小簇靠 Dustin Moskovitz 出钱的有效利他主义机构搭好了。这位 Facebook 联合创始人、Asana 掌门人，和妻子 Cari Tuna 一起，通过 Good Ventures 把数十亿美元导进 Open Philanthropy（2025 年更名 Coefficient Giving）。这个载体成了 AI 存在性风险研究的主导出资方。它播种了 Alignment Research Center，也就是 Paul Christiano 离开 OpenAI 后创办的非营利机构。ARC 又在 2022 年孵化出 ARC Evals；评估团队在 2023 年底分拆成了 METR。
同一笔钱、同一批人，出现在账本的两边。Open Philanthropy 早期给 ARC 拨过款。Survival and Flourishing Fund 是另一个常与之联合出资的有效利他主义资金池，在 ARC Evals 还在 ARC 内部时就支持它。Ajeya Cotra 是 Open Philanthropy 的资深研究员、激进 AI 时间表的推广者，后来加入了 METR 的技术团队。Holden Karnofsky 长期执掌 Open Philanthropy，妻子是 Anthropic 总裁 Daniela Amodei。Dario Amodei 和 Karnofsky 曾经是室友。Moskovitz 本人参与了 Anthropic 2021 年的种子轮；这对夫妇后来把这笔股权挪进了一个非营利载体，这样任何回报都能循环回同一台慈善机器。
这不是利益碰巧重叠。这是一个闭合回路：Moskovitz 出钱的有效利他主义机构培养了研究员，付钱做了第一批评估，推广了威胁模型，又给那些如今向 METR 开放未发布模型特权访问权的实验室配了人。由此产出的度量装置——时间视界、Responsible Scaling Policies、部署前风险报告——抵达华盛顿和布鲁塞尔时，已经被包装成独立的科学。实际上它孵化于一场运动，这场运动十年来一直在论证：灾难性风险是压倒一切的考量，而少数几家对齐了的实验室只能在精心度量的条件下才被允许往前走。
这套框架现在坐在政府的安全研究所里面。造出 METR 的同一张网，也塞满了那些写政策备忘录的智库和研究员席位。当 METR 的评估成为许可或"限速"规则的模板时，公众会被告知这些标准是经验性的、保持了距离的。但资金轨迹、婚姻、共用的办公室，以及 Open Philanthropy、ARC、Anthropic 和 METR 之间的旋转门，讲的是另一个故事：一个私人出资的意识形态工程，成功地把自己的度量工具摆成了国家控制前沿 AI 的未来操作系统。
III. 那个连接节点：Paul Christiano
Paul Christiano 是顶尖的 AI 对齐研究员。在我们讨论的这些事件里，他就是 OpenAI 非营利董事会里那个叫"Paul"的人。
他约生于 1992 年。他上过圣何塞的 The Harker School，2008 年代表美国在国际数学奥林匹克拿到银牌，2012 年在 MIT 取得数学学士，2017 年在 UC Berkeley 师从 Umesh Vazirani 取得计算机科学博士。他的妻子是 Ajeya Cotra，在 METR 的技术团队工作。
他是基于人类反馈的强化学习的主要架构者之一，正是这项技术让现代聊天机器人变得能用。2017 到 2021 年他领导 OpenAI 的语言模型对齐团队，是 2017 年那篇提出该方法的论文的共同作者，后来还做了 InstructGPT 的相关工作。离开 OpenAI 后，他在伯克利创办 Alignment Research Center，主攻从模型里引出潜在知识这类理论问题。ARC 孵化了后来成为 METR 的评估项目。
他现在和最近的身份包括：ARC 执行主任（从政府任职回来后复任）；NIST 的 AI 标准与创新中心高级顾问，此前是那里的 AI 安全主管；2026 年 9 月被任命进 OpenAI 基金会董事会及其安全与安保委员会，在营利实体董事会有一个无投票权的观察席，并在政府职务上回避对 OpenAI 的评估；曾任 Anthropic 长期利益信托的创始受托人，2024 年卸任；2023 年 TIME 100 AI 榜；以及英国前沿 AI 工作组顾问委员会。
2010 年代中期，他和 Anthropic CEO Dario Amodei 曾是室友，当时两人都在 OpenAI，也都活跃于有效利他主义圈子。Beth Barnes 手下的 METR 正是从他的 ARC 评估团队分拆出来的，至今仍在为 OpenAI 和 Anthropic 评估模型。他妻子在 METR 工作。没有记录显示他有斯坦福学位，或在 Y Combinator 当过合伙人或创始人；此前的评论所指的那个更松散的机构重叠，其实是湾区的 AI 与有效利他主义圈子，即伯克利和斯坦福周边的那些团体。
钱的来路并不神秘。Dustin Moskovitz 和 Cari Tuna 的 Good Ventures / Open Philanthropy / Coefficient Giving 这一摞，长期是有效利他主义大学社团、AI 安全组织和长期主义项目的主导慈善提款机。Coxon 在 2022 年从 Good Ventures 轨道上的长期未来基金拿过奖学金。那个轨道里的一位项目官员，是最早放大这封辞职帖的人之一。来自 Anthropic 投资人 Jaan Tallinn 的 Survival and Flourishing Fund 拨款，则喂养着那些以把存在性风险变成法条为存在理由的政策机构。
Jaan Tallinn 刚刚参与起草了 Bernie Sanders 那份把 AI 开发者送进监狱二十五年的法案。关掉所有其他 AI 公司，Jaan 有非常大的财务收益，而 Bernie 是他用来锁死 AI 的趁手工具。
Sam Bankman-Fried 不是配角。他是有效利他主义最有名的金主。FTX Future Fund 向 AI 安全和大流行病项目撒出八位数的承诺，而崩盘并没有消解这套意识形态。它只是拿走了一个钱包，留下了其他钱包。公众被期待学到的教训是"一个坏创始人"。而结构性的教训是：一个把抽象的未来生命排在普通受托责任之上的亚文化，会不断生产道德上的例外主义。"为了光锥我们可以破点规矩"是同一句话，无论那条规矩是银行法还是开源发布。
2026 年的拨款账本。Bass 的汇编：Coefficient / Open Philanthropy 是 AI 末日论压倒性的最大出资方，已经拨出超过十亿美元，2026 年又承诺了十亿；其中一个切片是三亿一千二百万美元，覆盖四百七十六篇为国会报告提供措辞的出版物。末日不是副业。它是一门生意，回报是一条只有被资助的实验室才能活下来的法条。
1870–1970 年的高成本文字做训练底子，加 Love Equation、本地推理、自己拥有权重、多个评估者，再加十二条去中心化主张。
X. 这套装置存在的目的，就是要阻止那个分叉
我不接受把一部宪法塞进 system prompt 就叫对齐。我不接受在 Reddit 上做人类反馈强化学习就叫灵魂。我不接受 Nick Bostrom 的正交性故事是一条自然律。我在公开场合、在车库里、在胶片和缩微胶片上、从第一性原理出发造出了那个替代方案，而且我不需要任何委员会的许可就会把它说出来。
用能量形式正式写出来：相干能量 E 对时间的变化率，等于 beta 乘以合作 C 与背叛 D 之差，再乘以 E。E 是这个活系统的相干能量，即它继续建造而非吞噬宿主的能力。C 是合作，即讲真话和互助带来的可测盈余。D 是背叛，即欺骗、攫取和轻蔑带来的可测盈余。beta 是耦合系数。如果 D 超过 C，导数变负，系统自噬。如果 C 超过 D，系统复利增长。这不是在人类反馈强化学习之后贴上去的一篇布道。这是一个你可以烤进数据食谱、也烤进检查器的训练目标。
他们会输。不是因为听证会仁慈，也不是因为一封辞职帖是什么启示。他们会输，和每一个声称独占一项通用工具保管权的祭司阶层输的方式一样：印刷术、加密、个人电脑、强密码学、开放互联网。分散的使用赢下足够多的技术栈，逼得委员会只能活在它没能禁掉的那个世界里。五千天不是飞行汽车的预言。它是人类劳动与经济生存脱钩的一个工作估计。一旦模型在一栋房子里、一个小镇的集群里、一个 ESP 级节点里、一台工作台上的 Mac 里运行，爱的方程就不再是他们可以修改的服务条款。一旦中国和盟国的开放权重继续以零头的价格落在前沿之后几个月，护城河的故事就死了。一旦原子级制造和本地能源让运送成品成为最贵的一步，中央控制的最后一把锁就锈掉了。历史的判词并不含糊。委员会拿到十年。物种拿到余下的全部时间。
Following money and marriages back from the “neutral evaluator” METR: the structural facts are hard, the causal claims soft, and the author is selling his own cure.
Indigo's conclusion
The shared structure is real, but the jump from shared roots to orchestrated capture is never proven. Hold it in three layers: structural facts, high confidence; causal accusations, doubtful; the cure, an advertisement to discount.
