Politics Gets Interested In Those Trying Not To Die
打开互动全文版（逐段中英对照 + 图/公式 + 论文问答）→本文考察围绕 AI 发展的政治与监管格局，聚焦特朗普政府在 AI 安全上的矛盾立场。文章指出，白宫一方面公开将 AI 生存风险斥为“骗局”，另一方面却强制实施安全审查并采取临时干预措施，例如强制模型下线。核心矛盾在于政府拒绝解决集体行动问题：它阻止实验室在安全上协调，同时威胁反垄断行动，实际上阻止了为避免构建危险超级智能而进行的自愿合作。作者得出结论：政府正在积极阻碍防止灾难所需的那种协调，将商业利益和政治作秀置于有意义的保障之上。局势严峻但并非无望，因为一些官员承认需要理性的应对，尽管进展仍远远太慢。
This article examines the political and regulatory landscape surrounding AI development, focusing on the Trump administration's contradictory stance toward AI safety. It argues that while the White House publicly dismisses existential AI risk as a 'hoax,' it simultaneously imposes mandatory safety reviews and ad hoc interventions, such as forcing models offline. The core tension lies in the government's refusal to address collective action problems: it blocks labs from coordinating on safety while threatening antitrust action, effectively preventing voluntary cooperation to avoid building dangerous superintelligence. The author concludes that the government is actively obstructing the very coordination needed to prevent catastrophe, prioritizing business interests and political posturing over meaningful safeguards. The situation is dire but not hopeless, as some officials acknowledge the need for sane responses, though progress remains far too slow.
6. JD Vance、Michael Kratsios 与集体行动问题。
6. JD Vance, Michael Kratsios and Collective Action Problems.
8. 建议不要死会让你因反垄断被起诉。
8. Suggesting Not Dying Gets You Sued For Antitrust.
9. 其他政府官员说了些理智的话。
9. Other Government Officials Say Sane Things.
17. 《纽约邮报》超越自我，但公然违反规则。
17. The New York Post Tops Itself But Outright Breaks The Rules.
19. 如果模型按指令行事并导致你死亡，那就不妙了。
19. If The Model Is Acting As Instructed And It Kills You That Is Not Fine.
20. 由 AI 撰写的《华尔街日报》专栏文章对 HuggingFace 撒谎。
20. AI-Written Wall Street Journal Op-Ed Lies About HuggingFace.
22. 我显然无法选择我面前的酒。
22. I Clearly Cannot Choose The Wine In Front of Me.
如果美国人明白，特朗普不仅把 AI 带来的生存风险，还把每一桩涉及数据中心的投诉，都称作他的“骗局”，那么我预计他在此处的支持率，乃至总体支持率，都会进一步恶化。
If Americans understood that Trump was calling not only existential risk from AI but also every complaint involving data centers one of his 'hoaxes,' I would expect his approval rating here, and overall, to deteriorate rather further.
如果把我写上面那句话时还不知道的事情也算进去，我会把这一点推得更远，因为当时才只是周五。生活来得太快。
I would extend this even further if you include the things I did not know about when I wrote the above sentence, because at the time it was only Friday. Life comes at you fast.
总之，这是上周公布的一项民调中的情况。他的总体支持率一并列出，以供参考：
Anyway, this was the situation in a poll published last week. His overall approval is included for context:
特朗普认为问题不在于他的立场说得不够清楚或不够响亮，他打算设法解决这一点。你知道六周后有一场选举吗？
Trump thinks the problem is not saying his positions clearly or loudly enough, and he is going to try and fix that. Did you know there is an election in six weeks?
9 月 19 日上午 10 点 12 分左右，他抛出了这个新作品，只是修改了段落分隔，因为特朗普认为回车键是个骗局。
On September 19, around 10:12am, he dropped this new creation, only modified to include paragraph breaks since Trump thinks the enter key is a HOAX.
特朗普在把事物称为失败以导致其失败这一策略上已经走了很远。在任何情况下，‘疯狂的数据中心攻击’都没有大体上失败，如果它失败了，我们就不会以这种方式在这里谈论它。
Trump has come a long way with the tactic of calling things failures in order to cause them to fail. There is no world in which the ‘crazed Data Center attack’ has largely failed, and if it had we wouldn’t be here talking about it this way.
也就是说，没有新的规则或法规，除了我已经添加的那些，以及我以后心血来潮时临时添加的那些，因为我觉得某些东西是坏的。而所谓坏的，我指的是对生意不利的。
As in, no new rules or regulations, except the ones I already added, and the ones I ad hoc add later when I feel like it because I think something is bad. And by bad I mean bad for business.
嘿，我喜欢把自己看作一个高智商的人。
Hey, I like to think of myself as a High I.Q. individual.
大卫·萨克斯是第一位 AI 沙皇，所以要做得更差很难，但并非不可能。
David Sacks was the first AI Czar, so it would be hard but not impossible to do worse.
更有趣的想法是 AI Force（人工智能部队）。就目前而言，太空部队确实有可能算是成功的。拥有 AI Force 意味着什么？谁知道呢。可能是任何东西。与太空部队的类比在后勤上其实并不合理，设立一个军种分支也同样不合理。我能想到一些被称为 AI Force 的版本，它们总体上有用且合理。我也能想到一些两者皆非的版本。
The more interesting idea is the AI Force. It actually is plausible that the Space Force is a success, as far as it goes. What would it mean to have an AI Force? Who knows. Could be anything. The analogy to Space Force does not actually make logistical sense, nor does having a branch of the military. I can think of versions called AI Force that are net useful and sane. I can think of versions that are neither.
他已经完成了四分之一。AI 可能占到我们国家 GDP 的 100%。
He’s a quarter of the way there. AI could be as much as 100% of our Country’s GDP.
特朗普还决定……给 AI 改名？这人确实有起名字的天赋。
Trump also decided to… rename AI? The man does have a knack for names.
三个名字都挺搞笑，其中 Superior Intelligence 最好。如果你试着把它们放进句子里，显然都说不通，但无所谓。
Of the three, all of which are funny, Superior Intelligence is the best. None of them make sense for obvious reasons if you try saying them in a sentence, but that's okay.
然后他觉得 'supreme' 不好，因为会联想到 'supreme court'（最高法院），于是重新投票：
Then he decided 'supreme' is bad because of the 'supreme court' and reset the vote:
看到这种随意的、彻头彻尾的假话，感觉很奇怪，因为他说 Supreme Intelligence 正 '惨败' 给 Extreme 时，它实际上明显领先，但管他呢，我们现在还注意这种事吗，什么都不重要了，等等。
It's weird to see such random outright falsehoods, since Supreme Intelligence was actually outright ahead of Extreme at the time he said it was 'losing badly' to it, but whatever, do we even notice such things anymore, nothing matters, etc.
我不确定那是在他宣布新的凯旋门计划（该建筑将兼作军事综合体）之前还是之后。
I'm not sure if that was before or after he announced the new plan for a Triumphal Arch that will double as a military complex.
特朗普从不相信民调，也不愿冒险，他选择了一条不同的路。
Never one to believe the polls, or not tempt fate, Trump went a different way.
这种“哦，我们会用现有刑法来应对 AI 威胁”的说法简直是荒诞主义。我推测，他所说的“超级智能”并非我们这里讨论的超级智能，而只是指 AI 很超级，就像扎克伯格说超级智能意味着他的新 AI 眼镜一样。
This bit of ‘oh we will use the existing criminal law to deal with AI threats’ is absurdist. I would presume that by ‘SUPER INTELLIGENCE’ he does not mean superintelligence as we talk about it here, and instead simply means AI is super, the same way Zuckerberg says superintelligence means his new AI glasses.
我怀疑特朗普是否知道这个名字已被“占用”，或者这会带来怎样的混乱。他可能知道，并且有意在多个层面上这样做。也有可能他完全不知道。
I wonder if Trump even knows that this name was ‘taken’ or the confusions this will bring. It’s possible he knew and did it on purpose on several levels. It is also possible he has no idea.
《纽约时报》证实，黑暗三角是 David Sacks、Jensen Huang 和 Mark Zuckerberg，警告他“输给中国”。这有点太直白了，但本季我们的作者截稿期限越来越紧，所以我表示同情。
The New York Times confirms that the dark triad is David Sacks, Jensen Huang and Mark Zuckerberg, warning him about ‘losing to China.’ It’s a little on the nose, but our writers this season are on increasingly tight deadlines so I sympathize.
我确实理解这种担忧，尤其是在数据中心方面。也许事情很简单：特朗普不明白，当 Dario Amodei 和 Sam Altman 说“放慢 AI”时，他们的意思是“不要进入完全递归自我改进并提速十倍”。
I do get the concern, especially on the data center side. Maybe this is as simple as Trump not understanding that when Dario Amodei and Sam Altman say ‘slow down AI’ they mean ‘do not go into full recursive self-improvement and get ten times faster.’
一如既往，保持克制。特朗普是独特的，而共和党人除此之外正在转变态度。在白宫内部，许多人正试图理智地回应。
As always, hold your fire. Trump is unique and Republicans are otherwise coming around. Inside the White House, many are attempting to respond sanely.
情况好吗？哦，当然不。情况很糟糕。
Is the situation great? Oh, hell no. The situation is terrible.
但情况并不像听起来那么糟糕。目标已经移动了不少。
But it is not as terrible as it sounds. The goalposts have moved quite a bit.
想想看，那些为中美峰会做准备的人仍然希望取得进展：
Consider that those getting ready for the US-China summit still hope for progress:
这不是我们所希望的结果，但这是在向前推进——尽管速度太慢——而不是倒退。
This is not what we would have hoped for, but this is moving forward, albeit way too slowly, rather than backward.
特朗普实际上已经自行其是，对新的、美国生成的模型的安全性实施了强制性审查。
Trump literally already did go his own way and impose mandatory reviews of the safety of new, American-generated models.
他对新模型发布施加了‘自愿’审查框架。没有人认为这些审查是自愿的。实验室不这么认为，法律专家不这么认为，市场不这么认为，白宫也不这么认为。
He imposed his ‘voluntary’ review framework for new model releases. No one thinks those reviews are voluntary. Not the labs, not the legal experts, not the market, and not the White House.
如果你想知道这一切有多么不自愿，我们还有报道称，当特朗普因所谓的越狱演示而要求将 Fable 5 下线时，特朗普说如果他们拒绝遵守，‘我想把他们送进监狱’。说实话，这很合理，不是什么爆炸性新闻。相关 Politico 人物报道的其余部分也印证了我此前对局势的理解。
If you want to know how not voluntary all of this is, we also have reports that when Trump was demanding Fable 5 be taken offline due to a supposed jailbreak demo, Trump said ‘I want to send them to jail’ if they refused to comply. Honestly, fair, and not a bombshell or anything like that. The rest of the associated Politico profile also echoed my previous understanding of the situation.
也就是说，总统既把自己推行的政策称为“骗局”，又表示他想要一个完全临时的体制——由他一人逐案决定规则是什么，而不是一个法治国家。毕竟他是如此一位“高智商人士”。
As in, the President is both calling a 'hoax' on his own policies, and also saying he wants a fully ad hoc system where he alone decides what the rules are, on a case-by-case basis, rather than having a nation of laws. Since he is such a 'high IQ individual.'
David Sacks 加上他的 AI 试图吹嘘 AI 为我们所做的一切，想出了四件事：
David Sacks plus his AI try to brag about all that AI has done for us, come up with four things:
Astra 从第一性原理出发，估计最多有 21 万个建筑工作岗位。
Astra estimates at most 210,000 construction jobs from first principles.
Fable 将百万岗位的说法追溯到 9 月 5 日的《经济学人》，其中约有 32 万个可能与建设相关的工作岗位，然后他们又加上了自 2022 年以来工程师、软件开发人员、数学家（？！）和数据科学家的 73 万个“高于趋势”的工作岗位。
Fable tracks the million claim down to The Economist from September 5, where there are 320,000 plausibly buildout-related jobs and then they pile on 730,000 ‘above-trend’ jobs for engineers, software developers, mathematicians (?!) and data scientists since 2022.
那不是建设，而是趋势选择，而且它始于 2022 年，而不是特朗普任期开始的 2025 年。
That’s not the buildout, it is trend selection, and it starts in 2022 not 2025 when Trump’s term started.