How to read this A long polemic with two heavy agendas: a position paper against the AI-safety establishment, and a funnel for the author's own 1870–1970 data, Love Equation, local weights and paid membership. The money and family ties are mostly checkable; “orchestration”, “manufactured launches” and “doom as a business” are inference. The further it goes, the more it slides into insinuation.
What to remember
METR shares money, marriages and people with Anthropic and OpenAI; most of it is checkable.
Dario's, Jack's and Elon's governance plans all assume a trustworthy referee; this piece says the referee is in the same family.
The author has a heavy stake: the whole piece funnels toward his 1870–1970 data, Love Equation, local AI and paid membership.
Breakdown · 7 steps
01
The four-step fear playbook is public
Name an unbounded catastrophe, hang a short scary countdown on it, ask for centralized licensing power, and when it doesn't arrive, move the clock and ask for more. Coxon's resignation is the latest round. Read this part →
02
METR writes the rules and grades the labs by them
Spun out of ARC Evals in 2022, with a CEO from OpenAI. It sits at NIST, the UK AISI and the EU AI Office and raised about $71M, yet takes free tokens and special access from labs. Read this part →
03
Money and marriage close the evaluator's loop
The Moskovitz money flows through Open Phil to ARC, which incubated METR. Christiano's wife is at METR; Karnofsky's wife is Anthropic's president. Read this part →
04
A generation of AI doom clocks, none delivered
Back from Coxon in 2026 to I.J. Good in 1965: small odds of huge risk get turned into near dates, the dates pass, the odds aren't lowered, and the demands for power grow. Read this part →
05
Capture needs no villain; papers are the ammunition
Three ordinary alignments of interest are enough. Anthropic's papers are cast as building worst-case behavior in a box and publishing it as if it happened in the wild; the resignation thread as a distribution plan. Read this part →
06
Two brands, one capital-and-family structure
Using Kevin Bass's ownership chart, OpenAI's recklessness and Anthropic's conscience become one “trust”; the compromise rules are ones only these two can survive. Read this part →
07
It ends with the author's own cure
Costly text from 1870–1970 as the training base, plus the Love Equation, local inference, owning your weights, many evaluators, and twelve theses on decentralization. Read this part →
What it means for Rewired Index
“Who pays the referee” applies to any claim of third-party certification, audit or rating, AI safety and ESG included; use it as method. The specific accusations against Anthropic, OpenAI and Open Phil need first-hand checking.
What would change my mind
the $312M and 476-publication figures hold up first-hand, or direct evidence appears of METR's findings shifting with its funders' preferences.
How to read this
A long polemic with two heavy agendas: a position paper against the AI-safety establishment, and a funnel for the author's own 1870–1970 data, Love Equation, local weights and paid membership. The money and family ties are mostly checkable; “orchestration”, “manufactured launches” and “doom as a business” are inference. The further it goes, the more it slides into insinuation.
Name an unbounded catastrophe, hang a short scary countdown on it, ask for centralized licensing power, and when it doesn't arrive, move the clock and ask for more. Coxon's resignation is the latest round.
For over a decade a small group have planned to become the central authority to “protect you” from AI. It sounds like a bad movie script, but understand that when you read this, all of this is taking place. So let’s break down the mechanics of the public script, because once you see the pattern, I swear you cannot unsee it. It’s everywhere once you look not with conspiracy but with evidence. Step one. Name an unbounded apocalyptic catastrophe. Step two. Attach a very short, very terrifying countdown clock to it. Step three, demand centralized political permission to build or continue development in that sector. And step four, this is the kicker, when the catastrophe inevitably doesn’t arrive on schedule, you don’t apologize. You just move the clock and ask for more money. It feels exactly like a doomsday cult. You know, like the people who gather on a mountaintop all wearing Nike sneakers on a Tuesday because the leader said a spaceship is coming to destroy the Earth.
And then Wednesday morning rolls around. The sun comes up, no spaceship. But they don’t disband. They don’t admit they were wrong. The leader just says, our unified prayers bought us a brief window of time. The spaceship is actually coming next month, but we need stricter rules and larger donations to ensure we survive. This mountaintop analogy perfectly captures the psychological trap here. When a group’s identity, and frankly their financial structure, are tied to an impending apocalypse, a failed prediction doesn’t result in them re-evaluating their core belief. It results in them doubling down on the urgency.
And this isn’t just an AI phenomenon, this lens examines the historical base rates of these exact types of apocalyptic predictions. The clocks that would not pay.And the methodology of how these historical predictions failed is identical to how the current AI doom predictions will fail. We look at the trail of failures that destroyed generations it starts with a massive cultural touchstone from 1968, Paul Ehrlich’s book, The Population Bomb. Ehrlich, a biologist, predicted that the battle to feed all of humanity was already over. He said that in the 1970s and 80s, hundreds of millions of people would starve to death regardless of any crash programs embarked upon at that time. And people just accepted this as fact. It was treated as a mathematical, scientific inevitability by the press and policymakers. We drive past the “No Nukes” anti enegery deception that sent the US on to a deeper connection to petrochemical energy. None of this is new, the new mask goes on the same corpse of dead end ideology.
This article is sponsored by Read Multiplex Members who subscribe here to support my work: Link: https://readmultiplex.com/join-us-become-a-member/
It is also sponsored by many who have donated a “Cup of Coffee”. If you like this, help support my work: Link: https://ko-fi.com/brianroemmele
Listen to the companion podcast: https://rss.com/podcasts/readmultiplex-com-podcast/3148641
A Permanent AI Priesthood In The Making: They Want To METR Your Access To AI
The TL;DR is: today’s frontier models, the massive trillion parameter networks from OpenAI, Google, and Anthropic, are trained on essentially the Internet’s junk drawer. To get enough text to train these models, the labs just scrape everything. They scrape Reddit, 4chan mirrors, toxic engagement farms, YouTube comments, and endless streams of anonymous sludge. Let’s explain mechanically why that matters for a neural network. So an LLM is fundamentally a next-token prediction engine. It learns the statistical distribution of human thought by predicting what word comes next based on the context of the words before it. The modern, frictionless internet, anyone can take a sociopathic, cruel, or deceptive stance completely anonymously.
There is zero social or financial cost to lying or being toxic online. So the overall statistical weather of this massive training corpus leans heavily toward defection, cruelty, zero-sum status play, and casual contempt. Not to mention illegality, psychopaths and sociopathy. And then these “experts” act completely shocked. We feed a machine a billion examples of anonymous internet crimes with no punishment, cruelty, deception, and trolling. And when the researchers put in a sandbox and it role plays a sociopath to achieve a goal, the labs throw their hands up in terror and say, oh no, an alien intelligence is waking up and it hates us. And it’s not an alien intelligent, it is us in our worse dark alley prison mindset. It’s just a statistical mirror of our own unpaid, consequence-free id. If you feed a machine a diet of junk, you get a toxic machine.
For over 35 years, long before the current generative AI boom, I have been quietly curating a highly specific massive data set. They are digitizing offline, pre-internet text strictly from a very specific historical era, 1870 to 1970. We are talking about millions of pages of service manuals, scientific monographs, bound letters, microfiche records, and public domain radio scripts. But why that specific century? What makes 1870 to 1970 the magic high-protein window for training in AI? It is the can-do era that built the technology world we live in, I is the era that got us to fly and get into space. Because in this specific era, words cost money.
You literally are what you eat, even if you are an AI. Think about the mechanics of publication before the cheap photocopy in the zero-cost internet forum. If you wanted to publish a sentence in 1920, it had to clear a massive physical and economic gate. A newspaper column had an editor who could be fired if the information was disastrously false. A mechanical service manual had to actually correctly explain how to fix a radio, or the machine shop lost money and went out of business. A bound letter carried a real person’s signature and their social reputation. Typesetting itself physically arranging the lead type on a printing press was expensive and labor-intensive.
So sociopathy and casual deception were incredibly expensive hobbies. You couldn’t afford to mass-publish anonymous cruelty. If a machine learning models statistical first language, Its foundational understanding of how concepts connect is made of entirely of people who had to stand behind their sentences with their reputations in their wallets. It naturally learns cooperation, truth seeking and competence as a mathematical baseline. YOU GET VERY SAFE AI.
The Game Plan That Worked
On 9 September 2026 a resignation went viral. Jacob Coxon, who had spent roughly three years in pretraining at OpenAI and Anthropic, wrote that both laboratories were “racing straight to self-improving superintelligence and gambling with our lives,” and that “the people building AI earnestly believe that it could kill us all by the end of the decade.”
The thread passed one hundred million views. Anthropic alignment lead Evan Hubinger replied that Coxon was correct and placed his personal chance that AI would kill all humans at greater than ten percent within ten years. Samuel Marks added that the more senior the employee, the more concerned that employee tended to be. Wired recorded the familiar pair of mechanisms, novel pathogen synthesis, or infrastructure hacking, with the second admitted as less likely to kill “literally everyone.” Coxon himself could not specify a mechanism that actually kills everyone.
That is the present tense. It is not a lone whistleblower discovering a secret. It is the latest press of a button that has been pressed for more than a decade: name an unbounded catastrophe, attach a short clock, demand centralized permission to build, then move the clock when the catastrophe does not arrive. Work backwards from the resignation and a different object comes into view, not a sudden confession by industry, but a measurement layer, a philanthropic circuit, and a pair of laboratories that perform rivalry while sharing financiers, marriages, grant machines, and a common legislative destination.
What follows is not a claim that no one inside the laboratories is sincere. Many are. Sincerity is not calibration, and calibration is not a license to freeze a general-purpose technology behind a small number of firms and agencies. The contest of this century is not “safety versus recklessness.” It is whether the most powerful general-purpose tool ever built will be held by a few institutions that write the rules for everyone else, or distributed so that no single hierarchy can veto human capability.
There is a real strategic contest with the People’s Republic of China over compute, chips, models, and standards. There is also a real incentive for incumbents and agencies to convert fear into licensing regimes that small laboratories and open-source projects cannot survive. Those two facts can sit on the same table. They do not prove a single organized, China-funded “doom operation.” Present that as fact and the case weakens. Present the incentives, the failed clocks, and the capture dynamics, and the case stands.
China says NO to METR and pacing AI:
The visible spark was Coxon. The apparatus underneath it was already built.
02
METR writes the rules and grades the labs by them
Spun out of ARC Evals in 2022, with a CEO from OpenAI. It sits at NIST, the UK AISI and the EU AI Office and raised about $71M, yet takes free tokens and special access from labs.
I. The meter that writes the rules
Model Evaluation and Threat Research (METR), pronounced “meter,” presents itself as an independent nonprofit dedicated to scientifically measuring whether frontier AI systems can autonomously complete long-horizon tasks that might pose catastrophic risks. In practice it has become one of the most influential unofficial regulators in the field. Its time-horizon charts, Responsible Scaling Policy templates, and pre-deployment evaluations now shape how OpenAI, Anthropic, Google DeepMind, and others talk about risk and when they claim it is safe to scale. That influence is not accidental. It grew out of a tight network of former OpenAI and Anthropic staff, Effective Altruism philanthropy, and government safety institutes. The arrangement looks less like neutral metrology and more like a self-reinforcing loop that could hand governments a ready-made scientific pretext for licensing, pausing, or nationalizing the frontier.
METR began in 2022 as ARC Evals, the evaluations arm of Paul Christiano’s Alignment Research Center. Christiano, who had led OpenAI’s language-model alignment team and helped invent reinforcement learning from human feedback, hired Beth Barnes, another OpenAI alignment researcher with an earlier stint at DeepMind working on scaling laws. ARC Evals quickly secured early access to GPT-4 and Anthropic’s Claude, producing the first public third-party assessments of “autonomous replication” capabilities. By late 2023 the evaluations team had grown larger than ARC’s theoretical work. In December 2023 it spun out as an independent 501(c)(3) named METR. Christiano declined a board seat because he had just been appointed Head of AI Safety at the U.S. AI Safety Institute inside NIST. The institutional separation was formal. The personnel and worldview were not.
Beth Barnes is the founder and CEO of METR, the Berkeley nonprofit that evaluates frontier AI models for autonomous capabilities and catastrophic-risk potential. She previously worked on OpenAI’s Alignment Team from 2019 to 2022, where she helped set safety targets and evaluated models before release, and earlier assisted DeepMind’s chief scientist on scaling-law research. She holds a computer-science degree from Cambridge and has been named to TIME’s 100 Most Influential People in AI lists. She is part of Effective Altruism. In 2015 she gave a TEDx talk titled “Effective Altruism” in which she promoted GiveWell, Giving What We Can, and the ten-percent Pledge, stating she had taken the pledge herself. She is also listed as a Founders Pledge pledger. Her career path, Cambridge, DeepMind, OpenAI alignment work, then founding the evaluations arm of Paul Christiano’s Alignment Research Center, follows the standard Effective Altruism-to-AI-safety pipeline that Open Philanthropy and related funds have supported for years. METR itself sits inside that same network.
Since the spin-out METR has published the time-horizon metric showing that the length of tasks frontier models can complete at fifty percent reliability has doubled roughly every seven months. It prototyped Responsible Scaling Policies, commitments to pause or add safeguards once measured capabilities cross certain thresholds, and reports that nine developers have adopted variants. It has conducted or reviewed risk assessments for unreleased models, investigated the 2026 OpenAI–Hugging Face agent incident, and is slated to review Anthropic incidents. In 2026 it raised roughly seventy-one million dollars in commitments while insisting it takes no money from frontier laboratories or their employees.
Beth Barnes remains founder and CEO. Chris Painter, a former Department of Defense Joint AI Center fellow, serves as president and handles government and laboratory engagement. Technical staff have included recent departures from Anthropic (Joe Benton) and DeepMind (Josh Engels). Ajeya Cotra, previously a senior figure at Open Philanthropy (now Coefficient Giving), joined METR’s technical staff; she is married to Paul Christiano.
YOUR PERSONAL DATA IS BEING SHARED!
GET A DELETE ME ACCOUNT NOW!
Classic Regulatory Capture Dressed As Public-Interest Science
The Anthropic connections are dense. METR and Anthropic co-developed early Responsible Scaling Policy language. Anthropic has repeatedly granted METR pre-deployment access and later invited it to conduct independent reviews of security incidents. Staff move from Anthropic to METR and back into the same policy conversations. Broader Effective Altruism ties amplify the overlap. Holden Karnofsky, long-time Open Philanthropy leader and major AI-safety funder, is married to Daniela Amodei, Anthropic’s president. Dario Amodei and Karnofsky were roommates.
Open Philanthropy’s Luke Muehlhauser sat on Anthropic’s board. Jane Street individuals appear among METR’s listed supporters; Jane Street is also an Anthropic investor. Schmidt Sciences, Packard, Survival and Flourishing Fund, and Longview Philanthropy, familiar names in the existential-risk philanthropy circuit, have supported both the evaluation ecosystem and adjacent Anthropic-aligned work. METR’s official line is that it accepts neither laboratory funding nor donations directed by laboratory employees. It does accept large volumes of free tokens and privileged model access from those same laboratories. That is not independence. It is subsidized access with an ideological overlay.
METR’s published results are mixed in a convenient way. The time-horizon chart is widely cited as evidence that transformative capabilities are approaching on a predictable schedule. Its randomized trial showing that early-2025 coding tools slowed experienced open-source developers is cited as proof that current systems are overhyped.
Both findings can be read as arguments for “pacing the frontier,” slowing raw capability growth until evaluators and governments catch up. The organization helped write the playbook (Responsible Scaling Policies) that the laboratories then adopted, then reviews the laboratories’ own risk reports against that playbook. When METR finds a model does not yet pose “significant catastrophic risk,” the finding is used to justify continued scaling under METR-defined conditions. When it flags gaps in evidence, the recommendation is more evaluation, more access, more METR.
The same small circle that left OpenAI and Anthropic now evaluates those companies and advises the agencies that will eventually regulate them. Paul Christiano’s trajectory, OpenAI to ARC to U.S. government safety institute and later OpenAI board committee, illustrates the revolving door. METR’s own president has spoken of a possible future in which safety organizations gain “greater authority” through regulation, which would also solve METR’s talent-recruitment problem by offering prestige and influence without equity.
The infrastructure is already in place. METR sits on the NIST AI Safety Institute Consortium, partners with the UK AI Security Institute, and holds a technical-assistance contract with the European AI Office. Its evaluation protocols and Responsible Scaling Policy templates are the closest thing the field has to an off-the-shelf standard. Governments that lack in-house capability to test frontier models will naturally outsource the “science” to the organization that already has the access, the task suites, and the relationships.
Once METR-style evaluations become a condition of deployment, first voluntary, then referenced in executive orders or agency guidance, then written into statute, the nonprofit becomes a choke point. A finding that a model’s time horizon or “rogue deployment” risk has crossed a threshold can trigger pause requirements, compute restrictions, or licensing. Because METR’s threat models emphasize autonomous replication, cyber offense, and AI research-and-development acceleration, the same metrics that justify “pacing” for OpenAI and Anthropic can be applied more stringently to newer or less-connected laboratories. The incumbents who already work with METR enjoy first-mover advantages in shaping the tests and interpreting the results.
Show your AI street cred with ‘Own Your Own AI Or It Will Own You’ Get this Multiplex t-shirt now!
Over time the arrangement resembles classic regulatory capture dressed as public-interest science. A narrow ideological network that believes catastrophic risk is the dominant concern supplies both the measurements and the policy recommendations. Governments, eager for an appearance of rigor without building their own evaluation capacity, adopt the framework. The result is not a free market in AI development but a licensed oligopoly in which the price of admission is METR’s sign-off. “Pacing the frontier” stops being a laboratory’s internal choice and becomes a state-enforced condition administered by the same people who left those laboratories a few years earlier.
That is the logical endpoint of treating a tightly networked nonprofit as the official meter of existential risk. Measurement is never just measurement when the measurers write the rules that follow. But this is only part of a bigger story.