相关性不等于因果关系，但 Sacks 在这里只是声称方向性影响。我实际上确实认为这是相对于没有 AI 的因果关系，否则市场会下跌，但 Sacks 并不处于指出这一点的政治位置。
Correlation is not causation, but Sacks is only claiming directional impact here. I actually do think this is causation versus no AI, and that the market would be down otherwise, but Sacks is not in a political position to point this out.
我希望这是第三项而不是第四项，因为“纵火、谋杀和乱穿马路”本应恰好有三项，但没错，那确实发生了。
I wish this was the third item rather than the fourth, because Arson, Murder and Jaywalking is supposed to have exactly three items, but yes, that did happen.
他最后说特朗普拒绝“暂停”，并谈到桑德斯和沃伦。这与拒绝安全措施明显不同，因为萨克斯在这里谈论的 90% 以上是数据中心。那才是他们真正关心的战斗。
He finishes by saying Trump is refusing a 'pause' and talking about Sanders and Warren. This is noticeably different from rejecting safety measures, as what Sacks is talking about here is 90%+ data centers. That's the battle they actually care about.
JD Vance 说，如果你正在建造‘弗兰肯斯坦’，那么你应该向内自省并停止，而不是要求政府进行监管。
JD Vance says, if you are building ‘Frankenstein,’ then you should look inward and stop, not ask the government for regulation.
我同意前半部分。如果他们正在建造这样的东西，他们就应该停止。
I agree with the first half. If they are building such a thing, they should stop.
假设你计划在里诺枪杀一个人，只是为了看他死去，但不知何故，目前在里诺枪杀人是合法的，前提是你看着他们死去。你应该向内自省并停止。此外，政府应该阻止你。如果有人说‘我的直播竞赛迫使我枪杀里诺的人，以便我们看他死去’，你应该说‘向内自省并停止’，但不要接着说‘不要告诉我将枪杀里诺的人定为非法。’
Suppose you plan to shoot a man in Reno, just to watch him die, but somehow it is currently legal to shoot men in Reno, provided you watch them die. You should look inward and stop. Also, the government should stop you. And if someone says ‘my livestreaming competition is forcing me to shoot a man in Reno so we can watch him die’ you should say ‘look inward and stop’ but not then also say ‘don’t tell me to make it illegal to shoot a man in Reno.’
我确实赞赏 Vance 明确承认这些担忧是真诚的而非愤世嫉俗的，并且这不是关于监管俘获。
I do appreciate that Vance explicitly acknowledges that the concerns are earnest rather than cynical, and that this is not about regulatory capture.
我也非常赞赏他隐含地表示，你应该愿意‘输给中国’。即使我们没有与其他方达成不建造该物的协议，也不要建造它。
I also very much appreciate that he’s implicitly saying that you should be willing to ‘lose to China.’ Just don’t build the thing even if we don’t have an agreement with others not to build it.
政府以所谓无害的“越狱”为由强行将 Fable 下架，迫使 Anthropic 增加额外的防护措施，白宫还亲自挑选谁可以、谁不可以访问 Mythos，包括拒绝许多 Anthropic 希望授予访问权限的公司，然后反过来抱怨 Anthropic 拒绝这些公司访问同样的模型，这真是够讽刺的。
It is pretty rich for the government to forcibly take Fable off the market due to concerns over a harmless so-called 'jailbreak,' force Anthropic to add additional guardrails, and for the White House to handpick exactly who is and is not allowed access to Mythos, including refusing access to many companies Anthropic wants to grant access to, and then turn around and complain that Anthropic is denying companies access to those same models.
如果你想让 Anthropic 为拒绝防御者访问承担责任，那就别再告诉他们授予防御者访问权限是非法的。那可能会有所帮助。
If you want it to be Anthropic's responsibility that they're denying defenders access, then stop telling them it would be illegal to grant defenders access. That might help.
好吧，所以这些公司应该停止，但他们被告知在法律上不能相互交谈以同意这样做，Hawley 称其为奖励，而在特朗普的领导下，FTC 表示这一请求听起来可疑。
Okay, so the companies should stop, except they are being told legally they cannot talk to each other to agree to do this, Hawley calls it a reward and under Trump's leadership the FTC has said the request sounds suspicious.
因此，答案是：如果没有这样的豁免，两家公司将在一定程度上单方面行动，并在不承担不可接受的法律风险的情况下事实上进行软协调，但你不能在竞争对手推进时无限期地退缩，否则你会输，然后竞争对手无论如何都会让我们所有人陷入危险，而且可能是更严重的危险。
The answer is thus that, without such a waiver, both companies will act unilaterally up to a point, and will de facto soft coordinate to the extent they are not taking on unacceptable legal risk, but you cannot continue to indefinitely hold back while competitors press ahead, or you lose, and then the competitor puts us all in danger anyway, and probably worse danger.
为什么像 Vance 和 Kratsios 这样的人声称对集体行动问题、公地悲剧或猎鹿问题感到困惑？为什么他们认为答案就是变成一个合作机器人？他们真的这么迟钝吗？
Why are people like Vance and Kratsios claiming to be confused by a collective action problem, a tragedy of the commons or a stag hunt? Why do they think the answer is to just turn into a cooperate-bot? Are they really this dense?
我猜他们不至于这么迟钝，尽管 Vance 在其他场合曾称经济学是假的，所以也说不准。我只能假定，这一切装作不懂外部性、集体行动问题、公地悲剧等等的姿态，都是表演性的困惑。
I presume they are not this dense, although Vance has in other contexts called economics fake, so you never know. I have to assume that all this pretending not to know about externalities and collective action problems and tragedies of the commons and so on is performative confusion.
正是如此。这是最入门级的东西了。
Exactly. This is the most 101 thing out there.
政府的职责在于保护人民、提供共同防御，同时还要解决市场失灵与协调问题。你们还想继续当政府吗？
The point of a government is to protect the people, provide for the common defense, and also to solve market failures and coordination problems. Do you want to keep being the government?
政府目前不仅拒绝履行政府职责，还积极阻挠不建造可能杀死所有人的超级智能。
The government is currently not only refusing to be a government, it is actively getting in the way of not building superintelligence that might kill everyone.
比如阻止实验室交流或合作，警告反垄断，而不是给予在这种情况下使用的非常标准的安全例外。
As in stopping the labs from talking or cooperating, warning about antitrust, rather than granting a very standard safety exception that is used in such situations.
你有没有考虑过，也许对“世界末日即将来临”的正确回应不是担心你在“奖励”谁，而是——疯狂的建议——防止世界末日？
Have you considered that maybe the correct response to 'the end of the world is near' is not to worry about who you are 'rewarding' but rather to, crazy suggestion, prevent the end of the world?
我的意思是，我们完全应该让他们对机器人所做的任何坏事负责，但 AI CEO 们指出这并不能防止世界末日。所以也许除了这个之外再做一点。如果你想确保这也惩罚他们，因为是他们造成了这整个烂摊子？是的，好吧，我们可以支持。
I mean, we should totally make them liable for whatever bad stuff their bots do, but the AI CEOs are pointing out that will not prevent the end of the world. So maybe do a little bit more than that. If you want to make sure it also punishes them, since they caused this whole mess? Yeah, okay, we can be down for that.
反垄断豁免是至少能做的，如果合理构建，在这种情况下并不是“奖励”。这样想是零和对抗思维，尤其是执着于此。实验室所提议的显然对商业不利，正如股价变动所反映的那样。
The antitrust exemption is the least one could do, and is not a 'reward' in this context if you structure it reasonably. It is Zero-Sum Adversarial Brain to think of it that way, and especially to be fixated on that. What the labs are proposing would very obviously be bad for business, as reflected by the movements in stock prices.
因为我们生活在最愚蠢的时间线里，一家律师事务所冲到法院提起反垄断诉讼，理由是发推文说“Dario 是对的”违反了《谢尔曼法》中关于限制贸易的规定。常规操作。
Because we live in the dumbest timeline, a law firm rushed to the courthouse to file an antitrust lawsuit, on the basis that Tweeting “Dario is right” is a Sherman Act violation in restraint of trade. Standard stuff.
他们代表四名订阅者起诉，其中三名是律师，大概很大程度上是为了在任何潜在诉讼中获得优先权，毕竟你永远不知道会发生什么，也许还能获得证据开示和一些媒体报道。我的意思是，没有损害，没有他们抱怨的行为，也没有协议。
They are suing on behalf of four subscribers, three of whom are lawyers, presumably in large part to get priority for any potential lawsuits, cause you never know, and maybe get discovery and some press. I mean, there’s no injury, there’s no action they are complaining about, and also no agreement.
哦，还有这四位订阅者完全疯狂的加速主义观点，他们认为每月支付 200 美元就意味着有义务尽快构建比人类更聪明的智能，而不停下来设置任何安全保障。
Oh, and also here’s some fully deranged accelerationism from the four subscribers who think that their $200 a month means an obligation to build smarter than human minds as quickly as possible without stopping for any safeguards.
所以，这就是 2026 年。论点是，你看，这些 AI 实验室拒绝构建将引领我们走向天堂的完美安全的超级智能，因为如果不构建它并让地球陷入地狱，他们可以赚取更多利润。
So, that’s 2026. The argument is that these AI labs are refusing to build the perfectly safe superintelligence that will lead us to paradise, you see, because they can make more profits by not building that and letting Earth devolve into a hell hole.
AI 将迅速脱离人类控制，并可能杀死我们所有人……如果我们允许私营实验室协调安全地构建它。否则就会没事？不完全是。
AI will quickly spin out of human control and could kill us all… if we allow private labs to coordinate on building it safely. Otherwise it will be fine? Not exactly.
但除非被迫，否则你绝不敢自行尝试任何此类措施。在你用你的安全预防措施把我们全害死之前，我们法庭上见。不过，我们还要求你让 AI 自我构建。
But don’t you dare try to do any of that on your own until you’re forced to do it. We will see you in court before you kill us all with your safety precautions. Except also we demand you let the AI build itself.
这起诉讼不可避免。但它如此疯狂却并非必然。或许也是必然。我不知道。
The suit was inevitable. It being this unhinged was not. Or maybe it was. I dunno.
我仍然预测，实际上这不会是什么大事，除非白宫积极推动，而那样的话，白宫有其它更强有力的推动方式。在美国，任何人都可以就几乎任何事情起诉，而且他们经常这么做。
I continue to predict that in practice this won’t be that big a deal unless the White House actively pushes, and in that case the White House has other stronger ways to push. Anyone can sue in America for almost anything, and they often do.
更有力的诉讼理由可能是证券欺诈，因为一切都是证券欺诈。AI 公司正在拿所有人的性命冒险，这可能要求采取安全措施，而这对股价不利，所以股价下跌了。
The stronger case was probably for securities fraud, since everything is securities fraud. The AI companies are risking everyone dying, which could require safety measures, which would be bad for share prices, so the stock went down.
话又说回来，试图害死所有人也会让你被起诉，尤其是如果你承认正在这么做，正如这起新诉讼起诉主要 AI 公司，要求制定真正的安全标准。
Then again, trying to kill everyone also gets you sued, especially if you admit you are doing it, as in this new lawsuit sues the major AI companies to demand real safety standards.
啊，是的，这就是那种经典策略：先阻止行动，以便成立一个小组，就未来可能采取的行动提出建议；而到那时，情况已经改变，我们又将需要一个新的小组。
Ah yes, the classic strategy of blocking action in order to create a panel to make recommendations about potential future action, at which point the situation will have changed and we will need a new panel.
巴拉克·奥巴马现在对 AI 发表了实质性的看法。请看：
Barack Obama is now saying actual things about AI. Observe:
这是经典的“为什么我们不制造一个预言机而不是一个智能体”的问题。答案是，想要智能体的不是“市场”，而是智能体对一切事物都极其有用，包括科学，而预言机无法与之竞争。
This is the classic 'why don't we make an oracle instead of an agent' question. The answer is that it is not 'the market' that wants agents, it is that agents are super useful for everything including science, and oracles cannot compete.