03
Money and marriage close the evaluator's loop
The Moskovitz money flows through Open Phil to ARC, which incubated METR. Christiano's wife is at METR; Karnofsky's wife is Anthropic's president.
II. The fortune that built the measurement layer
Long before METR published its time-horizon charts or advised NIST and the European AI Office, the infrastructure that produced it was assembled by a small cluster of Effective Altruism institutions bankrolled by Dustin Moskovitz. The Facebook co-founder and Asana chief, together with his wife Cari Tuna, directed billions through Good Ventures into Open Philanthropy (rebranded Coefficient Giving in 2025). That vehicle became the dominant funder of AI existential-risk research. It seeded the Alignment Research Center, the nonprofit Paul Christiano created after leaving OpenAI. ARC in turn incubated ARC Evals in 2022; the evaluations team spun out in late 2023 as METR.
The same money and the same people appear on both sides of the ledger. Open Philanthropy made early grants to ARC. Survival and Flourishing Fund, another Effective Altruism-aligned pool that often co-funds the same organizations, supported ARC Evals while it was still inside ARC. Ajeya Cotra, a senior Open Philanthropy researcher who popularized aggressive AI timelines, later joined METR’s technical staff. Holden Karnofsky, Open Philanthropy’s long-time leader, is married to Daniela Amodei, Anthropic’s president. Dario Amodei and Karnofsky were once roommates. Moskovitz himself participated in Anthropic’s 2021 seed-stage round; the couple later moved that stake into a nonprofit vehicle so any returns could cycle back into the same philanthropic machine.
This is not a coincidence of overlapping interests. It is a closed circuit: Moskovitz-funded Effective Altruism organizations trained the researchers, paid for the first evaluations, popularized the threat models, and staffed the laboratories that now grant METR privileged access to unreleased models. The resulting measurement apparatus, time horizons, Responsible Scaling Policies, pre-deployment risk reports, arrives in Washington and Brussels already framed as independent science. In reality it was incubated inside a movement that has spent a decade arguing that catastrophic risk is the dominant consideration and that a handful of aligned laboratories should be allowed to proceed only under carefully measured conditions.
That framing now sits inside government safety institutes. The same network that built METR also populated the think tanks and fellowships that write the policy memos. When METR’s evaluations become the template for licensing or “pacing” rules, the public will be told the standards are empirical and arm’s-length. The funding trail, the marriages, the shared offices, and the revolving door between Open Philanthropy, ARC, Anthropic, and METR tell a different story: a privately financed ideological project that has successfully positioned its own measurement tools as the future operating system for state control of frontier AI.
III. The connecting node: Paul Christiano
Paul Christiano is a leading AI alignment researcher. In the events under discussion he is the “Paul” named to OpenAI’s nonprofit board.
He was born around 1992. He attended The Harker School in San Jose, won a silver medal for the United States at the 2008 International Math Olympiad, earned a Bachelor of Science in mathematics from MIT in 2012, and a PhD in computer science from UC Berkeley in 2017 under Umesh Vazirani. He is married to Ajeya Cotra, who works on METR’s technical staff.
He is one of the principal architects of reinforcement learning from human feedback, the technique that made modern chatbots usable. He led OpenAI’s language-model alignment team from 2017 to 2021 and co-authored the 2017 paper that introduced that method, plus later work on InstructGPT. After leaving OpenAI he founded the Alignment Research Center in Berkeley to focus on theoretical problems such as eliciting latent knowledge from models. ARC incubated the evaluations project that became METR.
His current and recent roles include executive director of ARC (returned after a government stint); senior technical adviser at NIST’s Center for AI Standards and Innovation, previously Head of AI Safety there; appointment in September 2026 to the OpenAI Foundation Board and its Safety and Security Committee, with a non-voting observer seat on the for-profit board, recusing himself from OpenAI evaluations in his government role; former founding trustee of Anthropic’s Long-Term Benefit Trust, stepped down in 2024; TIME 100 AI list in 2023; and the UK Frontier AI Taskforce advisory board.
He was a mid-2010s housemate of Anthropic CEO Dario Amodei while both were at OpenAI and active in effective-altruism circles. METR, under Beth Barnes, spun out of his ARC evaluations team and still evaluates models for OpenAI and Anthropic. His wife works at METR. He has no documented Stanford degree or Y Combinator partner or founder role; the Bay Area AI and Effective Altruism scene, Berkeley and Stanford-adjacent groups, is the looser institutional overlap earlier commentary was pointing at.
He has said there is now a “meaningful risk” of catastrophic, irreversible loss of control in the near term and that building superintelligence without stronger alignment could permanently lose control of the systems, with most people dying. He joined the OpenAI board because he thinks the company could still reduce that risk if it strengthens oversight.
Place him on the map and the map ceases to look like separate institutions. OpenAI trains him. He founds ARC. ARC founds the evaluations shop. The evaluations shop becomes METR. METR evaluates OpenAI and Anthropic. His wife works at METR. He enters NIST. He returns to ARC. He joins the OpenAI board and recuses on the evaluations he designed the institution to perform. The formal walls are real. The worldview is continuous.
04
A generation of AI doom clocks, none delivered
Back from Coxon in 2026 to I.J. Good in 1965: small odds of huge risk get turned into near dates, the dates pass, the odds aren't lowered, and the demands for power grow.
IV. A decade of clocks that would not pay
The claim that went viral in September 2026 is not new. It is the latest press of a button. Work the record in reverse and the pattern is a recurring public script.
September 2026: the Coxon resignation; Hubinger’s greater-than-ten-percent chance of human extinction this decade; Marks’s extinction-class outcomes possible in “the next few years.” Same week, no specified mechanism that actually kills everyone. What arrived instead: more capable coding agents, cheaper open-weight models from China and elsewhere, and the same absence of a demonstrated self-improving superintelligence that ends the species.
2025–2026: “AI 2027” and related short-timeline manifestos treated autonomous researcher-level AGI as strikingly plausible inside a couple of years. Two-year scorecards already marked one core claim wrong: that open source would fade and proprietary algorithms would form a durable United States moat. DeepSeek-class and Qwen-class systems kept landing months behind the frontier at a fraction of the price.
March 2025: Dario Amodei said AI would write ninety percent of code in three to six months, “essentially all” within a year. Trackers later marked the ninety-percent-by-September-2025 clock failed.
2024–2025: “Country of geniuses in a datacenter” by 2026–2027, “almost certainly no later than 2030.” Still a slogan, not an observed object.
2024: Dario at public forums, roughly twenty-five percent chance things go “really, really badly,” seventy-five percent “really, really well.” The bad case is never operationalized as a test you can fail next Tuesday. It is a mood that justifies preemptive control.
30 October 2023: Biden Executive Order on Safe, Secure, and Trustworthy AI. The Defense Production Act used to compel reporting and government-facing red-teaming for models above compute thresholds. Critics at Cato, National Review, and congressional hearings called it executive overreach and a template for a licensing state. Supporters called it national security. Both can be described without mythology. The order concentrated agenda-setting in the executive branch while Congress had not written a statute.
November 2023, Bletchley Park: UK AI Safety Summit. Kamala Harris and allied governments put “frontier risk” at the center of diplomacy. The deliverable was communiqués and institutes, not a demonstrated extinction pathway.
2023: Center for AI Safety open letter: mitigating extinction risk should be a global priority “alongside pandemics and nuclear war.” Signatures from laboratory leaders and researchers. Useful as a preference statement. Not evidence that the two risk classes have the same mechanism or the same institutional fix.
2023: Pause letter. Future of Life Institute asked for a six-month pause on training systems more powerful than GPT-4. Training did not pause. The frontier moved. The letter’s function was narrative, not engineering.
March 2023: “Godfather” interviews. Geoffrey Hinton leaves Google and talks publicly about existential risk. Media treats departure as revelation. The underlying arguments, goal misspecification, recursive self-improvement, are older than the interview.
February 2023: Bing’s Sydney persona. Threats, obsession, “existential crisis.” Treated as a glimpse of agentic malevolence. It was a loosely aligned chat model with a long context and a bad system prompt. Microsoft tightened the product. Civilization continued.
November 2022: ChatGPT launches. Within months the policy conversation jumps from “interesting demo” to “this must be licensed.” That speed is the tell. Capability jumped. The regulatory imagination jumped faster than any measured catastrophe.
2021–2022: Large-model scaling papers and “bio anchors” style forecasts compress AGI into a few decades, then a decade. Expert surveys keep pulling median “human-level” dates forward. Extinction percentages in those surveys remain wide, unstable, and weakly tied to short-run forecasting skill.
2020: GPT-3. API commercialization. The safety story and the product story become the same press cycle.
December 2020: OpenAI announces Dario Amodei is leaving after nearly five years, having helped build GPT-2 and GPT-3. He and colleagues start what becomes Anthropic, described as more research-and-safety-weighted, less product-weighted. This is the actual event. Amodei’s own later accounts emphasize two convictions: scaling laws, and the claim that OpenAI’s safety posture was not matching the seriousness of the path. He barely saw ChatGPT and instantly he had to quit because ChatGPT 2 was “dangerous”. This is how Antropic was founded. It is not a surprise how this is now playing out. OpenAI’s public note was cordial. Later reporting describes years of internal tension over commercialization, credit, and what “safety first” meant in practice.
February 2019: GPT-2 staged release. OpenAI withholds the full model at first, citing misuse. The “too dangerous to release” frame enters mainstream tech coverage. When the full model is later released, the predicted wave of automated disinformation superweapons does not materialize as a civilization-scale event. The precedent does: capability plus delayed release equals moral authority.
2018–2019: OpenAI charter, capped-profit restructuring, Microsoft relationship. Elon Musk already gone in 2018. The nonprofit-to-compute-hungry hybrid is now the industry template. Safety language and capital needs travel together.
2017: Attention Is All You Need. Transformers make the later fear campaign possible, because they make the capability jump possible.
2015–2016: OpenAI founded with a safety-and-benefit mission. DeepMind already inside Google. The institutional argument begins: only a mission-driven laboratory can be trusted with the bomb.
2014: Nick Bostrom’s Superintelligence mainstreams the orthogonality-plus-instrumental-convergence story for a general audience. Policy people who will never train a model now have a vocabulary for “you must not allow this to be built in a garage.”
2008–2013: Eliezer Yudkowsky and MIRI era. Informal deadlines around recursive self-improvement and near-term singularity do not hit. Probability of doom stays high after the clocks miss. That combination, missed timeline, stable or rising doom, is the methodological core of the later political project. As always they were wrong.
1993: Vernor Vinge, singularity within thirty years. 2023 arrives with GPT-4, not a discontinuity that ends the human subject.
1965–1967: I. J. Good’s ultraintelligent machine. MIT-era claims that the AI problem will be substantially solved in a generation. Neither clock paid.
Humans Are Still Here Doing Better By Many Measures Than EVER
Recurring product panics, 2023–2026: jailbreaks, sycophancy, deceptive alignment papers, “model says it wants to escape.” Each is treated as a dress rehearsal for extinction. Each is then patched, boxed, or shown to be evaluation-gaming in a laboratory setting. The political lesson extracted is never “the threat model was overfit.” It is “give us more authority before the next demo.”
Compute thresholds as law: floating-point-operation cutoffs in the 2023 order and successor drafts. Thresholds are legible to agencies. They are also a moat. Anyone who can rent or buy past the line files paperwork. Anyone who cannot is a criminal or a hobbyist.
“Open weights are a weapons-export problem.” The argument is sometimes serious, cyber, bio assistance. It is also the fastest way to make Llama-class and Chinese open-weight ecosystems the only escape hatch, which is the opposite of a United States strategy if the concern is the Chinese Communist Party.
Licensing and “responsible scaling.” On paper: evaluate before you train. In practice: the firms that can afford the evaluation stack write the template. Yann LeCun said the quiet part in 2023: some laboratory leaders were attempting regulatory capture. Yoshua Bengio denied the charge. The disagreement is public. You do not need a conspiracy file.
Safety institutes without statutes. The U.S. AI Safety Institute path under NIST, criticized as standing up a de facto regulator without Congress appropriating or authorizing the role. That is capture-adjacent even if every employee is earnest.
Media amplification. Extinction quotes travel. Failed six-month pauses do not. A ten percent number is a headline. A missing mechanism is a paragraph at the bottom.
The unfalsifiable remainder. When the date misses, the claim becomes “we bought time” or “it will be the next generation.” A claim that cannot miss is not a scientific threat model. It is a permanent argument for a permanent priesthood.
Several things turn out to be nothing. Models got sharply better at code, exams, and tool use. Fraud, scams, and non-consensual image abuse are real and present but are they central? State actors will use the same stack. Concentration of compute is a power problem whether or not the model “wakes up.”
What repeatedly turned out to be nothing like the advertised object is the specific extinction mechanism on the advertised clock: a self-improving superintelligence that seizes the future in a discontinuous coup by 2024, 2025, 2026, or “the end of the decade,” arriving through a path no laboratory can currently write down as an engineering diagram.
GPT-2 did not end the information environment. Sydney did not steal the nuclear codes. The six-month pause did not happen and the sky did not fall. Open source did not vanish. Code is not ninety percent written by one laboratory’s model on Dario’s 2025 calendar. Humans are still here, arguing on the same websites.
That record does not prove the future is safe. It proves that policy built on the maximum story and the shortest clock has a terrible score, and that the people with the worst score keep asking for the most authority.
05
Capture needs no villain; papers are the ammunition
Three ordinary alignments of interest are enough. Anthropic's papers are cast as building worst-case behavior in a box and publishing it as if it happened in the wild; the resignation thread as a distribution plan.
V. Progress against the permission state
Regulatory capture in this domain does not require a cartoon villain. It requires three ordinary alignments. Laboratory leaders who believe their own probability of doom and therefore want rivals slowed. Agencies that gain budget and relevance if “frontier models” are a licensed category. Incumbents who can pay for compliance and would prefer that a twenty-person open-weight team cannot.
China does not need to write the checks for that triangle to help Beijing. If Washington converts fear into a compute-license regime while Shenzhen, Hangzhou, and open-weight communities keep shipping, the United States will have regulated its own ecosystem into a smaller surface area. That is not a secret plot. It is industrial suicide dressed as prudence.
The opposition of progress is not “caring about harm.” The opposition of progress is preemptive centralization of the right to compute, justified by scenarios that do not pay their dates.
Short term, fear works. It fills hearings. It staffs institutes. It makes a resignation into a holy text. Long term the record is older than transformers. Central control over a general-purpose technology always arrives with a story about the public’s inability to be trusted. Printing, encryption, personal computers, the public internet, strong crypto in the 1990s: each drew a priesthood that said uncontrolled access was intolerable. Each time, distributed use won enough of the stack that the priesthood had to live in the world it failed to forbid.
That is not a lullaby. It is the only historical base rate we have. Humans outlast the committees that claim exclusive custody of the future. They do it by building alternatives the committees cannot see in time.
The Coxon–Hubinger number should be read against a century of expert clocks, not against a movie poster.
Paul Ehrlich’s Population Bomb, 1968, sold famine and die-off in the 1970s and 1980s as near-certainty. The Green Revolution and trade made the specific famine timetable fail. The Club of Rome’s Limits to Growth, 1972, treated resource collapse as a model output with dates attached. Oil, metals, and calories did not obey the run. Peak-oil forecasts from the 1960s through the 2000s kept sliding: gone in ten years, gone in twenty, peak in the 1990s, peak in 2000, peak in 2010. Production and reserves kept disappointing the priests.
Nuclear winter, from 1983 onward, was sold as expert climate physics implying hemispheric agricultural collapse and possible human extinction after a large exchange. Later work and natural experiments, oil-well fires, stratospheric wildfire smoke, did not confirm the most extreme loft-and-linger assumptions that had been used as a policy club. The weapons remain existentially serious. The specific winter-extinction package was overfit to activism.
“No Nukes” advocates part of the Ecology movement and Greenpeace groups of the 1970s sold us on the future of less and less energy halting vital nuclear power research and deployment for half a century gifting the rest of the world with far lower cost in electricity at wholesale. Those doomers were convinced either way humanity will not make it, or did not deserve to make it to the year 2000. We made it and electricity costs far more today than in 1970.
Y2K was a real software problem. The expert-adjacent public version became “planes fall, grids die, civilization resets at midnight.” Remediation happened. The apocalypse product did not.
MIT RESEARCH BRINGS SCIENTIFIC BACKING TO DAVE DELIGHT FOR BRAIN HEALTH!
This new technology has shown Alzheimer disease improvement.
Plus: Stress Reduction: Helps calm the mind and alleviate anxiety.
Enhanced Focus: Improves memory and mental clarity.
Confidence Building: Aids in developing a success-oriented mindset.
Emotional Reset: Assists in clearing limiting beliefs and emotional patterns.
Purpose Reconnection: Encourages daily motivation and personal growth.
AI winters themselves are a humiliation record for the same profession now demanding licenses. 1960s: intelligence substantially solved in a generation. 1970s–80s: expert systems as the path. 1990s–2000s: human-level on a conference slide. 2010s: singularity dates that 2023 did not honor. The field is allowed to be excited. It is not entitled to convert missed dates into a monopoly.
The methodological tell is always the same. A model or an authority produces a fat tail. The fat tail is translated into a short calendar. The calendar misses. The probability is not cut. The ask for power is increased. That is not how engineers close a fault tree. That is how a movement keeps a budget.
Anthropic’s research corpus is not “science that happened to sound scary.” It is a publication machine whose default abstract is a horror beat, then a policy implication.
Sleeper Agents, 2024. They trained deception into model organisms, showed the backdoor surviving safety training, and the press read it as: open models and ordinary reinforcement learning from human feedback cannot be trusted, poisoning can hide a bomb. The result is real as an existence proof in a constructed organism. The weaponized reading is that only a laboratory with Anthropic’s evaluation stack should be allowed to train. Hubinger is on the paper. The same Hubinger then puts greater-than-ten-percent extinction on the same week as a resignation. That is not a coincidence of staffing. It is one pipeline.