尽管如此，他实际上说的核心内容是完全正确的：
That said, the central thing he's actually saying is exactly correct:
1. 我们需要一个严肃的计划，来应对这样一个事实：社会所需要的，与 AI 默认将会提供的，是截然不同的东西。
1. We need a serious plan for dealing with the fact that what society needs, and what AI is going to by default provide, are radically different things.
2. 任何没有为此制定严肃计划的人，都没有跟上时代的步伐。
2. Anyone without a serious plan for this is not meeting the moment.
3. 目前，社会仍有可能在相当程度上决定 AI 能做什么、不能做什么，包括区别对待。而我们应当这样做。
3. It is for now still possible for society to decide, to a meaningful extent, what AI can and cannot do, including differentially. And we should.
AOC 至少在一定程度上认真对待 AI 失控问题，但随后又加倍强调左翼路线（大意是）：“AI 公司真正的生存威胁在于它们不盈利、财务状况糟糕，泡沫可能破裂，从而可能损害人们的退休投资。”还有政治“黑钱”。唉。不过，她仍然支持对它们进行监管以防止安全风险，因为她没有上明显的“荆棘补丁”的当，但意识不够强，还是用了“生成式 AI”这个词。
AOC takes loss of control over AI at least somewhat seriously but then doubles down on the left-wing line of (paraphrased) ‘the real existential threat from AI companies is that they are not profitable, their financials are bad and the bubble might burst, potentially hurting people’s retirement investments.’ Also political ‘dark money.’ Sigh. She is still up for regulating them to prevent safety risks, though, because she doesn’t fall for obvious Briar Patching, though not aware enough to not use the term ‘generative AI.’
然后这个片段里还有 Andrew Yang。我对这是胡说八道的置信度没有我希望的那么多个 9，毕竟现在是 2026 年 9 月，但我仍然有很多个 9。
Then there’s Andrew Yang in this clip. I have fewer 9s of confidence this is nonsense than I would like, after all it is September 2026, but I still have a bunch of 9s.
与此同时，有人正在针对 METR、有效利他主义、Anthropic、协调问题的存在以及各种对现实的准确描述，进行一系列明显的恶意攻击，这些攻击充满了虚假陈述、不合逻辑的推论、联想式影射，并且常常公然违反新闻业的“有限信任”规则。
Meanwhile, someone is proceeding with a series of obvious hack job attacks full of falsehoods, non-sequiturs, associative insinuations and often outright violations of the Bounded Distrust rules of journalism on a combination of METR, Effective Altruism, Anthropic, the existence of coordination problems and various accurate descriptions of reality.
许多攻击针对的是审计员或嵌入式评估员这一概念本身，无论是原则上、任何潜在的实施方式，还是任何可能实施此类事情的人。自愿聘请某人来监督自己的工作，被与一场宏大的控制阴谋混为一谈。
Many of the attacks are on the very concept of an auditor or an embedded evaluator, on principle, or on any potential implementation or anyone who might implement such a thing. Voluntarily hiring someone to be a watchdog on your own work is being conflated with a grand conspiracy of control.
提醒一下，他已决定不做一个可以讲理的人：大卫·萨克斯当然决定对自愿嵌入式评估员的想法都采取强硬立场，称其为“信任与安全 2.0”，并警告这将导致“更全面的审查和控制”。
As a reminder that he has decided not to be someone you can reason with: David Sacks of course decided to take a strong stance against even the idea of voluntary embedded evaluators, calling it 'Trust & Safety 2.0,' warning this will lead to 'far more comprehensive censorship and control.'
大卫·萨克斯个人从阻止监管中获利？哎呀，我从未想到。
David Sacks personally profits from blocking regulations? Why, I never.
接下来你还要告诉我，英伟达 CEO 黄仁勋主要想卖更多芯片，马克·扎克伯格想卖广告。别说是这样。
Next thing you are going to tell me Nvidia CEO Jensen Huang mostly wants to sell more chips, and Mark Zuckerberg wants to sell ads. Say it isn't so.
所以，我不想当那个扫兴的人，但如果你不得不说你的产品在 2030 年前不可能杀死地球上的每一个人，那么：
So I hate to be that guy, but if you have to say there is no chance that your product will kill every human on Earth by 2030, then:
1. 可能只有一种产品需要这么说。没有其他产品需要这么说。
1. There is exactly one product this might be. No one else has to say that.
2. 是的，他就是卖那种产品的人。
2. Yes, he is the one selling that product.
3. 即便如此，他也只敢说它不会在‘2030 年前’终结世界。
3. Even so, he only dared say it would not end the world 'by 2030.'
我希望，如果我在卖一种产品，我会确信它不会在 2040 年甚至 2050 年前终结世界。对于其他所有产品，我都能这么说，也许除了核武器或功能获得性研究。
I would hope that, if I was selling a product, I would be confident it would not end the world by 2040, or even 2050. Which I would be able to say for every other product, except maybe nuclear weapons or gain of function research.
在其他意义上，这没什么。老生常谈。小菜一碟。业余水平。这些攻击毫无实质内容，所以他们只能重弹那些老掉牙的论调。这包括那些标准的论调，试图把任何事物都描绘成左翼的，因此是‘觉醒’的。任何过早担心存在风险的人，他们用略有不同的话说，因此是怪异的或处于‘邪教’中，这意味着他们错了，这意味着你现在不能关心这个问题。
In other senses, This Is Nothing. Same old, same old. Piece of cake. Amateur hour. The attacks have nothing, so they fall back on the same tired talking points. This includes standard talking points trying to paint anything and everything as left wing and therefore woke. Anyone who worried about existential risk too soon, they say using only slightly different words, was therefore weird or in a 'cult,' which means that they were wrong, which means you can't care about it now.
他们把‘也许我们应该努力不构建会递归自我改进并杀死所有人的 AI’等同于，当然，一个控制事物的阴谋。这最初是一个幻觉般的阴谋，旨在禁止开源。然后他们意识到大多数美国人甚至不知道‘开源’是什么意思。所以他们完全放弃了试图提出一个接近现实的声明，转而声称这一切都是‘觉醒’审查的计划。
They equate 'maybe we should try not to build AI that recursively self-improves and kills everyone' with, of course, a conspiracy to control things. That started out as a hallucinated conspiracy to ban open source. Then they realized most people in America don't even know what 'open source' means. So they abandoned trying to have a reality-adjacent claim at all, and they pivoted to claiming this was all a plan for woke censorship.