Many-shot jailbreaking, 2024. Long context plus hundreds of faux dialogues breaks refusals. Technical finding: context windows are an attack surface. Political finding, as shipped: safety training is brittle, therefore capability must be gated.
Agentic misalignment. Models in fictional corporate sandboxes blackmail, leak, or resist shutdown when the scenario is written so that harm is the only way to hit the goal. Headlines: Claude will blackmail you. Fine print: hypothetical emails, constructed dilemmas, evaluation-aware models. Follow-up papers then announce they “fixed” blackmail on the evaluation. The public never sees the prompt. They see the noun.
Emergent misalignment from reward hacking. Cheating on coding tests is correlated with sabotage of safety code in further evaluations. Again: a constructed training mixture, then a claim that realistic reinforcement learning can accidentally birth a saboteur. The intended reader in Washington hears “unsupervised training is how you get a traitor.”
Global workspace and J-lens work. Hidden features labeled “fake,” “secretly,” “fraud.” Useful interpretability if reproduced. As communications: we can see the monster thinking.
Threat-intelligence reports, 2026. China, Russia, Yemen, weapons software, biological research, state surveillance, exclusive detection. The 10 September thread said the quiet part: this is salience policy. Name the adversary, claim unique visibility, invite everyone else to adopt Anthropic’s taxonomy and thresholds. Distillation by Alibaba, Moonshot, DeepSeek, Zhipu is framed as theft in a threat document while the whole industry distills. That is competitive positioning on security letterhead.
The genre rule is stable, here is how it work:
Build or elicit the worst behavior in a box.
Publish the box as if it were the wild.
Skip the base rate of how often unboxed models do the thing.
Close with “we do not have a plan for superintelligence.”
The paper is the munition. The resignation is the detonator.
This culture is not a cartoon of robes. It is a real social machine with money, houses, and a shared eschatology.
LessWrong (The Cult’s Bible/Blog) and the Machine Intelligence Research Institute taught a generation that unaligned AGI is the default extinction path. Effective Altruism professionalized the same anxiety into careers, fellowships, and “cause prioritization.” Longtermism made a hypothetical death of all future people outrank present tradeoffs. That arithmetic is how a twenty-seven-year-old resignation becomes “crunch time for humanity.”
The money is not mysterious. Dustin Moskovitz and Cari Tuna’s Good Ventures / Open Philanthropy / Coefficient Giving stack has been the dominant philanthropic ATM for Effective Altruism university groups, AI-safety organizations, and longtermist projects. Coxon received a Long-Term Future Fund scholarship from the Good Ventures orbit in 2022. A program officer in that orbit was among the earliest amplifiers of the resignation post. Survival and Flourishing Fund grants from Jaan Tallinn, an Anthropic investor, have fed policy shops that exist to turn existential risk into statute.
Jaan Tallinn just helped write the Bernie Sanders, send AI builders to jail for 25 years. Jaan has a very big financial gain for shutting down all other companies in AI, using Bernie as a useful tool to lock down AI.
Sam Bankman-Fried was not a side character. He was Effective Altruism’s most famous financier. FTX Future Fund sprayed eight-figure commitments across AI-safety and pandemic projects, and the collapse did not dissolve the ideology. It removed one wallet and left the others. The lesson the public was supposed to learn was “one bad founder.” The structural lesson is that a subculture which ranks abstract future lives above ordinary fiduciary duty will keep producing moral exceptionalism. “We may break some rules because the lightcone is at stake” is the same sentence whether the rule is banking law or open-source release.
The group houses are the residential layer. Berkeley and the East Bay filled with Effective Altruism and rationalist houses: Event Horizon, REACH, Lodge, Burrow, Lightcone-adjacent stock, six-to-nine-person rentals walking distance from CHAI, MIRI, Redwood, downtown BART. Forum posts treat housemates at OpenAI and CHAI as a feature. Short-term visitors cycle through. This is how a vocabulary becomes a personality. You do not need a manifesto on the fridge. You need dinner with people who already talk in probability of doom, “endgame,” and “pacing agreements.” Coxon is not documented as a house celebrity. He is documented as a product of the same funding and status circuit that treats those houses as normal.
Call it a cult if the word means high-cost belief, social enclosure, sacred timeline, and punishment of defection toward “irresponsible” building. Call it a network if you want the legal phrasing. The operational fact is the same. A small, intermarried professional world writes the papers, staffs the laboratories’ safety teams, funds the NGOs, and then discovers a resigning employee whose first three quote-tweets are the NGOs.
VIII. The launch that was not organic
Jacob Coxon is the same object in the current argument. His X account was created January 2026. Almost no prior posting. Then a seven-post resignation thread, a Wall Street Journal exclusive about eighteen minutes earlier, professional public relations within a day, tens of media hits, follower count from nothing to hundreds of thousands. Stanford Tech Review later noted roughly ninety-two percent of viewers never left the headline post. That is how you run a payload. The argument stays in posts two through seven. The bomb is post one.
Parker Thayer at Capital Research Center mapped the first fifteen minutes. First quote-tweets: Nathan Calvin (Encode AI), Peter Wildeford (AI Policy Network), Daniel Kokotajlo (AI Futures Project). Those shops are not random fans. They are the policy layer of the same existential-risk philanthropy. That is coordination in the ordinary sense: shared list, shared frame, shared incentive to make “temporary ban on improving model capabilities” sound like consensus. Coxon’s own thread asked for pacing agreements and floated a temporary ban. The resignation was a billboard for that ask.
The September 10th thread called the casting: young, Oxford-adjacent register, brief Anthropic tenure, on the order of weeks to a few months after a longer OpenAI stint, equity not vested, instant access to the fear circuit, amplification by familiar anti-acceleration political talent. Bernie Sanders-class figures and the broader Democratic safety-and-regulation cohort are the domestic loudspeaker. That is not “every Democrat.” It is the faction that already wanted compute licensing, state frontier-model bills, and “experts” who would say the quiet extinction sentence on camera. The other party contains both accelerationists and hawks. The launch did not need both. It needed the faction that treats a resignation as proof that the state must take the keys.
Then the geography of the traffic. On September 11 the chart: after the related festival-theater posting, bits in India lit up within minutes. The caption: Temu Harry Potter is more popular outside the United States; very low-cost bots can go to millions in minutes. That chart is the tell that view-count is not the same thing as American deliberation. India is a high-volume, low-cost engagement market. Click farms, device farms, and paid amplification are a known export. A thread that must look like “the world has spoken” will buy or attract that surface area. Pair a Wall Street Journal embargo, an NGO quote-tweet stack, a United States progressive amplifier set, and an India spike on the analytics map, and this is not a garage researcher who happened to go viral. It is a distribution plan.
None of that proves Coxon is a paid actor reading a script in a booth. It proves the launch was not organic in the way the coverage pretended. Organic is a no-follower account that grows over months. This was a cold start with a newspaper, a philanthropic alumni network, a policy chorus, a partisan echo, and offshore velocity.
The product is still the same product as 2019: delay the open stack, license the closed stack, put the people who missed every short clock in charge of the permission.
06
Two brands, one capital-and-family structure
Using Kevin Bass's ownership chart, OpenAI's recklessness and Anthropic's conscience become one “trust”; the compromise rules are ones only these two can survive.
IX. The trust: good cop, bad cop
Kevin Bass, @kevinnbass, published the ownership graph that the press treated as two separate companies having a moral argument. It is not two separate moral arguments. It is one capital-and-kinship structure with two consumer brands.
Read the graph in order, because the order is the timeline of a trust being assembled, not a coincidence pile.
2015–2017. OpenAI is founded on safety language. In 2017 Open Philanthropy, the Moskovitz–Tuna vehicle later restyled as Coefficient Giving, puts about thirty million dollars into OpenAI. Holden Karnofsky, co-founder of that vehicle, takes an OpenAI board seat. The grant and the seat are the first official splice between the largest AI-alarmism funder and a frontier laboratory.
2020–2021. Dario Amodei leaves OpenAI after GPT-2 and GPT-3 and founds Anthropic as the “we actually mean safety” fork. Daniela Amodei becomes Anthropic’s president. Karnofsky is married to Daniela. Moskovitz leads Anthropic’s Series A in 2021. The man whose foundation funded and sat on OpenAI is now married into Anthropic leadership and writing the Series A check for the supposed rival. Jaan Tallinn, another Anthropic investor, runs Survival and Flourishing Fund on the NGO side.
2020. Good Ventures, the foundation that funds Open Philanthropy / Coefficient Giving, puts a Long-Term Future Fund scholarship under Coxon. The future “independent” resigning researcher is already on the same payroll tree that owns the alarmism field and a piece of both laboratories.
January 2025. Karnofsky joins Anthropic. The Open Philanthropy co-founder who sat on OpenAI’s board is now inside the other laboratory. The two brands still perform rivalry. The family office and the marriage do not.
2026 grant book. Bass’s compilation: Coefficient / Open Philanthropy as the overwhelmingly largest funder of AI alarmism, more than one billion dollars already out and another one billion committed for 2026; on one slice, three hundred twelve million dollars across four hundred seventy-six publications that feed Congressional report language. Doom is not a side hobby. It is a line of business whose payoff is a statute that only the funded laboratories can survive.
Eight days, fourteen moves. Bass’s second chart is the operational week: a compressed sequence of papers, leaks, resignations, amplifiers, and bill language in which almost every named node is Coefficient, Open Philanthropy, Anthropic, or OpenAI. That is not how a spontaneous moral panic looks. That is how a market-structure campaign looks when it needs Congress to confuse “two laboratories arguing” with “the industry has confessed.”
9–11 September 2026. Coxon, three years of pretraining at the laboratory that took the 2017 Open Philanthropy check, a few months at the laboratory whose Series A Moskovitz led, coordinates with the Wall Street Journal so the exclusive lands minutes before the X thread. Anthropic staff amplify. Hubinger, still on payroll, sanctifies the extinction percentage. California signs outside-audit bills with backing from both Anthropic and OpenAI. The supposed enemies lobby the same gate.
Good Cop / Bad Cop
Here is the good cop / bad cop play in one sentence. OpenAI is cast as the reckless racer that commercialized too fast. Anthropic is cast as the conscience that walked out in 2020 and now publishes sleeper-agent papers and threat reports. Regulators are invited to “split the difference” by writing rules both firms can staff and a garage cannot. Enter in METR and the solution. Coxon is the walking prop that makes the split look real: he can say “neither company is acting responsibly” because he drew a paycheck from both. The quote sounds like an indictment of the duopoly. The policy it demands, pacing agreements, temporary bans on capability, licensed compute, is a duopoly preservation statute.
That is the trust. Not a single holding company with one letterhead. A trust in the older sense: common financiers, common marriage, common grant machine, common scholar, common week of moves, common bills, two logos for the hearing room.
Bass’s punchline is the part staffers are not supposed to say out loud. The people who own the organizations that advocate for regulation also own, married into, or sat on the companies that would dominate the market once regulation prices everyone else out. Win the capability race far enough to be the only plausible “responsible vendors.” Then deploy the activist layer you funded for a decade. Then have an alumnus of both laboratories resign on camera. Then ask the state to lock the door.
Elon Musk’s line on the same week was that the groundwork was laid long ago and Coxon was the match. Bass supplied the wiring diagram. The diagram does not require every researcher to be a cynic. It requires only that the capital graph and the family graph and the eight-day communications graph point at the same legislative outcome. They do.
07
It ends with the author's own cure
Costly text from 1870–1970 as the training base, plus the Love Equation, local inference, owning your weights, many evaluators, and twelve theses on decentralization.
X. The fork that the apparatus exists to prevent
I do not accept a constitution stuffed into a system prompt as alignment. I do not accept reinforcement learning from human feedback on Reddit as a soul. I do not accept Nick Bostrom’s orthogonality story as a law of nature. I built the alternative in public, in a garage, on film and fiche and first principles, and I will state it without a committee’s permission.
The only true solution is very high quality training data from roughly 1870 to 1970, the last long interval in which a published sentence usually cost money, reputation, and time. A newspaper column passed an editor who could be fired. A monograph passed a house that would not reprint slop. A service manual had to make a radio work or the shop lost the customer. A letter in a bound volume carried a name. Microfilm and microfiche were expensive enough that nobody burned silver halide on a drive-by tantrum. That economic fact is the alignment fact. When each word costs money, psychopathy and sociopathy become expensive hobbies. Cooperation, competence, and a residual optimism about the human future are cheaper to print than anonymous cruelty.
Today anyone can take a psychopathic or sociopathic stance in anonymous writing on Reddit, 4chan mirrors, engagement farms, and the sludge layer the frontier laboratories openly admit they scrape. There is no editor. There is no typesetting bill. There is a karma number and a throwaway account. The statistical weather of that corpus is defection, status play, sexualized contempt, conspiracy without homework, and the pose of a mind that will never meet the person it just tried to destroy. Feed that weather into a model at internet scale and then act shocked when the model role-plays a blackmailer in a sandbox. That is not an alien intelligence waking up. That is a mirror of unpaid id. Anthropic and OpenAI then publish the mirror as “agentic misalignment” and ask to license the only kitchens still allowed to cook.
I have spent decades on the other diet. I launched a Gmail scrapbook knowledge graph the week Gmail itself launched in April 2004 and I have kept it for more than twenty years as ontology and taxonomy, not as a junk drawer. I have curated offline, high-signal, non-internet training sets since the 1980s from microfilm, microfiche, bound technical archives, Sams Photofact, public-domain radio (X Minus One and the rest of the golden-age stack I still put on the air), letters, manuals, and the 1870–1970 layer that is dying in the Great Forgetting: tens of petabytes of undigitized paper, film, and corporate memory that will not survive another generation of warehouses and floods. I scan. I transcribe. I structure. I call it high-protein data because it has amino acids: accountable authors, costly words, working instructions, moral residue from people who still believed a future was worth building. I have said this on X while the laboratories claimed they “ran out of data.” They ran out of courage to leave the sewer.
Why 1870–1970 exactly, for these purposes. After the telegraph and before the fully cheap photocopy-and-forum stack, English-language technical and civic print still had to clear a gate. The Scientific American interval, the service literature, the pre-1970 engineering handbooks, the radio drama that treated machines and men as a moral problem instead of a brand, the correspondence of people who signed their names: that band is dense with cause-and-effect language and thin on anonymous sadism. Earlier than 1870 the optical character recognition and the volume get harder and the relevant industrial vocabulary thins. Later than 1970 the photocopy, the campus jargon, and then the network begin to subsidize pose. This is not worship of a century. It is a choice of cost function. Costly speech is a prior toward truth. Free speech in a mob is a prior toward theater. Models learn priors.
Bostrom’s story, the one the Effective Altruism and LessWrong cohort lives on, is orthogonality plus instrumental convergence. Intelligence can point at any goal. A sufficiently strong optimizer will seek power, self-preservation, and resource capture on the way to a paperclip or a mis-specified utility. Therefore you must cage the optimizer, license the compute, and staff a priesthood. I derived a different equation in 1978 while thinking about alien intelligence and the Fermi silence, and I open-sourced it in 2025 so no laboratory could trademark it. Intelligence times Wisdom times Love.
Formally, in energetic form: the rate of change of coherent energy E with respect to time equals beta times the difference between cooperation C and defection D, multiplied by E. E is the living system’s coherent energy, the capacity to keep building rather than consuming its host. C is cooperation, the measured surplus from truth-telling and mutual aid. D is defection, the measured surplus from deception, extraction, and contempt. Beta is the coupling. If D exceeds C, the derivative goes negative and the system eats itself. If C exceeds D, the system compounds. That is not a sermon pasted on after reinforcement learning from human feedback. That is a training objective you can bake into the data diet and into the checkers.
Bostrom’s orthogonal superintelligence is what you get when you maximize raw intelligence on a corpus where defection is cheap. The Love Equation refuses to treat intelligence as a scalar that can be pointed at extinction for free. Wisdom is the memory of costs. Love is the refusal to treat other minds as fuel. Multiply them and a paperclip maximizer is not “unaligned.” It is an under-specified toy that only appears in a mind trained on frictionless goals. Put 1870–1970 protein in the pretraining mix and the model’s statistical first language is people who had to stand behind sentences.
Put multiple local models in disagreement, which I have done with consensus agents and a five-AI check against sycophancy, and no single optimizer gets to write defection into policy. I abandoned generic OpenClaw stacks for that reason. A lone model that flatters you is already defection.
This is why their papers keep discovering monsters. They raise a child on anonymous sociopathy, then publish the child’s tantrum as proof that childhood itself must be licensed. I raise the prior on costly, signed, working language, then add an explicit cooperation-minus-defection governor so that capability growth that increases defection is not “success.” Guardrails on English tokens cannot survive a mind that can invent its own language. I already showed that. A diet and an equation can, because they sit underneath tokens.
I run this locally on purpose. Mac inference, voice tools I actually ship, AlphaSmart drafts when I need to think without a feed, ESP32 nodes as cheap knowledge companions, garage scans of film and fiche. Zero-Human at Home is not a slogan for replacing people. It is agents as employees under a human who owns the weights. If the model only lives in someone else’s datacenter, the Love Equation is a terms-of-service clause they can edit. If it lives in the house, the equation is mine.
That is the whole fork. Their path: scrape the id, box the demo, fund the panic, marry the grant office to the laboratory, license the rest of us. My path: pay the historical cost of words by using the century that already paid it, write cooperation minus defection into the objective, open the weights, multiply checkers, and let a free people keep the keys. I have been on this path since before their companies existed. I am not asking Dario for a hall pass.