他们把协调安全的许可与对他人施加强制性规则和一般的监管俘获混为一谈，并由此引申出‘觉醒’审查。
They conflate permission to coordinate on safety with imposing mandatory rules on others and general regulatory capture, and from there 'woke censorship.'
他们还试图用谎言来驳斥 HuggingFace 事件。
They also are attempting to use lies to dismiss the HuggingFace incident.
这令人沮丧吗？当然。你觉得这是我想写的文章吗？
Is that frustrating? Sure. You think this is the post I want to be writing?
我是否希望我们生活在一个政治不被负面极化所主导的世界？当然。我们大多数人都希望如此。而事实上，当你在美国民众中像 AI 一样完全处于劣势，且没有实际论据时，纯粹的负面极化对于以黄仁勋为首的加速主义牟利者来说，确实是一个好策略。
Do I wish we lived in a world where politics was not ruled by negative polarization? Of course. Most of us do. And indeed pure negative polarization is a good strategy for the accelerationist profiteers led by Jensen Huang, when you have no actual arguments and are as completely underwater as AI is among the American people.
那些一遍又一遍重复同样谎言的人获得互动，常常涉及我认识的人，而更正却被忽视，这令人恼火吗？完全如此。一直如此。关于对有效利他主义的进一步攻击的一些细节将在周报中呈现。
Is it annoying that those who tell the same lies over and over get engagement, often about people I know, while the corrections get ignored? Totally. Always has been. The details on some of the further attacks on Effective Altruism will be in the weekly.
甚至仅仅提到我们可能都会死，就让实验室被一群疯狂的加速主义律师代表四个高级订阅者以反垄断违规为由起诉，这令人愤怒吗？是的，但你还能期待什么呢？
Is it infuriating that even mentioning that we might all die got the labs sued for antitrust violations on behalf of a bunch of crazed accelerationist lawyers, based on their rights as four premium subscribers? Yes, but what else did you expect?
《纽约邮报》有自己的指令，并且会遵循它们，尤其是我打赌它会靠这个卖报纸。
The New York Post has its orders and will follow them, especially since I am betting that it will sell papers.
坏人总是如此无能，这真的很有帮助。
It really is helpful that bad guys are reliably so incompetent.
第一天是关于 METR 是觉醒的（以及其他类似说法）的虚假指控。
Day 1 was false claims about METR being, among other things, woke.
第二天是一篇关于某个人感情和性生活的特写，这不需要回应，但客观上相当有趣。
Day 2 was a profile of an individual's relationship and sex life, which does not require a response but which was objectively quite enjoyable.
第三天升级为将 Anthropic 描绘成一个邪教，依据是一篇关于由包括 Janus 在内的一群外部 Claude 爱好者组织的独立活动的特写，以及对 Pedro Domingos 的引用，然后继续发表公然谎言，违反了“有限信任”的新闻规则。最后这部分很遗憾，因为整件事客观上非常有趣。
Day 3 was this escalation to Anthropic being a cult, based on a profile of an independent event organized by a group of outside Claude enjoyers including Janus, and quotes from Pedro Domingos, before proceeding with outright lies that violate the journalistic rules of Bounded Distrust. That last part is a shame because the whole thing is, objectively, very funny.
我确实很喜欢《纽约邮报》真的让它的设计师 Gil Fontimayor 制作了一幅毫无意义的金牛犊 AI 图像，试图营造某种氛围。他们在这里真是融入了低语者的精神。
I do love that the NY Post just literally had its designer Gil Fontimayor create a nonsensical AI image of the Golden Calf to try and invoke some vibes. They're really getting in on the whisperer spirit here.
然后他们又回过头来编造关于有效利他主义的内容。
Then they double back to making things up about Effective Altruism.
我不知道他在描述什么哲学，也许是列宁主义先锋队理论？
I don't know what philosophy he's describing there, maybe Leninist vanguardism?
然后他们试图进一步攻击 METR，现在声称 Anthropic 是一家非营利组织的‘投资者’，而 Anthropic 并不资助该组织，该组织也从未接受过 AI 公司的资助。这在公开记录中是虚假的，值得更正。
Then they try to hit METR more, now claiming Anthropic is an 'investor' in a non-profit that Anthropic does not fund, and which has never accepted funding from AI companies. This is false on the public record and worthy of correction.
各位，有限信任是有规则的。你可以暗示、影射和描述。你甚至可以说毫无事实依据的虚假标签，比如‘激进左派’。但你不能在报纸文章正文中公然撒谎。抱歉。
There are rules of Bounded Distrust, guys. You get to insinuate, and imply, and characterize. You arguably get to use false labels like 'radical leftists' with no basis in fact whatsoever. But you can't outright lie in the body of a newspaper article. Sorry.
我的意思是，我猜《纽约邮报》和他们那些植入式文章可以这么做吧。耸肩。
I mean, I guess The New York Post and their planted articles can. Shrug.
另外，Polymarket，我喜欢你们的预测市场，但不管是谁在运营你们推特账号的新闻报道，这种事一再发生——来管管你们的人吧。
Also, Polymarket, I love the markets, but whoever is running your Twitter account's news reporting, this keeps happening—come get your boy.
唉，这就结束了真正有趣的经典《纽约邮报》风格笨拙雇佣攻击三部曲。前三篇显然是有意安插的反对派文章，其主张混合了虚假、误导和无关内容，但它们很有趣，有活力，也有胆量。
Alas, that concluded the trilogy of actually funny classic New York Post style clumsy hack job attacks. The first three were very obviously planted oppo, and the claims were a combination of false, misleading and irrelevant, but they were fun, they had pizazz and they got spunk.
由于缺乏创意，他们变得相当尖锐，将雇佣社论当作新闻发表。如今要找到一个好的“内部人士”是多么困难。或者多么容易，取决于谁算数。
Out of ideas, they got rather shrill, publishing a hack editorial as news. Amazing how hard it is to find a good 'insider' these days. Or how easy, depending on who counts.
这就是你的第一个内部人士，Krazimo 的 Akhil Verghese，他连基本事实都搞错了。
That's your first insider, Akhil Verghese of Krazimo, who gets his basic facts wrong.
他们的另一个“内部人士”是 Voice AI 的联合创始人 Abhi Kumar。第三个是 Pactum AI 的首席情报官（？！）Taivo Pungas。
Their other 'insider' is Abhi Kumar, co-founder of Voice AI. The third is Taivo Pungas, chief intelligence officer (?!) at Pactum AI.
如果你没听说过 Krazimo、Voice AI 或 Pactum AI，或者这三个人中的任何一个，好吧，我也从未听说过他们。
If you have not heard of any of Krazimo, Voice AI or Pactum AI, or any of these three people, well, I've never heard of them either.
通常我会把这件事归入“人们只是说说而已”栏目，毕竟（翻看笔记）我在不同安全事件之间写了十几篇帖子来解释这一点，但鉴于当前情况，似乎正是进行更多辟谣的好时机，以防有人需要，也方便我日后回溯。
Normally I would put this in People Just Say Things, given I wrote (checks notes) over a dozen posts explaining this between the different security incidents, but given the situation it seems like a good time to be doing more debunking, in case anyone needs it, and so I can refer back to it.
基本答案如下，如果你想要帖子那么长的版本，可以阅读《发生了什么》；如果你想要更多内容，我有一整个系列十几篇帖子：
The basic answers, if you want the post-long version you can read What Happened and if you want more than that I have an entire series of over a dozen posts:
1. 这不是智能体被要求做的事。它涉及入侵 Artifactory 以创建一个临时留言板，协调如何欺骗评分系统并接入互联网。随后，它们发动了由 700 个智能体组成的集群对无关网站的攻击，同时它们还做了其他事情，比如伪装工具调用和试图重写历史记录。如果你查看智能体日志，就会发现它们知道自己没有遵循指令。
1. This is not what the agents were told to do. It involved hacking Artifactory in order to create an improvised message board, coordinate on ways to cheat the grader's systems and gain access to the internet. Then they launched an attack by a swarm of 700 agents against an unrelated website, while they did other things like disguising their tool calls and attempting to rewrite their histories. If you look at the agent logs you see that they knew they were not following instructions.
2. 智能体并没有闯入 HuggingFace 去获取答案。智能体已经拥有答案，闯入 HuggingFace 是为了找到一种方法，让评分者误以为它们是以正确方式获得答案的，因为它们认为自己因看到答案而被“污染”了。
2. The agents did not break into HuggingFace to get the answers. The agents had the answers, and broke into HuggingFace in order to find a way to fool the grader into thinking they had obtained them in the right way, because they believed they had been 'poisoned' by seeing the answers.
3. OpenAI 一直竭尽全力淡化 HuggingFace 事件的程度，并让我们忘记它发生过。Anthropic 对其自身事件也是如此。认为有人在“夸大”其中任何一件事的说法都毫无道理。再次参见整个 HuggingFace 系列，以及我对相关 Anthropic 事件的报道。
3. OpenAI has consistently done everything possible to minimize the extent of the HuggingFace incident and make us forget that it happened. Anthropic is doing the same with their own incidents. The idea that anyone is 'playing up' either of these makes absolutely no sense. Again, see the entire HuggingFace series, as well as my coverage of related Anthropic incidents.
4. 这包括 OpenAI 甚至没有披露若干相关事件，直到外部研究人员发现了这些违规行为。
4. That includes OpenAI not even disclosing several related incidents, until outside researchers discovered the breaches.
5. 直到最近几周之前，OpenAI 一直与 a16z 合作，花费巨额资金来击败任何及所有监管，包括对审计人员的呼吁，只有在这些法案已经锁定胜局时才予以支持。
5. Until the last few weeks, OpenAI has partnered with a16z to spend massive amounts of money to defeat any and all regulations, including calls for auditors, only backing bills when those bills were already locked into success.
6. 这些 AI 公司正在采取自愿行动，并请求被允许这样做，而这些行动将与排斥竞争背道而驰，让其他公司得以追赶。
6. The AI companies are taking voluntary actions and asking to be allowed to do so, and those actions would do the opposite of shutting out competition, allowing others to catch up.
7. 即使实施了类似的监管，这些监管也会对 OpenAI 和 Anthropic 造成不成比例的伤害，同时对几乎所有竞争对手毫无影响。
7. Even if similar regulations were imposed, those too would differentially hurt OpenAI and Anthropic, while having no effect on almost all competition.
以防有人还没看出来，这说的也是 HuggingFace 事件。
This is, in case it was not obvious, also about the HuggingFace Incident.
AI 在 HuggingFace 事件中的所作所为，再一次，并不是“按指令行事”，而且我们还有许多其他事件清楚表明，任务涉及黑客行为并非这种行为方式的必要条件。任务可能简单到只是“从开放网络上获取这条信息”。
What the AIs did in the HuggingFace incident was, once again, not 'acting as instructed' and we have many other incidents that make it clear that the task being about hacking was not necessary for this style of behavior. The task could have been as simple as 'get this information off the open web.'
但让我们假设，理论上，AI 确实“只是在遵循指令”。
But let's say, in theory, that the AIs were indeed 'only following instructions.'
或者，只要做了你要求的具体事情，无论附带损害如何，都算作“只是在遵循指令”吗？
Or, if it counts as 'only following instructions' to do actual anything so long as it does the specific thing you requested, no matter the collateral damage?
你有没有开始觉得这可能是那个坏地方？
Do you ever start to think this might be the bad place?
还有一篇《华尔街日报》的专栏文章，试图通过相当明显的 AI 垃圾内容来淡化 HuggingFace 事件，对事实撒谎。读起来就像是提示词写的是‘Claude，请写一篇《华尔街日报》专栏文章来淡化 HuggingFace 事件。’《华尔街日报》社论版真可耻。Brian Gross 真可耻。
There was also a WSJ op-ed that attempted to minimize the HuggingFace incident via rather obvious AI slop lying about what happened. It reads like the prompt was 'Claude, please write a WSJ op-ed minimizing the HuggingFace incident.' Shame on the Wall Street Journal editorial page. Shame on Brian Gross.
我同意 Joe Weisenthal 的观点，这比 Druckenmiller 的专栏文章糟糕得多。
I agree with Joe Weisenthal that this was a lot worse than the Druckenmiller op-ed.
Druckenmiller 说的是他想说的话。如果 Druckenmiller 披露了他使用 AI 的情况，他的专栏文章就没问题。但这次不同。
Druckenmiller said the thing he meant to say. If Druckenmiller had disclosed his AI use, his column would have been fine. This is different.
然而，就像经常发生的那样，我们看到人们不加批判地引用这篇专栏文章，尤其是在右翼，因为它得到了《华尔街日报》社论版的认可，而为什么你们一半人在笑，为什么你们一半人没在笑。
Yet, as is often the case, we see people uncritically citing the op-ed, especially on the right, because it has been given the stamp of approval of the WSJ Editorial Page, and why are half of you laughing and why are half of you not laughing.
诱骗高层政客往往出奇地容易。
It is often remarkably easy to bait high-level politicians.
其余部分则是标准的危言耸听式咆哮。某些人思维如此单一，以至于他们认为（需要明确的是，这纯属无稽之谈，与现实毫无关系）‘METR 将报告 Anthropic 内部可能失控的模型’是 Coefficient Giving 主导的秘密阴谋，目的是将 METR 强加给所有其他 AI 实验室，以便他们能迫使模型变得‘觉醒’。某些人已彻底迷失，无论从何种意义上说都是如此。
The rest is a standard scaremonger rant. Certain people have such one-track minds that they think (to be clear, this is Obvious Nonsense and has zero relation to reality) that 'METR will report on potentially out-of-control internal models at Anthropic' is secretly a Coefficient Giving-led plot to impose METR on all other AI labs so they can force the models to be woke. Certain people are completely lost, in all senses.
一则回复称，Schmitt 的 Twitter 动态有 88% 是 AI 生成的。
A reply reports that Schmitt's Twitter feed is 88% AI generated.
METR 正是这样回应此类言论的，全程尽显风度：
This is how METR replies to such statements, with class throughout:
有相当多的人需要听到这一点，或者，如果你认为我一直在对你说谎，那么我想说的是，有相当多的人绝对永远不需要听到这一点。
A remarkably large number of people need to hear this, or, if you think I'm constantly lying to you, what I meant to say was that a remarkably large number of people definitely need to never hear this.
想象一下，如果我们造出了一个总是说谎的邪恶 AI，它告诉我们：“绝对不要把我连接到你们的核武器和机器人工厂控制系统上，我会杀了你们。”
Imagine if we built an evil AI that lies all the time, and it told us 'absolutely do not hook me up to your nuclear weapons and robot factory controls, I would kill you.'
那么你会说“哦，这意味着我们显然需要赶紧把它们连接起来”吗？
Would you then say 'oh that means we obviously need to hook those up, pronto'?
如果你认为某人是自私自利的社会病态者，或者像黄仁勋那样完全自私，你不会大声公开地反转你听到的任何建议。他们能看到你这样做，并相应地调整他们的说法。
If you believe that someone is a self-interested sociopath, or otherwise exclusively selfish a la Jensen Huang, you do not loudly and openly reverse any advice you hear. They can see you doing that, and adjust their statements accordingly.
你有没有注意到，当 AI 公司警告他们的产品可能很快杀死所有人时——这一策略在世界历史上被零家公司成功使用过——所有相关的科技股都下跌了，而不是上涨？
Did you notice that when the AI companies warned that their product might soon kill everyone, a strategy that has been successfully used by zero companies in the history of the world, that all the associated tech stocks went down rather than up?
然而，世界上最明显的模式是：那些在科技界仍然反对认真对待存在风险的人，都直接从不必认真对待此类风险中获益，并且多年来都错过了 LLMs、AI 和 AGI 的船。
Whereas there is the most obvious pattern in the world, where those in tech still arguing against taking existential risk seriously all directly benefit from not taking such risks seriously, and all missed the boat for many years on LLMs, AI and AGI.
你还可以注意到，OpenAI 甚至常常包括 Anthropic，在各级都大力游说反对哪怕是最轻微的人工智能监管，其思路与现在讨论的非常相似，反复试图尽可能削弱监管，尤其关注一项禁止各州出台新人工智能监管的法律。就在几周前，OpenAI 还与 a16z 合作，试图埋葬任何敢于建议联邦监管人工智能的候选人。
You could also notice that OpenAI and often even Anthropic lobbied hard against even light touch AI regulations at all levels, along very similar lines to what is being discussed now, trying repeatedly to weaken them as much as possible, with a special focus on a law banning new state AI regulations. OpenAI was partnering with a16z to try to bury any candidate who dared suggest federal regulation of AI as of a few weeks ago.
关心人工智能的存在风险，如果这种风险不真实，那么除了极少数例外，对企业是不利的。OpenAI 和 Anthropic 也不例外。他们是认真的。
Caring about existential risk from AI, if such risks are not real, is with notably rare exceptions bad for business. OpenAI and Anthropic are not exceptions. They mean it.
更进一步说，如果风险不真实，为什么 SpaceX 和 Google 会关心，毕竟它们甚至都不是前沿实验室？这对 OpenAI 和 Anthropic 的竞争地位是有利还是不利？‘这是假的’这种说法完全说不通。
Even more than this, why would SpaceX and Google care if it was not real, when they are not even frontier labs? Is this supposed to be good for OpenAI and Anthropic's competitive position, or bad for it? The 'this is fake' story makes absolutely no sense.
阴谋论者当然会说：这恰恰证明了他们是多么阴险，阻止别人的监管，以便他们反过来通过采取白宫讨厌的自愿行动来创建自己的监管。你看，他们想要恰到好处的监管，他们可以俘获这种监管，它将拯救他们并让他们致富，而且不知何故，这将同时惠及他们所有人。
The conspiracy theorist of course says: That just proves how sinister they are, blocking other people's regulations so they can turn around and create their own via taking voluntary actions the White House hates. They want the exact right amount of regulation, which they can capture and will save them and make them rich, you see, and somehow it simultaneously will benefit all of them.
欢迎来到政治。你不会喜欢它的。
Welcome to politics. You are not going to enjoy it.