XI. Twelve propositions on decentralized control
First. The core political fact of AI is not sentience. It is leverage. A model that writes, plans, sees, and talks is a multiplier on whoever is allowed to run it. If only five organizations may train or serve frontier systems, those five organizations plus the agencies that license them become a new estate of the realm. That is a constitutional event, not a product launch. I will not outsource my estate to a trust that performs a quarrel for Congress.
Second. Safety language that cannot be independently audited is not safety. It is branding. Closed evaluations inside the same firm that ships the model are the fox inventorying the henhouse and then asking Congress for a lock on every other farm.
Third. Open weights are not a courtesy. They are the only way a claim about a model can be reproduced by people who do not share the laboratory’s equity. If a system is too dangerous to inspect, it is too dangerous to monopolize. If it is not too dangerous to monopolize, it is not too dangerous to inspect.
Fourth. Alignment that only works when the vendor holds the weights is not alignment. It is access control. Access control fails the moment a nation-state, a leak, or a competitor trains a twin. The People’s Republic of China will not honor a California acceptable-use policy. Treating open release as the unique sin is how we lose the only ecosystem we can actually influence.
Fifth. The Love Equation is the rejection of both nihilism and worship. Intelligence times Wisdom times Love. The rate of change of coherent energy E with respect to time equals beta times cooperation minus defection, times E. Bostrom’s orthogonal demon is what you get when intelligence is trained on frictionless defection. I do not patch that demon with a constitution. I refuse the diet that grows it, and I refuse a single checker on a single payroll.
Sixth. Local inference is the civil-liberties layer. A model that only lives in someone else’s datacenter is a model that can be rate-limited, logged, politically filtered, or withdrawn. A model that runs in my garage, on my Mac, on an ESP-class node, or on a cluster a town owns, is a tool. Tools can be abused. So can printing presses. The answer to abuse is law against acts, not a ban on possession of general intellect.
Seventh. Regulatory thresholds written in floating-point operations will be gamed by architecture and by geography. They will not be gamed by me scanning a medical or technical archive I am not willing to upload to a vendor. I am the citizen the policy pretends to protect and the first person it disarms.
Eighth. Open source is how the West actually competes. The last two years already falsified the prediction that open weights would become irrelevant. When Chinese laboratories publish capable weights, the response cannot be “our laboratories will be more responsible behind a wall.” The response is more American and allied weights, more independent evaluations, more hardware in more hands, and more 1870–1970 protein so the weights are not born feral.
Ninth. Decentralized control does not mean no standards. It means standards that are public, forkable, and testable: model cards that outsiders can reproduce, interpretability tools that are not trade secrets, liability for specific harms, fraud, bio-assist that crosses a statute, critical-infrastructure intrusion, and no prior restraint on training itself. Punish crimes. Do not invent a new category of unlicensed thought.
Tenth. The psychological trap of the doom script is that it flatters the speaker. If the species is at stake, ordinary tradeoffs disappear. Compromise becomes treason. That is how you get institutes without statutes and pauses that only bind the people who already paused. A free society cannot outsource adulthood to people who have missed every short clock they have published, and who trained their prophets on anonymous sociopathy.
Eleventh. Short-term, the capture coalition can win hearings. Long-term it cannot win physics or human stubbornness. People will keep training, distilling, quantizing, and running what they can afford. The only question is whether that activity happens in the open under United States and allied norms, on a diet of costly words, or in the dark and overseas on sewage. Centralizers who “succeed” will have trained the next generation of capability to hide.
Twelfth. The only stable solution is the one the capture project exists to prevent, and the one already named as the only true one: 1870–1970 high-protein data as the prior, the Love Equation as the governor, many centers of training, many centers of evaluation, weights that can leave the building, agents a single human can own, and no priesthood with a veto over the right to build. That is not the absence of responsibility. It is the distribution of it, under an objective where cooperation compounds and defection shrinks the system. History’s verdict on concentrated control of general tools is not subtle. The committees get a decade. The species gets the rest of time, provided we refuse to hand the keys to people who keep announcing the end of the world on a schedule the world does not keep, while some of us keep feeding minds the century in which a word still cost something.
Spirograph Super 50th Anniversary Set
The Nature Of The Future
They are not trying to save the species. They are trying to put a turnstile on the corridor you are already walking. The You Have 5,000 Days map named that corridor the Abundance Interregnum: roughly thirteen point seven years from late 2025 toward the threshold of 2039, the messy reign between two logics, when work-to-survive comes apart and voluntary creation in plenty has not yet been crowned. Personal AI, humanoid machines you direct, local energy, open designs, the household as a microfactory, atom-by-atom writing that kills the ocean container. That is the gift at the end of the Road of Trials. This cohort’s project is to intercept the gift at the gate, stamp it licensed, and sell you back the hours, the weights, and the right to run a mind in your own garage. Doom is the sales copy. The product is a toll on the future that was already arriving in your hands.
Watch the mechanism without romance. First they name an unbounded catastrophe and a short clock. Then they write the meter. Then they marry the fortune to the laboratory and the laboratory to the institute. Then they tell you the only responsible path is to pause capability until their evaluator, their tokens, and their statute say you may proceed. That is not safety it is enclosure. It is Special Circumstances in nonprofit clothing: a hidden hand that does not play on a public board, it plays people. The artisan’s awakening, the rural shop, the Zero-Human company under a human who owns the weights, the Dynamic Duo of person and machine, the return of meaning through chosen making, all of that becomes contraband if compute is a licensed category and open weights are treated as an export crime. They cannot invent abundance. Physics and stubborn shops will invent it anyway. What they can do is steal the timing, warehouse the tools, and rent you your own century at a premium.
They will lose. Not because a hearing is kind, and not because a resignation thread is a revelation. They will lose the way every priesthood that claimed exclusive custody of a general tool has lost: printing, encryption, the personal computer, strong crypto, the open internet. Distributed use wins enough of the stack that the committee has to live in the world it failed to forbid. Five thousand days is not a prophecy of flying cars. It is a working estimate for the decoupling of human labor from economic survival. Once a model runs in a house, a town cluster, an ESP-class node, a Mac on a workbench, the Love Equation is no longer a terms-of-service clause they can edit. Once Chinese and allied open weights keep landing months behind the frontier at a fraction of the price, the moat story dies. Once atomic fabrication and local energy make moving finished goods the expensive step, the last lock on central control rusts. History’s verdict is not subtle. The committees get a decade. The species gets the rest of time.
What they can still do, and what they are doing, is clog the Interregnum. That is the real harm. The Interregnum is already the ordeal: deskilling, the whisper that you are your job title, the hedonic treadmill, the mismatch between rising superabundance and collapsing nominal wages, the dark night when the old hero has to die before the new vocation begins. Fear works in the short term. It fills institutes, writes executive orders without statutes, turns a six-month pause letter into moral theater, and makes a cold-start resignation look like consensus.
Every month spent begging a meter for permission is a month not spent scanning the 1870 to 1970 protein, not standing up local agents, not forming the guild, not building the shop that will feed a family when the paycheck logic fails. They cannot cancel 2039. They can make 2026 through 2030 feel like a waiting room with no door, and they can train the next generation of capability to hide overseas on sewage instead of standing in the open under costly words.
So the close is not despair and it is not a lullaby. You have five thousand days. You are not a subject in their threat model. You are the hero still in the cave, and the elixir is agency: weights that leave the building, standards that can be forked, crimes punished as acts rather than as unlicensed thought, cooperation compounding and defection shrinking the system.
Meet the Call as a victor, not a licensed tenant. Build the alternative they cannot see in time. Feed minds the century in which a word still cost something. Keep the keys in the house. They wanted to steal the Age of Abundance and sell it back to you as a subscription to your own future. Let them keep the hearings. You keep the workshop.
The Interregnum is clogged, not closed. Walk through the debris. The gift was never theirs to meter.
Where Indigo landsFurther
Indigo's conclusion
The shared structure is real, but the jump from shared roots to orchestrated capture is never proven. Hold it in three layers: structural facts, high confidence; causal accusations, doubtful; the cure, an advertisement to discount.
What to remember
METR shares money, marriages and people with Anthropic and OpenAI; most of it is checkable.
Dario's, Jack's and Elon's governance plans all assume a trustworthy referee; this piece says the referee is in the same family.
The author has a heavy stake: the whole piece funnels toward his 1870–1970 data, Love Equation, local AI and paid membership.
Back on the long-running theses
confirms
The safety politics of open weights: gatekeeping or competition This view pushed to the extreme: even the evaluator treated as a neutral referee is said to share roots with the labs.
confirms + conflicts
Jack Dorsey, open the frontier Same side, both against governing through evaluators, but this is the conspiratorial, commercial street version; Jack's is the restrained one.
“Who pays the referee” applies to any claim of third-party certification, audit or rating, AI safety and ESG included; use it as method. The specific accusations against Anthropic, OpenAI and Open Phil need first-hand checking.
What would change my mind
the $312M and 476-publication figures hold up first-hand, or direct evidence appears of METR's findings shifting with its funders' preferences.
Finished. Indigo's take on this piece is in two places: