Here is a strange fact to start with.先说一个奇怪的事实。 The only sample of an alien ocean that human beings have ever gotten hold of was not scooped up by a lander or carried home in a flask.人类迄今获得的唯一一份外星海洋样本，既不是着陆器舀取的，也不是装进瓶子带回来的。 It was tasted in flight, by a spacecraft that is now dead, out of grains of ice smaller than specks of dust.它是在飞行途中被尝到的——由一艘如今已经报废的航天器，从比尘埃颗粒还小的冰粒中尝到的。 The ocean belongs to Enceladus, a small icy moon of Saturn.这片海洋属于土卫二（Enceladus），土星的一颗小小的冰质卫星。 And a study published this week says that those grains, the only sample we have, have been quietly misleading us. Not lying, exactly.本周发表的一项研究说，那些冰粒——我们拥有的唯一样本——一直在悄悄误导我们。倒不完全是在撒谎， Just sorted. Let me set the scene, because it is one of the great pieces of luck in planetary science.只是经过了筛选。让我先把场景交代一下，因为这是行星科学中一桩极其幸运的事。

Enceladus is tiny, about five hundred kilometers across, small enough to tuck inside a mid-sized country. It is wrapped in a shell of ice.土卫二很小，直径约五百公里，小到可以塞进一个中等大小的国家里。它外面裹着一层冰壳。 Underneath that shell is a global ocean of liquid salt water.冰壳底下是一片全球性的液态咸水海洋。 And out of long cracks near its south pole, that ocean sprays into space in continuous geysers, hundreds of kilometers high.而从它南极附近的几道长长的裂缝里，这片海洋以连续不断的间歇泉喷向太空，高达数百公里。 You do not have to drill through the ice to sample the sea. The sea comes to you.你不必钻穿冰层去取海水。海水会主动找上门来。

The Cassini spacecraft flew straight through those plumes, more than once, in the years before it was deliberately crashed into Saturn in 2017.在 2017 年被有意坠入土星之前的那些年里，卡西尼号（Cassini）航天器不止一次地径直穿过那些羽状喷流。 On board was an instrument called the Cosmic Dust Analyzer.它搭载了一台名为宇宙尘埃分析仪（Cosmic Dust Analyzer）的仪器。 Each ice grain that struck it was vaporized on impact, and the instrument read the chemistry of the flash.每一粒撞上它的冰粒都会在撞击瞬间汽化，仪器则读取这道闪光的化学成分。 That is where the famous Enceladus headlines came from. The ocean is salty, like ours. There are organic molecules in it.那些著名的土卫二头条新闻就是这么来的。这片海洋是咸的，和我们的海一样。里面有有机分子。 A couple of years ago the same team found phosphates, one of the basic ingredients of life as we know it.几年前，同一个团队发现了磷酸盐——我们所知的生命的基本成分之一。 Every one of those was a real detection, and every one made Enceladus a little more interesting.这些发现每一项都是真实的探测结果，每一项都让土卫二变得更有意思一点。

But all of those headlines rested on a quiet assumption. The assumption was that each grain of ice is a tiny, faithful scoop of the ocean.但所有这些头条都建立在一个不动声色的假设之上。这个假设是：每一粒冰都是海洋忠实的一小瓢。 Taste the grain, and you taste the sea it came from.尝一尝这粒冰，就等于尝到了它所来自的那片海。 The new study, led by Frank Postberg in Berlin with colleagues in the United States, Japan, China and Britain, and published in Science Advances, says that assumption does not hold.这项新研究由柏林的 Frank Postberg 领衔，合作者来自美国、日本、中国和英国，发表在《科学进展》（Science Advances）上，研究说这个假设站不住脚。 The grain is not a scoop of the ocean. It is a scoop that the moon has already run through a freezer.这粒冰并不是海洋的一瓢。它是这颗卫星已经先过了一遍冷冻机的一瓢。 And freezing, it turns out, sorts chemistry. Here is the mechanism, and it is something you have seen in your own kitchen.而事实证明，冷冻会对化学成分进行筛选。机制是这样的，而且是你在自家厨房里见过的东西。

Think about ice cubes. The ones from your freezer are usually cloudy and white in the middle.想想冰块。你家冰箱里出来的冰块，中间通常是浑浊发白的。 But the clear, glassy ice in a good cocktail bar is frozen slowly, on purpose. Why the difference?但高档鸡尾酒吧里那种清澈、像玻璃一样的冰，是特意慢慢冻出来的。为什么会有这种差别？ When water freezes, the growing ice crystal is picky. It wants pure water.水结冰时，不断生长的冰晶很挑剔。它只想要纯水。 It shoves almost everything else, the dissolved air, the minerals, the salt, ahead of itself into the liquid that has not frozen yet.它会把几乎所有其他东西——溶解的空气、矿物质、盐——统统推到自己前方，推进还没结冰的液体里。 Freeze fast, and all that stuff gets trapped in place, and you get cloudy ice.冻得快，这些东西就被原地困住，你得到的就是浑浊的冰。 Freeze slowly, and the impurities get pushed steadily into the last pocket of water to solidify, and the rest comes out clear.冻得慢，杂质就被稳步地推进最后一小洼凝固的水里，其余部分则冻得清澈。 The salt and the ice separate. The ocean does this on Earth too.盐和冰分离开来。地球上的海洋也会这么做。

Sea ice is nearly fresh, because as it forms it rejects the salt back down into the water below. Engineers use the same trick on purpose.海冰几乎是淡的，因为它在形成时会把盐排回下面的水里。工程师们也特意利用同样的把戏。 It is called zone refining, and it is how you purify the silicon in a computer chip.这叫区域提纯（zone refining），电脑芯片里的硅就是这样被提纯的。 You melt a thin band of a rod and move the band slowly along, and the impurities ride inside the molten stripe and pile up at one end.你把一根棒料上的一小段窄带熔化，再让这条熔带缓缓移动，杂质便随着熔融的条带一起走，最后在一端堆积起来。 Slow freezing is a sorting machine. Now put that in a geyser.缓慢冻结是一台分选机。现在把它放进一座喷泉里。

A droplet of Enceladus ocean water starts to freeze as it rises through the cracks in the ice.一滴土卫二的海洋水在穿过冰层裂隙上升的过程中开始结冰。 If it freezes slowly, and the study's lab work and models say it often does, the different salts do not stay mixed.如果它冻得很慢——而这项研究的实验和模型表明情况往往如此——各种盐分就不会继续均匀混在一起。 They separate into different regions inside the droplet, because each kind of salt crystallizes out at its own temperature as the leftover brine gets more and more concentrated.它们会在液滴内部分离到不同的区域，因为随着剩下的卤水越来越浓，每一种盐都在各自的温度下结晶析出。 Then the half-frozen droplet is flung upward, accelerating to something like a thousand kilometers an hour, and it shatters against the walls of the vent into a spray of tiny shards.接着，这个半冻结的液滴被向上甩出，加速到大约每小时一千公里，撞在喷口壁上，碎裂成一团细小的碎片喷雾。 Each shard comes from a different region of the sorted droplet.每一块碎片都来自这个已被分选过的液滴的不同区域。 So one grain is rich in table salt, the next in carbonates, the next in phosphates, the next in potassium.于是一颗颗粒富含食盐，下一颗富含碳酸盐，再下一颗富含磷酸盐，再下一颗富含钾。 Not because the ocean has those regions. Because the freezing made them. They checked this two ways.这不是因为海洋本身有这些区域，而是因为冻结过程造出了它们。他们用两种方法验证了这一点。

In the lab they froze droplets of water mixed to match Enceladus's ocean, a couple of hundred micrometers across, and watched the salts split apart whenever the cooling was slow, around ten degrees a minute or less.在实验室里，他们冻结了一些按土卫二海洋成分配制的水滴，直径约几百微米，观察到只要冷却速度很慢——大约每分钟十度或更低——盐分就会彼此分离。 Freeze faster and everything stayed evenly mixed.冻得更快，所有成分就保持均匀混合。 Then they went back to the Cassini archive and looked at the chemistry of nine hundred and sixty-one of the salt-rich grains.然后他们回到卡西尼号的数据档案，考察了其中 961 颗富盐颗粒的化学成分。 They fell into at least five distinct chemical families, exactly the kind of split you would expect if single droplets had been sorted and then smashed.它们至少可以归入五个截然不同的化学族——正是你所预期的那种划分，前提是单个液滴先被分选、再被打碎。

And here is the line in the paper that I think lands hardest.而论文里最击中我的，是下面这句话。 One of the authors, Yasuhito Sekine, said the surprise was that all of this diversity could come from droplets originating in essentially the same ocean water.作者之一关根康人（Yasuhito Sekine）说，令人意外的是，所有这些多样性都可能源自本质上来自同一种海洋水的液滴。 Read that twice. The grains look chemically diverse. The ocean they came from need not be diverse at all.请把这句话读两遍。颗粒在化学上看起来多种多样，而它们所来自的海洋却完全不必是多样的。 A single, well-mixed ocean, run through a freezer, produces a whole zoo of different-looking grains.一片单一的、均匀混合的海洋，经过一台“冷冻机”，就产出了一整群看起来各不相同的颗粒。 Which means the diversity in the data is not evidence of a diverse ocean.这意味着，数据中的多样性并不是海洋多样的证据。 The same spread of measurements is just as consistent with a plain, uniform sea and a sorting process sitting in between.同样这片离散的测量结果，也完全可以用一片朴素、均一的海洋加上夹在中间的一道分选过程来解释。 If you have ever tried to reason backward from what an instrument recorded to what was actually out there, you know this trap.如果你曾试着从仪器记录到的数据反推出外面实际存在的东西，你就知道这个陷阱。 Many different realities can leave the very same fingerprint. You cannot tell them apart by staring harder at the fingerprint.许多不同的现实可以留下完全相同的指纹。你没法靠更使劲地盯着指纹看就把它们区分开来。

So let me do the honest two-column accounting, because that is where this story earns its ten minutes.那么，让我老老实实地做一份两栏对账，因为这正是这个故事值得花这十分钟的地方。 What did the work actually show, and what is it being made to say? What it showed is solid.这项工作究竟证明了什么，又被用来说了什么？它所证明的部分是扎实的。

Freezing an Enceladus-like brine sorts its salts, in the lab, reliably, when the freezing is slow.冻结一份类似土卫二的卤水，在实验室里，只要冻得慢，就能可靠地把其中的盐分选开。 And the real Cassini grains come in families that match. That part is a measurement, and it is a careful one.而真实的卡西尼号颗粒也确实分成了与之吻合的几个族。这一部分是测量结果，而且是一次严谨的测量。

What is being claimed on top of that is softer, and worth pulling apart.在此之上所提出的说法则要软一些，值得拆开来看。 The hopeful headline is that this is great news for finding life, because if the freezer concentrates salts into separate grains, it should concentrate organic molecules, maybe even the chemical traces of microbes, into their own grains too.乐观的标题说，这对寻找生命是个大好消息，因为如果这台“冷冻机”能把盐分浓缩进一颗颗独立的颗粒，它应该也能把有机分子——甚至可能是微生物的化学痕迹——浓缩进各自的颗粒里。 A future spacecraft would not have to pick out a faint signal diluted across a whole ocean.未来的航天器就不必从被整片海洋稀释的微弱信号中去挑拣了。 The moon would have pre-concentrated the evidence for it. Postberg put it nicely.这颗卫星已经替它把证据预先浓缩好了。波斯特贝格（Postberg）说得很妙。 Enceladus does a lot of the sample preparation for us that would normally take a chemistry lab back on Earth.土卫二替我们做了许多样品制备工作，而这些通常需要地球上的一间化学实验室才能完成。

That might well be true, and it is a genuinely clever point.这很可能是对的，而且是个确实巧妙的观点。 But notice that it is an extrapolation from salts to biology, not a measurement of biology.但要注意，这是从盐类向生物学的外推，而不是对生物学的测量。 They did not freeze a droplet full of microbes and watch the microbes gather.他们并没有冻结一滴充满微生物的液滴，然后观察微生物聚集。 And the broader headline, that Enceladus is now even more promising for life, quietly swaps two different questions.而那个更宽泛的标题——土卫二如今对生命更有希望——悄悄偷换了两个不同的问题。 Whether we could detect life if it is there, and whether it is there, are not the same question. This study is about the first one.如果生命存在、我们能否探测到它，与它是否真的存在，这并不是同一个问题。这项研究关乎的是前者。 It says nothing, and the authors are careful to say it says nothing, about the second. There is even a cost hidden inside the good news.它对后者什么也没说，而且作者谨慎地指出它什么也没说。好消息里甚至还藏着一笔代价。

The same sorting that makes life easier to detect makes the ocean harder to read.正是那个让生命更容易被探测到的分选过程，使得海洋更难解读。 If each grain is a biased, concentrated sample of one part of the chemistry, then you can no longer take the average composition of the grains and call it the composition of the sea.如果每一颗冰粒都是对某一部分化学成分有偏向、经过浓缩的样本，那你就不能再把这些冰粒的平均组成当作海洋的组成。 You have to work backward through the sorting to recover what the ocean was actually made of.你必须反过来穿过这个分选过程，才能还原出海洋实际上由什么构成。 The salts and the phosphates really are down there. That part stands.盐类和磷酸盐确实就在那下面。这一点成立。 But the amounts, the ratios, the recipe, those now have to be un-mixed from a process that spent its whole time deliberately un-mixing them.但含量、比例、那份配方，如今都得从一个自始至终都在刻意拆解它们的过程里，反向解出来。

That is the real shape of this result, and it is a shape worth recognizing, because it turns up everywhere.这才是这个结果真正的形状，而且这是一个值得认清的形状，因为它随处可见。 Between the thing you want to know and the thing you can measure, there is almost always a process in the middle that transforms the signal.在你想知道的东西和你能测量的东西之间，几乎总有一个居中的过程，它会改变信号。 The ocean is the input. The slow freezing and the shattering are the transform. The grain is the output.海洋是输入。缓慢的冻结和碎裂是变换。冰粒是输出。 Reading the ocean off the grain is the same kind of problem as reading the rainfall off a river downstream, or reading a cause off a single outcome.从冰粒反推海洋，和从下游河流反推降雨、或从单一结果反推成因，是同一类问题。 You are never quite measuring the thing. You are measuring the thing after the world has had its way with it.你从来都不是在直接测量那个东西。你测量的是这个世界对它施加过影响之后的东西。 And the whole discipline is to model that middle step honestly, instead of pretending it is not there.而全部的功夫就在于诚实地为中间那一步建模，而不是假装它不存在。

Enceladus runs a free chemistry lab for us, out past Saturn, flinging prepared samples into space for anyone who cares to fly through.土卫二在土星之外为我们免费运营着一座化学实验室，把备好的样本抛向太空，供任何愿意飞掠而过的人取用。 This week we learned a little more about how that lab actually works.这一周，我们对这座实验室实际如何运作，又多了解了一点。 And the first thing any good lab teaches you is that the sample on the slide is never quite the thing itself.而任何一座好的实验室教给你的第一件事，就是载玻片上的样本从来都不完全等于那个东西本身。

There is a particular discipline that a diagnostic pathologist spends a career learning, and it is not the one you would guess. It is not how to see. It is how to not over-see. A pathologist sits at a microscope with a sliver of someone's tumor on a glass slide, and the whole job is to read exactly what is there and no more. Not to let the frightening case pull the diagnosis toward the frightening answer. Not to see a pattern because a pattern would be convenient. The skill is calibration. It is holding your reading of the evidence to precisely what the evidence supports, and no further.

Richard Scolyer was one of the best in the world at this. He read melanoma. He died this June, in Sydney, at fifty-nine, and the reason his death is worth ten minutes is not that he was famous, though by the end he was. It is that he spent the last three years of his life turning himself into the single kind of evidence he had spent the previous thirty years warning his colleagues never to trust. And he knew it. That is the whole story. A man whose life was the refusal to over-read a slide became a slide, and then refused to over-read himself.

Start with what he actually did, because the popular version skips it. Scolyer was a pathologist at Royal Prince Alfred Hospital and at the Melanoma Institute Australia, and with an oncologist named Georgina Long he helped build one of the real advances in cancer of the last twenty years. To understand it you need one idea. Your immune system has brakes. T cells, the immune system's attack cells, carry molecular switches that say stand down, do not attack, so that they don't turn on your own body. Many tumors survive by leaning on those switches. A class of drugs called checkpoint inhibitors takes the brakes off. The drugs do not attack the tumor themselves. They release the T cells to do it. That much was already known.

What Long and Scolyer helped establish was stranger and more useful. It was a matter of timing. If a patient has an operable melanoma, the obvious move is to cut it out and then give the drug to mop up whatever is left. They showed the reverse works far better. Give the drug first, while the whole tumor is still in the body, and only then operate.

Why would the order matter? Because the tumor is the training material. While it is still there, the released immune system gets to study the entire thing, the full set of molecular markings that identify these particular cancer cells, and build an army specific to them. Cut the tumor out first and you have thrown away the mugshot before the detective ever looked at it. There is a clean piece of evidence for this from brain cancer, a randomized trial run by Timothy Cloughesy in 2019. Same drug, two groups, the only difference being the timing. The group given the drug before surgery lived about thirteen and a half months. The group given it only afterward lived about seven and a half. Same molecule. The variable was when.

So Scolyer had spent a career on a treatment whose power came from a controlled comparison. Two groups, one difference, a clean readout. Hold on to that, because it is exactly what his own case could never be.

In June of 2023, at a conference in Poland, Scolyer collapsed with a seizure, and the scan showed a glioblastoma. This is the worst of the common brain cancers, and his was the worst subtype. The median survival, even with everything modern medicine can offer, surgery and radiation and chemotherapy together, is around fifteen months. And glioblastoma is the hardest possible target for the treatment he knew best. Immunologists call it a cold tumor. It sits behind the blood-brain barrier, which keeps immune cells out. It carries very few mutations, which means very few flags for the immune system to recognize. It is nearly empty of the very T cells the checkpoint drugs are meant to release. Everything that made melanoma a good target, glioblastoma reverses.

He and Long decided to throw the melanoma playbook at it anyway, and further than anyone had gone before. Before his surgery, he received not one checkpoint drug but three at once, each releasing a different brake. Then, twelve days later, the operation. Later still, a vaccine made from his own tumor's markings, to hand the immune system the flags the tumor refused to fly on its own. No glioblastoma patient had ever been treated this way before surgery. He was the first. Patient zero, in the most personal sense the phrase can carry. And then he did well. Month after month, no sign of the tumor coming back.

Seventeen months out, still clear, when the case was published. He had passed the median. He had passed it by a wide margin. The coverage almost wrote itself. The pathologist who invented a cure and used it on his own brain.

Here is where you have to be as careful as he was, because that sentence is not something anyone can actually know, and Scolyer, more than any journalist writing about him, understood exactly why.

He was one patient. One. And a single patient's survival, however long, cannot tell you whether the treatment did it. Glioblastoma survival is not a single number. It is a curve with a long tail. Most people die within a year or two, but a small minority, unlucky enough to have the disease and lucky within it, live for years on nothing but the standard care. Nobody fully knows why. Their tumor biology was gentler, or the surgery happened to get more of it, or some reason we have not yet named. Now put the two together. If you take one of those long survivors and you happen to have given them an experimental drug, the drug and the tail produce the identical picture. One person, alive at three years. You cannot look at that single point and read off which story put it there. The treatment worked. Or this was always going to be one of the lucky few. Or the surgery did it. The outcome looks the same in every case. The cause is simply not recoverable from it. This is why medicine runs randomized trials at all. Not out of bureaucratic caution.

It is the only known way to pull the drug apart from the tail, and it needs two groups precisely because a single person can never serve as their own comparison. Scolyer had built his career on exactly this logic in melanoma. And he said so, plainly and repeatedly, about himself. His case, he insisted, proved nothing about whether the treatment works. It could not. Proving it on him was never the point.

So what was the point? This is the part that both the triumphant version and the cynical version get wrong. The experiment did establish something real, and it is worth being exact about what. When they operated, they had his tumor in hand, and his blood, drawn before and after the drugs. And they measured it. In a tumor that is supposed to be immunologically cold, nearly empty of active immune cells, they saw the immune system switch on. T cells activating, the machinery of an immune attack turning over. That is a measurement, not a survival story. And it is attributable. You gave the drug, and days later, in the tissue, you saw the immune activation. That is a before and after you can genuinely read. Whether it bought him a single extra week of life, you cannot read at all.

That is the line the whole story turns on, and it is a line Scolyer drew himself, as a dying man, with total clarity. There is what you can measure and attribute, the mechanism, the immune system waking up inside his tumor. And there is what you cannot, the outcome, whether he personally lived longer because of it. The case report was honest about being a single patient whose job was to justify a proper trial, not to stand in for one. That trial is now running. It has other patients, and a comparison group, and it, not Richard Scolyer, is what will answer whether this works.

The tumor came back in March of 2025. He died fifteen months later, this past June. And the thing I keep returning to is that the discipline held. The whole of his professional life had been the refusal to let a slide say more than it can. At the end, the slide was him, the most emotionally loaded case he would ever read, his own brain, his own survival, every reason in the world to see the pattern he wanted to see. And he read it exactly. One patient. Encouraging mechanism, uninterpretable outcome, no conclusions about whether it works, run the trial.

If you have ever tried to reason backward from a single result to the thing that caused it, you know how hard that clarity is, and how rare. One number is consistent with many histories. You cannot invert it. Most of us, handed a single hopeful data point about our own life, would read it as proof. He was handed the most hopeful possible one, his own body outrunning its own prognosis, and he declined to call it proof. That, more than the three drugs or the vaccine, is the thing he was actually expert at.

Five years ago, one of the most exciting sentences in physics was about a wobble. A single particle, a muon, was spinning in a magnetic field, and it was wobbling a little faster than the rulebook said it should. In the spring of 2021 the news went around the world. A crack in the Standard Model. A fifth force of nature. The particle that would not obey. For a while it looked like the biggest thing to happen to physics in a generation.

Now the story has ended, quietly, and in almost exactly the opposite way. There is no new force. Nobody found a new particle. The crack has closed. And here is the part worth ten minutes of your evening: the anomaly did not go away because someone did a better experiment. It went away because physicists learned to compute one stubborn number a little better. The gap everyone thought lay between nature and our theory turned out to be, for the most part, a gap between two different ways of doing the same calculation. That is a very different kind of story, and I think it is the more useful one.

Let me build the picture from the ground up, because the picture is the whole point. Start with the muon.

A muon is, more or less, a fat electron. Same electric charge, same spin, but about two hundred times heavier, and it does not live long. Make one and it survives about two millionths of a second before it falls apart. In that flicker of time it does something very ordinary for a charged, spinning object. It behaves like a tiny bar magnet. Any little spinning charge is a little magnet, the same physics as a compass needle. Now put that tiny magnet in a strong magnetic field.

A spinning magnet in a field does not just line up. It wobbles, the way a spinning top wobbles when gravity pulls on it. The top does not fall over. Its axis sweeps around in a slow circle. Physicists call that precession, and the muon does the same thing. It spins, and its spin axis sweeps around, and the speed of that sweep is what the whole story turns on. How fast should it wobble?

There is a number that sets it, and physicists call that number g. And here is the first beautiful fact. If the muon were a perfect, simple point of charge, the cleanest theory we have, written down by Paul Dirac in the nineteen twenties, says g should be exactly two. Not roughly two. Exactly two, a clean whole number falling out of the mathematics. Except it is not exactly two.

In nineteen forty-eight a young physicist named Julian Schwinger worked out the first correction, and it is a small excess above two, about one part in a thousand. Schwinger was so fond of that little formula that the story goes it is carved on his gravestone. And the reason for the excess is the second beautiful fact, the one that makes this whole experiment worth building.

Empty space is not empty. What we call the vacuum is really a froth of particles flickering into being and vanishing again, too briefly to catch, borrowing energy from nothing and paying it back before nature notices. The muon is never truly alone. It sits inside this froth and constantly brushes against it, and every one of those brushes nudges its wobble. So g creeps above two. The size of the excess is written as a single quantity, defined as g minus two, all divided by two, which is where the experiment gets its name.

And now the payoff of the setup. That excess is a sum. Every kind of particle that exists contributes to the froth, and therefore to the wobble. The known ones, yes, but also, in principle, any particle we have not discovered yet. So if you can measure that wobble to absurd precision, and you can also calculate what the wobble should be from every particle you already know about, then any mismatch between the two is a fingerprint. It would mean there is something in the froth you have not accounted for. Something new. That is why this measurement matters. It is a way of weighing the entire particle zoo at once, including the animals still hiding in the dark.

The measurement itself is one of the loveliest things people have ever built. At Fermilab, outside Chicago, they send muons around a ring about fifteen meters across, a giant magnetic racetrack, and the muons go around thousands of times. As each one falls apart it throws off an electron, and the direction of that electron tells you which way the muon's spin was pointing at that instant. Watch millions of these and you can read the wobble directly. This year they published their final answer, pinned to a precision of about a hundred parts per billion. That is like measuring the distance from here to the moon and being unsure by a few dozen meters. The experiment is not the weak link; it is magnificent. So if the measurement is that good, where did the famous discrepancy come from? It came from the calculation.

Most of the froth is easy: the cleverest physicists grind it out with pen, paper, and computers, and trust the result to many decimal places.

That is the part made of electrons and photons. But there is one contribution that is genuinely hard, and it is the villain and the hero of this whole story. It comes from the strong force, from the flickering of quarks and gluons, the stuff inside protons. The strong force is not tame. You cannot just write down the correction and compute it the easy way. It refuses.

So physicists took two different roads to get that one hard number. The first road is to borrow from the real world. There is a related process you can actually measure in other experiments, electrons and their antiparticles colliding and turning into a spray of strongly interacting matter, and you can use those measurements as a stand-in for the froth you cannot compute. Call that the data-driven road. The second road is to simulate the strong force directly, from first principles, on a grid of space and time inside a supercomputer. Call that the lattice road.

For years, the data-driven road was the trusted one, and when you used it, the theory came out lower than the measurement, and the gap was large. Five standard deviations large, which in this field is the threshold people call discovery. That gap was the anomaly, the fifth force of the headlines. But here is what got lost in the headlines.

The two roads did not agree with each other. When a major team finished the lattice calculation, the simulated one, their number came out higher, close to the measurement, and the gap nearly vanished. So all along there were two theoretical predictions, not one, and they disagreed. The dramatic five-sigma anomaly was really the distance between the experiment and one particular way of computing the hard term. Choose the other way, and there was barely an anomaly at all. Over the last two years the field made its decision.

A fresh measurement of that borrowed process, from a group in Novosibirsk, came out high, agreeing with the lattice road rather than with decades of older data. The community's official prediction switched over to the lattice calculation. And this April, a first-principles computation of that hard term, sharpened to under half a percent, landed the theory within half a standard deviation of the experiment. The Standard Model, confirmed to eleven digits. No room left for the fifth force. The anomaly is, for now, gone.

But I promised you a story about where discrepancies really live, so I have to correct the correction. It would be just as wrong to say the mystery is solved. Because the reason the two roads disagreed has not been explained. The lattice simulations and that one new collision measurement agree with each other and with the wobble. But they all disagree with decades of careful older measurements of the same borrowed process. Somebody's data is wrong, and nobody yet knows whose. So we have not found new physics. What we have found is that we did not fully understand our own inputs. The puzzle did not disappear. It moved. It moved from a fight between nature and theory to a fight among our own instruments.

And that, I think, is the honest lesson, and it travels far beyond particle physics. A discrepancy between what you measured and what you predicted is only ever as trustworthy as the shakiest ingredient in the prediction. For fifteen years, a difference that people were ready to call a new force of nature was, in large part, sensitivity to a single modeling choice buried deep in the calculation. It took two independent roads, and finally a clean simulation, to see that. The signal was real. What it was a signal of was not the universe. It was the seam in our own arithmetic.

There is a new experiment being built in Japan to measure the wobble a completely different way, with results expected around the end of this decade. That is the right response. When your anomaly turns out to live inside your own method, you do not argue about it. You measure it again, by a road that cannot make the same mistake. Physics is slow on purpose. That is not a weakness. This month, it is the whole point.

On the eighth of September this year, OpenAI announced that one of its systems had proved something about the Navier-Stokes equations.今年9月8日，OpenAI 宣布，它的某个系统证明了关于 Navier-Stokes 方程的某项结果。 Within a day the announcement had hardened into a sentence you have probably seen.不到一天，这则消息就凝固成了一句你多半见过的话。 An artificial intelligence has solved a million-dollar Millennium Prize Problem. It is a wonderful sentence.一个人工智能解决了一道价值百万美元的千禧年大奖难题。这是一句绝妙的话。 It is also wrong, and the way it is wrong is more interesting than the way it would be right.它同时也是错的，而它错的方式，比它对的方式更有意思。

So tonight I want to take that sentence apart slowly.所以今晚，我想慢慢地把这句话拆开来看。 Because almost everything true about this story lives in a single short phrase that the headlines quietly dropped.因为关于这个故事，几乎所有真实的部分，都藏在一个被标题悄悄略去的短语里。 The phrase is: with a forcing term. Hold onto it. By the end it will do all the work. First, what the equations are.这个短语是：带一个强迫项（forcing term）。记住它。到最后，全部关键都落在它身上。先说这些方程是什么。

The Navier-Stokes equations describe how a fluid moves. Water down a pipe, air over a wing, blood in an artery.Navier-Stokes 方程描述流体如何运动。管道里的水、机翼上方的空气、动脉里的血液。 They were written down in the eighteen hundreds and they are used every single day by every engineer who touches a fluid.它们在 19 世纪被写下，如今每一位与流体打交道的工程师，每天都在用它们。 And yet no one has ever proved that they always behave. Here is the exact question, the one worth a million dollars.然而，从没有人证明过它们总是规规矩矩。下面是那个确切的问题，价值百万美元的那个。

It is one of the seven Millennium Prize Problems set by the Clay Mathematics Institute in the year two thousand. Take a fluid.它是克雷数学研究所（Clay Mathematics Institute）在 2000 年设立的七道千禧年大奖难题之一。取一团流体。 Give it a smooth, gentle starting flow, with a finite amount of energy. Then leave it completely alone.给它一个光滑、平缓的初始流动，能量有限。然后完全不去碰它。 No stirring, no pushing, nothing from outside. Now let the equations run forever. Does the flow stay smooth for all time?不搅动，不推挤，没有任何外来的作用。现在，让方程永远演化下去。这个流动会在所有时刻都保持光滑吗？ Or can it, at some finite moment, blow up? Blow up is a real technical term, and it is worth picturing.还是说，它会在某个有限的时刻爆破（blow up）？blow up 是一个真正的专业术语，值得在脑海里描绘一下。

It means the velocity, or the steepness of the velocity, becomes infinite at a point.它的意思是，速度，或者速度的陡峭程度，在某一点变成无穷大。 Imagine the swirling of the fluid concentrating into a smaller and smaller region, spinning faster and faster, until at one instant, at one point, it is spinning infinitely fast.想象流体的旋涡集中到越来越小的区域，转得越来越快，直到某一瞬间、在某一点上，它以无穷快的速度旋转。 That is a singularity. It is the equation predicting something physically impossible. If that can happen, the model breaks there.那就是一个奇点（singularity）。是方程预言出某种物理上不可能的东西。如果这真能发生，模型就在那里崩溃了。 If it can never happen, the equations are safe. Nobody knows which is true. That not knowing is the prize. Now, why is this so hard?如果它永远不会发生，那这些方程就是安全的。没有人知道哪个才是对的。这份不知道，正是大奖所在。那么，这为什么这么难？

The answer has a name, and the name matters for the whole story. The name is viscosity.答案有一个名字，而这个名字对整个故事都很重要。这个名字叫黏性（viscosity）。 Viscosity is internal friction, the stickiness of the fluid against itself. Honey has a lot, water has a little.黏性是内部摩擦，是流体对自身的黏滞。蜂蜜的黏性很大，水的很小。 And friction smooths things out. It drains energy away and spreads sharp features into soft ones.而摩擦会把一切抹平。它把能量耗散掉，把尖锐的特征铺展成柔和的。 Viscosity is the thing fighting against blow-up.黏性，正是那个与爆破对抗的东西。 There is a cousin of these equations, called the Euler equations, that has no viscosity at all, no friction.这些方程有一个表亲，叫 Euler 方程，它完全没有黏性，没有摩擦。 Those are easier to make misbehave.那些方程更容易出乱子。 The full Navier-Stokes equations, with viscosity switched on, are harder, precisely because the friction keeps trying to save you.完整的 Navier-Stokes 方程，把黏性打开之后，反而更难，恰恰是因为摩擦一直在试图救你。

So that is the real problem. A fluid, left alone, with friction working against it. Does it tear itself apart or not. And now the phrase.所以这才是真正的问题。一团流体，无人干预，摩擦与它作对。它会不会把自己撕裂。现在，回到那个短语。

With a forcing term. A forcing term is an outside push.带一个强迫项。强迫项就是一种来自外部的推动。 Instead of leaving the fluid alone, you reach in and keep stirring it, according to some rule, forever.你不再让流体自生自灭，而是伸手进去，按照某种规则，不停地搅动它，永远搅下去。 The equations with that extra push added are a different question. And here is the quirk that this entire episode turns on.加上这个额外推动的方程，是另一个问题。而下面，就是整期节目所围绕的那个奇特之处。 When Charles Fefferman wrote the official problem statement for the Clay Institute, he wrote four versions.当 Charles Fefferman 为克雷研究所撰写官方的问题陈述时，他写了四个版本。 Two of them say the flow stays smooth, with no outside force.其中两个说流动保持光滑，没有外力。 Two of them say the flow can break down, and those two were allowed to include a smooth outside force.另外两个说流动可能崩溃，而这两个被允许包含一个光滑的外力。 Many mathematicians now think that was a slip in the wording.如今许多数学家认为，那是措辞上的一处疏漏。 Because if you are allowed to keep pushing the fluid, the whole thing gets much easier.因为如果你被允许持续推动流体，整件事就会变得容易得多。

What OpenAI actually proved is blow-up for the Navier-Stokes equations with a smooth forcing term. A real result.OpenAI 实际证明的，是带有光滑强迫项的 Navier-Stokes 方程会发生 blow-up。这是一个真实的结果。 Checked by machine, which I will come back to. But it is the pushed version.由机器验证——这一点我稍后会再说。但它是被推动的那个版本。 It satisfies one of the four statements as literally written, and it does not touch the version that every working mathematician means when they say the problem.它满足四个表述中按字面写出的其中一个，却没有触及每一位在做这行的数学家提到这个问题时所指的那个版本。 The Clay Institute has not accepted it.Clay 研究所并没有接受它。 Its president, Martin Bridson, said the evaluation would be deliberately unhurried and absolutely rigorous.它的所长 Martin Bridson 说，评估将会刻意地不急不忙、绝对严谨。 The institute still lists Navier-Stokes as unsolved. Why does a push change everything? Here is the picture I would keep.研究所仍然把 Navier-Stokes 列为未解决。为什么一个推动会改变一切？这里有一个我想让你记住的画面。

Think of a child on a swing, and ask whether the swing can go all the way over the bar.想象一个荡秋千的孩子，然后问：秋千能不能一路荡过横杆的上方。 If no one is allowed to touch it, that is a deep question about the swing itself.如果谁都不许碰它，那是一个关于秋千本身的深刻问题。 But if you are allowed to push, in just the right rhythm, then of course it goes over. Everyone knew that.但如果你被允许去推，而且节奏恰到好处，那它当然会荡过去。这一点大家都知道。 The interesting question was always whether it happens on its own. A forcing term is a hand on the swing.真正有意思的问题始终是：它会不会自己发生。强迫项就是搭在秋千上的一只手。 It smuggles energy in from outside, and it can pour that energy into an ever-shrinking region on purpose.它把能量从外面偷偷送进来，而且可以故意把这些能量倾注到一个不断缩小的区域里。 Making a fluid blow up when you are allowed to feed it forever is a much smaller claim than making it blow up when it is sealed off and alone. 在你被允许永远给它喂能量的情况下让流体爆破，比在它被封闭、孤立无援时让它爆破，是一个小得多的论断。

So that is the first correction. The object is not the object.所以这是第一个更正。这个对象并不是那个对象。 The famous question, the undisturbed fluid, is still open, and still sitting there worth a million dollars.那个著名的问题——不受扰动的流体——仍然悬而未决，也仍然摆在那里，值一百万美元。

Now the second correction, which is about the machine. Did an artificial intelligence out-think the mathematicians? No.现在是第二个更正，关于机器。一个人工智能在思考上胜过了数学家吗？没有。 The idea, the actual mathematical construction, is human. It comes from two mathematicians, Diego Cordoba and Luis Martinez-Zoroa.那个想法，真正的数学构造，是人类的。它来自两位数学家，Diego Cordoba 和 Luis Martinez-Zoroa。 Charles Fefferman calls them the heroes of the story. Their method is beautiful and I can describe it. It is a cascade across scales.Charles Fefferman 称他们是这个故事的英雄。他们的方法很漂亮，我可以描述一下。它是一种跨尺度的 cascade。 Picture a large, slow swirl of fluid. That swirl stretches and spins up a smaller swirl inside it.想象一团巨大而缓慢的流体漩涡。这个漩涡拉伸并在它内部催生出一个更小的漩涡。 The smaller one spins up a still smaller one, faster again.更小的那个又催生出一个更小的，转得又更快了。 Energy funnels down the ladder, from big scales to tiny scales, and the steepness of the flow grows and grows at each rung.能量沿着这个阶梯向下汇聚，从大尺度到极小尺度，而流动的陡峭程度在每一级上都越来越大。 The trick is to arrange the ladder so that the whole infinite descent completes in a finite amount of time.诀窍在于把这个阶梯安排好，让整个无穷的下降在有限的时间内完成。 Add a carefully tuned push at each step to keep the books balanced. That is the human invention.在每一步加上一个精心调校的推动，让账目保持平衡。这就是人类的发明。 A storm breeding a tighter vortex breeding a tighter one, all the way down, in a finite instant.一场风暴孕育出一个更紧的涡旋，再孕育出一个更紧的，一路向下，就在有限的一瞬间里。

Two mathematicians extended that idea this summer. One of them is Levent Alpoge, and if that name rings a bell it should.今年夏天，两位数学家拓展了这个想法。其中一位是 Levent Alpoge，如果这个名字让你觉得耳熟，那是应该的。 He is the same person who, last year, used an AI system to shoot down the eighty-seven-year-old Jacobian conjecture, a story I told you back in episode seven.他就是去年用一个 AI 系统击落了那个已有 87 年历史的 Jacobian conjecture 的同一个人，这个故事我在第七集跟你讲过。 This time, with Tristan Buckmaster, and with AI help, he pushed the cascade construction onto the Euler equations, the frictionless cousin.这一次，他和 Tristan Buckmaster 一起，借助 AI，把这个 cascade 构造推到了 Euler 方程上——那个无摩擦的表亲。 Terry Tao wrote their work up in careful detail on the seventh of September.陶哲轩在 9 月 7 日把他们的工作详尽地写了出来。 OpenAI's contribution was to take the next step, to put viscosity back in and reach the real Navier-Stokes equations, still with the forcing term.OpenAI 的贡献是迈出下一步：把黏性重新放回去，得到真正的 Navier-Stokes 方程，同样带有强迫项。

And what did the machine actually do there? It ground and it verified.那机器在这里到底做了什么？它埋头苦算，然后加以验证。 OpenAI ran ten thousand automated agents for eighty-eight hours, exchanging nearly five million messages, at a cost of several million dollars, and produced a proof written in a formal language called Lean.OpenAI 运行了一万个自动化 agent，历时 88 小时，交换了近 500 万条消息，花费数百万美元，最终产出了一份用一种叫 Lean 的形式化语言写成的证明。 And here is my favorite detail, from Tao.而这里有一个我最喜欢的细节，来自陶哲轩。 The first draft the system produced was, in his words, the worst writeup we had ever seen in the history of mathematics.用他的话说，系统产出的初稿是「数学史上我们见过的最糟糕的写作」。 Human mathematicians then spent weeks rewriting it into something a person could read.随后人类数学家花了数周时间，把它重写成一个人能读懂的东西。

That formal-language part is why we can trust the mathematics while distrusting the marketing.正是这个形式化语言的部分，让我们既能信任其中的数学，又能不信任那些营销说辞。 This is the same asymmetry I keep coming back to. Finding something is hard, but checking it can be cheap and near-certain.这就是我反复回到的那种不对称性。找到某个东西很难，但检验它可以很廉价、且近乎确定。 A proof written in Lean is checked mechanically, line by line, by a computer that does not care who wrote it.一份用 Lean 写成的证明会被机械地、逐行地检验，由一台不在乎是谁写的计算机来完成。 So the theorem is very probably correct. The unreliable part of this story is not the theorem. It is the sentence wrapped around it.所以这个定理很大概率是正确的。这个故事里不可靠的部分不是定理本身，而是包裹在它外面的那句话。

I should be fair about the human drama too, and then leave it alone.我也应当公平地对待其中的人事纠葛，然后就此打住。 Buckmaster suggested OpenAI had heard about his team's progress and moved quickly to get ahead.Buckmaster 暗示 OpenAI 听说了他团队的进展，于是迅速行动以抢先一步。 OpenAI's Sebastien Bubeck denied using their work. That dispute is unresolved and I am not going to referee it.OpenAI 的 Sebastien Bubeck 否认使用了他们的工作。这场争议尚未解决，我也不打算去当裁判。

So let me set the two columns down plainly. What is real and new.那么让我把这两栏平实地摆出来。什么是真实且新颖的。 A hard blow-up result for a fluid equation, formally verified, produced over a single weekend.一个关于流体方程的、经过形式化验证的强 blow-up 结果，在一个周末内完成。 That is a genuine milestone for machine-assisted mathematics, and I do not want to wave it away. What is oversold.对于机器辅助数学而言，这是一个真正的里程碑，我不想把它一笔带过。什么是被过度吹捧的。 Solved the Millennium Prize. Out-reasoned humanity. Navier-Stokes is broken. None of those.解决了千禧年大奖难题。在推理上超越了人类。Navier-Stokes 被攻破了。这些都不是事实。

There is a lesson here that lands close to home for anyone who calibrates models against data.这里有一个教训，对任何拿数据来率定模型的人来说都格外贴切。 A result can be perfectly true to the letter of the question and still not answer the question you cared about, because an assumption got smuggled in through a side door.一个结果可以完全忠实于问题的字面，却依然没有回答你真正在乎的那个问题，因为某个假设被从侧门偷偷夹带了进来。 The forcing term is that side door. It is worth learning to hear the phrase that everyone else drops.强迫项就是那道侧门。值得学会去听出那句人人都略过的措辞。

Tao put the real point better than I can.陶哲轩把真正的要点说得比我更好。 The actual solving of these problems, he wrote, is only a proxy goal for the primary goal of developing mathematical understanding and insight.他写道，真正解决这些问题，只是首要目标的一个代理目标，而首要目标是发展数学的理解与洞见。 The forced cascade taught us something about how fluids can concentrate.受强迫的级联让我们了解到一些关于流体如何能够聚集的东西。 But the question that is still worth a million dollars is the quiet one.但那个仍然值一百万美元的问题，是安静的那一个。 Can a fluid, sealed off and left entirely alone, tear itself down to a single infinite point? Nobody pushed it. Nobody knows.一个流体，被密封隔绝、完全独自静置，能否把自己撕裂成一个单一的无穷点？没人推它。没人知道。

They called her a dreamer. Not as a compliment.他们说她是个空想家。这不是恭维。 For most of the nineteen eighties, when Ada Yonath told other scientists what she was trying to do, the polite ones changed the subject and the honest ones told her it could not be done.在整个二十世纪八十年代的大部分时间里，每当 Ada Yonath 告诉其他科学家她想做的事，客气的人会岔开话题，直率的人则告诉她这做不到。 She was trying to grow a crystal of the ribosome.她想培养出核糖体的晶体。 The consensus was that this was impossible, and the consensus was made up of serious, capable people who had tried and failed.共识是这不可能，而持这一共识的，都是些认真、有能力、尝试过却失败了的人。 Yonath died on the thirty-first of August this year, at eighty-seven.Yonath 于今年八月三十一日去世，享年八十七岁。 We are in the week before this year's Nobel Prizes are announced, so it is a good moment to tell her story, because in two thousand nine she won one, and because the way she won it is more interesting than the way it usually gets told.此刻正是今年诺贝尔奖公布前的一周，因此正适合讲讲她的故事——因为她在 2009 年获得了诺贝尔奖，也因为她获奖的方式比通常讲述的更有意思。

Start with what the ribosome is, because the whole story hangs on it.先从核糖体是什么讲起，因为整个故事都系于它。 Every living cell reads its genes and turns them into proteins, and the machine that does the turning is the ribosome.每个活细胞都会读取自己的基因并将其转化为蛋白质，而完成这一转化的机器就是核糖体。 It grabs a strand of messenger RNA, reads it three letters at a time, and threads out a chain of amino acids that folds into a working protein.它抓住一条信使 RNA，每次读取三个字母，再串出一条氨基酸链，这条链折叠成一个能发挥功能的蛋白质。 It is the factory floor of life. There are thousands of them in a single bacterium, millions in one of your cells.它是生命的生产车间。单个细菌里就有数千个，你的一个细胞里则有数百万个。 Nothing alive builds a protein without one. And until the very end of the twentieth century, nobody had seen one clearly.没有它，任何生命都造不出蛋白质。而直到二十世纪的最末尾，还没有人清楚地看见过它。 We knew roughly what it did. We did not know what it looked like, atom by atom, and without that you cannot really say how it works.我们大致知道它做什么。但我们不知道它逐个原子看上去是什么样，而没有这一点，你就无法真正说清它是怎么工作的。

To see something atom by atom, the tool of the century was X-ray crystallography. Here is the idea in one image.要逐个原子地看清一样东西，那个世纪的工具是 X 射线晶体学。用一幅图来说明这个原理。 A single molecule scatters X-rays far too faintly to detect.单个分子对 X 射线的散射太过微弱，根本无法探测。 But if you can persuade trillions of identical copies to line up in a perfectly repeating lattice, a crystal, their faint scattering adds up, and the beam comes out the other side as a pattern of bright spots.但如果你能让数万亿个完全相同的副本排列成一个完美重复的晶格——也就是晶体——它们微弱的散射就会叠加起来，射束从另一侧出来时便形成一幅由亮点组成的图样。 From that pattern you can, with a great deal of work, reconstruct where every atom sat.根据那幅图样，你可以——付出大量工作之后——重建出每个原子所处的位置。 It had been done for salt, for DNA, for small proteins. The ribosome was another matter.这在盐上、在 DNA 上、在小蛋白质上都已经做到过。核糖体则是另一回事。 It is enormous, a machine of hundreds of thousands of atoms in two loosely joined halves, and it is floppy.它极其庞大，是一台由数十万个原子构成、分为松散连接的两半的机器，而且软塌塌的。 It does not want to sit still in a neat grid. Coax it into a crystal and it slumps.它不愿意乖乖待在整齐的格子里。哄它进入晶体，它就瘫软下来。 That was the wall everyone hit, and it is why the word people used was impossible.这就是每个人都撞上的那堵墙，也是为什么人们用的那个词是「不可能」。

Now here is the first thing worth slowing down for, because it is the heart of the episode.接下来是第一件值得放慢来讲的事，因为它是本期节目的核心。 When a whole field tells you something is impossible, listen carefully to what they actually mean.当整个领域都告诉你某件事不可能时，要仔细听清他们实际指的是什么。 There are two very different claims hiding in that one word. One is that the object itself cannot be done.那一个词里藏着两个非常不同的说法。一个是，这个对象本身做不到。 The other is that the current method cannot do it. Those sound alike and they are not alike at all.另一个是，当前的方法做不到。这两者听起来相似，实则完全不同。 The ribosome was not refusing to crystallize because of some law of nature. It was refusing because of how everyone was preparing it.核糖体不肯结晶，并不是因为什么自然定律，而是因为大家制备它的方式。 And Yonath's genius was to treat impossible as a statement about the method, and then go change the method. She changed it in two ways.而 Yonath 的过人之处在于，她把「不可能」当作一个关于方法的论断，然后去改变方法。她从两个方面做了改变。

The first came, of all places, from bears.第一个改变的来处出人意料——来自熊。 Recovering from a bicycle accident, reading to pass the time, she came across an account of hibernating bears, and a detail stuck to her.在一次自行车事故后的康复期间，她靠阅读打发时间，读到了一段关于冬眠熊的描述，其中一个细节让她念念不忘。 Before a long hibernation, cells pack their ribosomes away in orderly, tightly stacked arrays, where they sit intact for months and then start working again. 在长时间冬眠之前，细胞会把自己的核糖体收拢成有序、紧密堆叠的阵列，它们在那里完好无损地待上数月，然后重新开始工作。 That was the clue. Ribosomes, under the right pressure, will pack themselves into order. Order is exactly what a crystal is.这就是线索。在合适的压力下，核糖体会把自己排成有序的结构。而有序正是晶体的本质。 So the question became, which ribosomes are the toughest, the most willing to hold their shape.于是问题变成了：哪种核糖体最坚韧、最愿意保持自己的形状。 Her answer was to go to the extremes of the living world. She took ribosomes from bacteria that live in punishing places.她的答案是走向生命世界的极端。她从生活在严酷之地的细菌中提取核糖体。 From hot springs near boiling. From the brine of the Dead Sea.取自接近沸腾的温泉，取自死海的卤水。 These organisms had been shaped by evolution to keep their machinery rigid under heat and salt that would tear an ordinary cell apart.这些生物经过进化塑造，能在足以撕裂普通细胞的高温和高盐下让自身的分子机器保持刚性。 A ribosome built to survive a hot spring is also a ribosome that will hold still in a crystal.一个为在温泉中存活而生的核糖体，也是一个能在晶体中保持静止的核糖体。 In nineteen eighty, after some twenty-five thousand attempts, she got her first crystals of a ribosomal subunit. They were poor.1980 年，在约 2.5 万次尝试之后，她得到了第一批核糖体亚基的晶体。它们质量很差。 But they existed, and their existence alone refuted the word impossible. The second change solved a crueler problem.但它们确实存在，仅凭其存在就驳倒了“不可能”这个词。第二项改变解决了一个更残酷的问题。

The very X-ray beam you need to see the crystal also destroys it.你用来观察晶体所必需的那束 X 射线，同时也在摧毁它。 The radiation knocks the delicate molecules apart faster than you can read them.辐射把这些脆弱的分子击碎的速度，比你读取它们的速度还快。 So Yonath did something that now sounds obvious and at the time sounded strange.于是约纳特做了一件如今听来理所当然、在当时却显得古怪的事。 She froze the crystals, hard, down near minus one hundred and eighty degrees, and collected the data while they were deep-frozen.她把晶体冻起来，冻得很硬，降到接近零下 180 度，并在它们深度冷冻的状态下采集数据。 The cold slowed the damage enough to finish the measurement before the crystal died. She called it cryo bio-crystallography.低温让损伤放缓到足以在晶体“死亡”前完成测量。她把这称为低温生物晶体学（cryo bio-crystallography）。 And this is the part that outgrew her own project entirely, because flash-freezing crystals is now simply how protein crystallography is done, everywhere, for almost everything.而这正是彻底超出她自己项目范畴的部分，因为快速冷冻晶体如今就是蛋白质晶体学的通行做法，无处不在，几乎适用于一切。 A technique she developed to catch one impossible molecule became the ordinary daily practice of an entire science.她为捕捉一个“不可能”的分子而开发的技术，成了整门学科的日常寻常操作。 That, quietly, may be the larger legacy.悄然之间，这或许才是更大的遗产。

Now I want to pull on a thread that runs straight into work some of you do, because a crystal structure is not a photograph.现在我想扯出一条直接关联到你们中一些人工作的线索，因为晶体结构并不是一张照片。 You never see the molecule.你从来看不到那个分子。 What the detector records is the pattern of spots, and for each spot it faithfully measures one thing, how bright it is.探测器记录下的是斑点的图样，而对每一个斑点，它忠实地测量一件事：它有多亮。 But a wave has two properties, its brightness and its phase, its timing, and the phase is exactly what the detector cannot record.但波有两个属性：它的亮度和它的相位，也就是它的时序，而相位恰恰是探测器无法记录的。 It is thrown away at the moment of measurement. To rebuild the structure you need both.它在测量的那一刻就被丢弃了。而要重建结构，两者你都需要。 So you are handed half the information and asked to recover the whole.于是你只拿到了一半信息，却被要求恢复出全部。 This has a name, the phase problem, and it is why crystallography is hard even after you have your crystal. It is an inverse problem.这有一个名字，叫相位问题（phase problem），这也是为什么即便你已经有了晶体，晶体学依然困难。它是一个反问题（inverse problem）。 You measure the output and must run the machine backwards to the thing that produced it, and the backward run is missing a piece.你测量的是输出，却必须把这台机器倒着运行，回推到产生输出的那个东西，而这一次倒推缺了一块。 For a molecule the size of the ribosome the phase problem is monstrous, and solving it took extra experiments and years of labor.对于核糖体这么大的分子，相位问题极其庞大，解决它耗费了额外的实验和数年的辛劳。 Keep that in mind whenever a structure is described as if someone simply looked.每当有人把一个结构描述得仿佛只是看一眼就得到了，请记住这一点。 It is a reconstruction, extraordinarily well constrained at high resolution, but a reconstruction, fit to data through an inverse problem that had to be solved before any atom could be placed.它是一次重建，在高分辨率下受到极其严格的约束，但终究是一次重建，是通过一个必须先解决、然后才能摆放任何一个原子的反问题去拟合数据得到的。

By two thousand and two thousand one the crystals had gone from poor to superb, and the structures came out at atomic resolution.到 2000 年和 2001 年，晶体已从质量低劣变得极为出色，结构以原子分辨率呈现出来。 Yonath shared the two thousand nine Nobel Prize in Chemistry with Venkatraman Ramakrishnan and Thomas Steitz.约纳特与文卡特拉曼·拉马克里希南（Venkatraman Ramakrishnan）和托马斯·施泰茨（Thomas Steitz）分享了 2009 年诺贝尔化学奖。 And here is where the payoff lands, the reason this was worth twenty years.回报就落在这里，这也是这一切值得花上二十年的原因。 Inside the ribosome there is a tunnel, the passage through which the freshly made protein snakes out. Yonath was the first to see it.核糖体内部有一条隧道，新合成的蛋白质就是从这条通道里蜿蜒而出。约纳特是第一个看到它的人。 That tunnel is where a large share of our antibiotics do their work.我们很大一部分抗生素正是在那条隧道里发挥作用的。 More than half of the antibiotics in a doctor's cabinet fight bacteria by jamming their ribosomes, and many of them, the ones related to erythromycin, plug that exit tunnel so the bacterium cannot finish building its proteins.医生药柜里超过半数的抗生素，都是靠卡住细菌的核糖体来对抗细菌的；其中很多——那些与红霉素相关的——会堵住那条出口隧道，让细菌无法完成蛋白质的组装。 Once you can see the tunnel atom by atom, you can see exactly how the drug sits in it, why a single mutation in the tunnel wall lets a resistant bug shrug the drug off, and how a human ribosome differs from a bacterial one so that the drug hits the germ and not you.一旦你能逐个原子地看清这条隧道，你就能精确地看到药物是怎样嵌在里面的，为什么隧道壁上的一个突变就能让耐药菌把药物甩开，以及人类的核糖体与细菌的核糖体有何不同，从而让药物只打击病菌而不伤到你。 That is not abstract. That is the difference between a working antibiotic and a useless one.这并不抽象。这正是一种有效的抗生素和一种无用的抗生素之间的区别。

So let me correct the way this story usually gets told, because the popular version does her a disservice.所以，让我来纠正一下这个故事通常的讲法，因为流行的版本其实对她不公。 The popular version is the lone woman, doubted by a sexist establishment, vindicated in the end by sheer stubbornness.流行的版本是：一个孤身奋战的女性，被充满性别歧视的建制派所质疑，最终仅凭一股倔劲得以正名。 Stubbornness she certainly had. But that framing is both unfair and, oddly, too small.倔劲她当然是有的。但这种叙述框架既不公平，而且说来奇怪，也把她说得太小了。 It was unfair to call it stubbornness, because what she actually brought were two specific, transferable ideas, the tough ribosome and the frozen crystal, not just refusal to quit.把它称作倔强是不公平的，因为她真正带来的是两个具体的、可迁移的思路——耐受的核糖体和冷冻的晶体——而不仅仅是拒绝放弃。 And the doubt was not mostly stupidity or malice. The problem really was brutally hard, and the doubters had the failures to prove it.而那些质疑也大多并非源于愚蠢或恶意。这个问题确实难得残酷，质疑者们手里也有一连串失败可以为证。 She did not out-argue them. She changed the preparation, and the preparation is what had been wrong.她并不是靠辩赢了他们。她改变了样品的制备方法，而此前出错的正是制备方法。 And it is too small a story because the reality is bigger. The Nobel was shared three ways.而说它是个太小的故事，是因为现实要更大。这个诺贝尔奖是三人分享的。 Getting from her crude nineteen eighty crystals to a finished atomic structure took two decades and many laboratories.从她 1980 年那些粗糙的晶体，到一个完成的原子级结构，花了二十年、经手了许多实验室。 The honest credit, the crystallization breakthrough and the freezing method that the whole field now lives by, is more impressive than the myth of a single obstinate genius, not less.而如实归功于她的东西——结晶上的突破，以及如今整个领域都赖以为生的冷冻方法——比那个孤身固执天才的神话更令人钦佩，而非更逊色。

She spent her last years arguing that the ribosome is a fossil, that its all-RNA core is a relic of a time before proteins, a clue to how life began.她把生命最后的几年用来论证：核糖体是一块化石，它那全由 RNA 构成的核心是蛋白质出现之前那个时代的遗迹，是生命如何起源的一条线索。 That is a hypothesis, a good one, not a settled fact, and she would have wanted it labeled that way.那是一个假说，一个不错的假说，而不是已成定论的事实，她本人也会希望它被如此标明。 But the sentence to carry out of this is the one she lived by.但从这一切中该带走的那句话，是她奉为信条的那一句。 When everyone tells you a thing is impossible, find out whether they are talking about the thing, or about their method.当所有人都告诉你某件事不可能时，去弄清楚他们说的究竟是那件事本身，还是他们的方法。 Ada Yonath bet her career that it was the method. She was right, and she froze a bear's trick into a crystal to prove it.阿达·约纳特赌上自己的职业生涯，赌问题出在方法上。她赌对了，而且她把一只熊的诀窍冻进了晶体里，以此加以证明。

Here is a puzzle that bothered astronomers for years. A supermassive black hole tears a star apart.这里有一个困扰了天文学家多年的谜题。一个超大质量黑洞把一颗恒星撕碎。 Sometimes, within days, it fires a jet of radio light out into space. Other times it does nothing.有时，几天之内，它就朝太空喷出一道射电光的喷流。另一些时候，它却什么都不做。 It shreds the star, and then it just sits there, dark and quiet, for months.它撕碎了那颗恒星，然后就那样静静待着，黑暗而沉寂，一待就是好几个月。 And then, with no new star and no new meal, it suddenly lights up and throws a jet after all.接着，没有新的恒星，也没有新的食物，它却突然亮了起来，终究还是抛出了一道喷流。 Same kind of black hole, same kind of event, completely different behavior, and a long, strange delay in the second case that nobody could explain.同一种黑洞，同一类事件，行为却截然不同，而在第二种情形里还有一段漫长而古怪、无人能解释的延迟。

This month a study in Nature Astronomy, led by Adelle Goodwin in Perth and Andrew Mummery in Princeton, proposes an answer.本月发表在《自然·天文学》上的一项研究，由珀斯的 Adelle Goodwin 和普林斯顿的 Andrew Mummery 领衔，提出了一个答案。 And the answer is lovely, because it turns out the delayed jet is not late. It is right on time.而这个答案很美妙，因为事实证明，延迟的喷流并不是迟到。它来得正是时候。 It is just waiting for a signal that only arrives once the feast is over.它只是在等一个信号，而这个信号只有等到盛宴结束才会到来。

Let me build the picture slowly, because the popular version gets the timing backwards. Start with the cartoon most of us carry.让我慢慢把这幅图景搭建起来，因为通俗的版本把时间顺序弄反了。先从我们大多数人脑中那幅漫画式的图景说起。

A black hole is a drain. A star falls in, the drain gulps it down, and out the other side comes a burp, a jet. Eat, then belch.黑洞是个下水道。一颗恒星落进去，下水道把它一口吞下，然后从另一头冒出一个嗝，一道喷流。先吃，后打嗝。 In that picture the jet should fire when the eating is most violent.在那幅图景里，喷流应当在进食最猛烈的时候喷发。 Hold onto that, because for the delayed jets it is exactly what the data does not show. First, some anchors.记住这一点，因为对于那些延迟的喷流来说，数据显示的恰恰不是这样。首先，先立几个基准。

When a star is torn apart, it does not vanish neatly. Roughly half of it is flung away.当一颗恒星被撕碎时，它并不会干净利落地消失。大约一半会被抛飞出去。 The rest spreads into a hot, swirling disk and spirals inward. That inward spiral is the feeding.剩下的部分铺展成一个炽热、旋转的盘，并螺旋着向内落。那向内的螺旋就是进食。 And there is a natural speed limit on how fast anything can feed. As matter falls in it glows, and light pushes back.而任何东西的进食速度都有一个天然的上限。物质向内落时会发光，而光会往回推。 Turn the feeding up high enough and the outward push of that light balances the inward pull of gravity.把进食调得足够猛，这道光的向外推力就会与引力的向内拉力相平衡。 That balance point is called the Eddington limit. Think of it as the black hole's top eating speed.这个平衡点被称为爱丁顿极限（Eddington limit）。可以把它想成黑洞的最高进食速度。

The important thing about that limit is that it scales with size.关于这个极限，重要的一点是它随尺寸而变。 A big black hole has a high limit, a small one a low limit, in simple proportion to its mass.大黑洞极限高，小黑洞极限低，简单地正比于它的质量。 So if you want to compare a black hole ten times the mass of the sun with one a million times heavier, you do not compare their raw feeding rates, which differ enormously.所以，如果你想把一个十倍太阳质量的黑洞和一个百万倍太阳质量的黑洞作比较，你不会去比它们各自原始的进食率——那差别大得惊人。 You compare each one to its own limit. You ask what fraction of full throttle it is running at. That fraction is a pure ratio.你要把每一个都拿来和它自己的极限相比。你要问它跑在满油门的百分之几上。这个比例是个纯粹的比值。 It does not care how big the black hole is. And that is the hinge of the whole story.它并不在乎黑洞有多大。而这正是整个故事的枢纽。

Now, the part that is not new, and here the headlines oversell a little.接下来是那个并不新鲜的部分，在这里那些标题党稍稍夸大了一点。 For decades we have watched small black holes in our own galaxy, ones only a few times the mass of the sun, sitting in binary systems where they slowly strip gas off a companion star.几十年来，我们一直在观测我们自己星系里的小黑洞，那些只有几倍太阳质量的黑洞，它们处在双星系统里，缓慢地从一颗伴星身上剥离气体。 These flare up and fade over months, and as they do, they run through distinct states.这些黑洞在几个月里爆发又消退，而在这个过程中，它们会经历一系列不同的状态。 When they are feeding hard, near their limit, the steady jet actually shuts off. Quiet.当它们猛烈进食、逼近自己的极限时，那道稳定的喷流反而会关闭。安静下来。 Then as the feeding fades back down, through a few percent of the limit, the jet switches on again.然后，随着进食减弱回落，穿过极限的百分之几时，喷流又重新打开。 Astronomers have mapped this again and again. The jet is tied not to the peak of the meal but to a transition on the way down.天文学家一次又一次地把这个过程标绘出来。喷流并不与进食的峰值绑定，而是与回落途中的一次转变绑定。

And back in 2003 and 2004, two teams noticed something bigger.而早在 2003 和 2004 年，两个团队注意到了某种更宏大的东西。 If you plot black holes by their radio output, their x-ray output, and their mass, the small ones and the giant ones fall on the same relationship.如果你把黑洞按它们的射电输出、X 射线输出和质量画出来，那些小黑洞和那些巨型黑洞落在同一条关系上。 They called it the fundamental plane of black hole activity.他们把它称为黑洞活动的基本面（fundamental plane）。 The message was that the physics of feeding and jet-launching seems to be scale-invariant, the same recipe from small to enormous, just stretched across size.其含义是，吸积（feeding）和喷流启动（jet-launching）背后的物理似乎是尺度不变的（scale-invariant）——从小到极大都遵循同一套配方，只是在尺寸上被拉伸。 So the idea that all black holes follow one rule is not the discovery here. It is more than twenty years old. Here is what is genuinely new.所以，所有黑洞都遵循同一条规律，这个想法并不是这里的新发现。它已经有二十多年历史了。下面才是真正新的东西。

Testing that idea at the giant end used to be impossible, because a supermassive black hole runs its cycle in slow motion.过去，在巨型这一端检验这个想法几乎不可能，因为一个超大质量黑洞是以慢动作在走完它的循环。 What a small black hole does in a few months, a giant one might take far longer than a human lifetime to do.一个小黑洞几个月里做完的事，一个巨型黑洞可能要花远超一个人一生的时间才能做完。 You cannot sit and watch it change state. But a shredded star is a stopwatch.你没法坐下来看着它改变状态。但一颗被撕碎的恒星就是一只秒表。 It dumps a burst of fuel all at once, and then the feeding fades on its own, and for a supermassive black hole that fade plays out over just a few years.它一次性倾泻出一阵燃料，然后吸积自行衰减，而对一个超大质量黑洞来说，这段衰减只在短短几年里演完。 Fast enough to watch.快到足以观测。 So a tidal disruption event lets you catch a million-solar-mass black hole running the same cycle its tiny cousins run, in real time.所以一次潮汐撕裂事件（tidal disruption event）让你能实时捕捉到一个百万太阳质量的黑洞，走完它那些微小同类所走的同一个循环。

Goodwin and Mummery gathered twenty of these events, watched across optical, ultraviolet, x-ray and radio light.Goodwin 和 Mummery 收集了二十次这样的事件，横跨光学、紫外、X 射线和射电波段进行观测。 Ten of them were clean enough to model both the feeding rate and the timing of the radio jets.其中十次干净到足以同时对吸积率和射电喷流的时序建模。 And in those ten, the delayed jets, the puzzling ones that switch on months to years after the star is gone, switch on just as the feeding rate falls through about two percent of the Eddington limit.而在这十次里，那些延迟喷流——就是令人困惑、在恒星消失之后数月到数年才启动的那些——恰好在吸积率跌穿约爱丁顿极限（Eddington limit）的百分之二时启动。 The same fraction, near enough, as the little black holes in our own galaxy. So look again at the cartoon.这个比例，大致相同，和我们自己星系里那些小黑洞是一样的。所以再看一眼那幅示意图。

There are really two separate jet moments here.这里其实有两个各自独立的喷流时刻。 There is a prompt one, right at the start, when the feeding is above the limit, over the top. That is the belch you expected. 有一个即时的（prompt），就在最开始、吸积高于极限、满溢的时候。那是你预料之中的那声饱嗝。 But the famous delayed jet, the one that made this a mystery, is the opposite. It does not fire at the feast.但那个著名的延迟喷流，就是让这件事成为谜题的那个，恰恰相反。它并不在盛宴时喷发。 It fires as the meal ends, at the moment the feeding rate sinks past that two percent mark on the way down.它是在这顿饭结束时喷发的——就在吸积率一路下降、跌过那个百分之二标记的那一刻。 The beacon lights at low tide, not high tide. The delay was never the black hole being slow.灯塔亮起在退潮，而非涨潮。这个延迟从来不是因为黑洞慢。 It was the black hole waiting for the flow to drop to the exact level that flips the switch.而是黑洞在等着流量掉到那个恰好能扳动开关的确切水平。

Now the careful part, because this is where you should hold the claim at arm's length. Nobody saw a jet launch.现在到了要谨慎的部分，因为这正是你该对这个结论保持一臂之距的地方。没有人看到过一次喷流启动。 Nobody has a gauge reading the feeding rate.没有人有一只测吸积率的仪表。 What you actually have is a radio flare that arrives months or years after the star was destroyed.你实际拥有的，是一次在恒星被摧毁之后数月或数年才到达的射电耀发。 From that, using a model of how the disk drains, you work backward to when the jet must have fired and how hard the hole was feeding at that moment.从这一点出发，借助一个描述吸积盘如何排空的模型，你反推出喷流必定是何时喷发的、以及那一刻黑洞吸积有多猛。 It is an inverse problem.这是一个反问题（inverse problem）。 You infer the state of an engine you cannot see from a delayed echo, and the answer leans on the model you chose for the disk.你从一个延迟的回声去推断一台你看不见的引擎的状态，而答案取决于你为吸积盘所选的那个模型。 A different disk model would move the two percent somewhat. So treat two percent as a good central estimate, not a measured constant.换一个吸积盘模型，会把这个百分之二挪动一些。所以把百分之二当成一个不错的中心估计，而非一个实测常数。

And the sample is ten.而样本只有十次。 Ten events, bridging a factor of a million in mass between these giants and the galactic black holes, is enough to draw a striking line.十次事件，在这些巨型黑洞和银河系里的黑洞之间跨越了百万倍的质量差，足以画出一条引人注目的线。 It is not enough to carve a law. The claim is also narrower than the headline.但还不足以刻出一条定律。这个结论也比标题所说的更窄。 It tells you when a jet fires, the trigger, the critical fraction.它告诉你的是喷流何时喷发——那个触发条件，那个临界比例。 It does not tell you how the jet is made, what actually launches it, or why that particular threshold and not another.它并没有告诉你喷流是如何形成的、究竟是什么把它发射出去的，或者为什么偏偏是那个特定的阈值而不是别的。 The engine is still hidden. And a shared number is suggestive of shared physics, but it is not proof of it.引擎依然是隐藏的。一个共同的数字暗示着共同的物理机制，但这并不是它的证明。 Two systems can hit the same mark for reasons that are not the same underneath. What survives all of that is still worth the ten minutes.两个系统可能出于底层并不相同的原因命中同一个标记。但即便如此，最终留存下来的东西仍然值得这十分钟。

The durable result is not a slogan about all black holes obeying one law.这个经得起时间检验的结果，并不是一句关于所有黑洞都遵循同一条定律的口号。 It is a clean demonstration that a single dimensionless ratio, feeding measured as a fraction of the maximum, seems to govern the same behavior across a factor of a million in size and a factor of a million in time.它是一个干净利落的演示：一个无量纲的比值——以最大值的比例来衡量的吸积速率——似乎在跨越百万倍的尺度和百万倍的时间跨度上，支配着同样的行为。 That is the kind of collapse scientists love, where wildly different objects fall onto one curve once you measure them in the right units.这正是科学家钟爱的那种坍缩现象：一旦你用正确的单位去衡量，千差万别的天体都会落到同一条曲线上。 It is the same move that lets an engineer test a small model in a water tank and trust it for a full-size ship, because what matters is a ratio, not the raw scale.这与工程师在水槽里测试小比例模型、然后相信它对全尺寸船舶同样适用是同一套做法，因为关键在于一个比值，而不是原始的尺度大小。

And the shape of the reasoning should feel familiar to anyone who works with hidden systems. You never see the thing itself.而这种推理的形态，对任何研究隐藏系统的人来说都应该感到熟悉。你从来看不到事物本身。 You see a delayed, smeared signal, and you run a model backward to recover what the machine was doing when it sent it.你看到的是一个延迟、模糊的信号，于是你反向运行一个模型，去还原这台机器在发出信号时正在做什么。 The honest worry is always the same one. Could a different model of the middle give you the same signal at the end?诚实的担忧始终是同一个：给中间环节换一个不同的模型，会不会在最后给出同样的信号？ Until more events fill in the long gap between the small black holes and the giant ones, two percent is a beautiful, provisional number.在更多的事件填补小质量黑洞与巨型黑洞之间那道漫长的空白之前，2% 是一个漂亮而暂定的数字。 A stopwatch made from a dying star, read carefully, and not oversold.一只用垂死恒星做成的秒表，被仔细地读取，而没有被过度吹嘘。

This week the headlines said something startling about the thing you are using to listen to me.本周的新闻标题，说了一件关于你正用来听我说话的那个东西的惊人之事。 They said the human brain is not one organ but two. Two separate brains, fused by evolution, sharing a skull.它们说，人脑不是一个器官，而是两个。两个独立的大脑，被进化融合在一起，共用一个头骨。 One does poetry and mathematics. The other keeps your heart beating.一个负责诗歌与数学，另一个让你的心脏持续跳动。 A team at Stanford, publishing in Nature Neuroscience on the eighteenth of September, was said to have rewritten the textbook.据说，斯坦福的一个团队在 9 月 18 日发表于《Nature Neuroscience》的论文中，改写了教科书。

It is a wonderful story. Most of it is even true.这是个精彩的故事，其中大部分甚至是真的。 But the interesting part is not the part that made the headline, and the part that made the headline is dressed in clothes two sizes too big.但有趣的部分并不是登上头条的那部分，而登上头条的那部分，穿着一件大了两号的外衣。 So let me do two things tonight. First, tell you plainly what these scientists actually found, because it is genuinely surprising.所以今晚我想做两件事。第一，坦白地告诉你这些科学家究竟发现了什么，因为它确实令人惊讶。 Then separate what they showed from what they, and the coverage, went on to claim. The gap between those two is the whole lesson.然后，把他们真正证明的东西，和他们以及媒体报道随后所声称的东西区分开来。这两者之间的差距，正是全部的教训所在。

Start with the embryo.先从胚胎说起。 Very early on, days into development, there is a flat sheet of cells that will become the entire brain and spinal cord.在非常早期，发育才进行到几天时，有一层扁平的细胞，它们将来会发育成整个大脑和脊髓。 The old picture, the one in the textbooks, is that this sheet begins as one continuous field. A single population of founder cells.旧的图景，也就是教科书里的那个，认为这层细胞一开始是一个连续的区域，一个单一的祖细胞群体。 And then, later, the field gets divided up.然后，在稍后，这个区域才被划分开来。 The front is told to become forebrain, the back is told to become hindbrain, and a sharp border is drawn between them somewhere in the middle.前部被指示发育成前脑，后部被指示发育成后脑，并在两者之间的某处画出一条清晰的边界。

What the Stanford group reports is that the division is not drawn later. It is there from the very beginning.斯坦福团队报告的是，这条分界线并不是后来才画出的，而是从一开始就存在的。 When they looked at the earliest moment they could see clearly, in mouse embryos, there were already two separate groups of founder cells.当他们在小鼠胚胎中观察能够清晰看到的最早时刻时，那里已经存在两群独立的祖细胞。 Not one field waiting to be split. Two crowds, standing front and back, that never mingle.不是一个等待被分开的区域，而是两群细胞，分立前后，从不混合。 One crowd carries a molecular flag, a gene switched on, called Otx2, and that crowd builds the front of the brain, the forebrain and the midbrain.其中一群带着一面分子旗帜——一个被开启的基因，叫做 Otx2——这群细胞构建大脑的前部，即前脑和中脑。 The other crowd carries a different flag, called Gbx2, and it builds the hindbrain, which becomes your brainstem and your cerebellum.另一群带着一面不同的旗帜，叫做 Gbx2，它构建后脑，也就是将来的脑干和小脑。 The two crowds sit side by side and do not trade members. Parallel tracks that run the whole way without crossing.两群细胞并肩而立，彼此不交换成员。就像两条平行的轨道，一路延伸而从不交叉。

And here is the mechanism that makes it stick. It is not just that the two groups have different flags flying.而这里就是让这一切固定下来的机制。不仅仅是两群细胞升起了不同的旗帜。 It is that inside each cell the genome itself is packed differently.而是在每个细胞内部，基因组本身的包装方式就不一样。 Think of the genome as a thick instruction manual, and imagine that in one group of cells a whole set of pages is glued shut, while in the other group a different set is glued shut.把基因组想象成一本厚厚的说明书，再想象在一群细胞里有一整套书页被粘死了，而在另一群细胞里被粘死的是另外一套书页。 The pages a cell can read decide what it can become.一个细胞能读到哪些书页，决定了它能变成什么。 Because the front cells and the back cells have different pages sealed off, a front cell physically cannot follow the back instructions even if you push it.由于前部细胞和后部细胞被封住的书页不同，即使你去逼迫一个前部细胞，它在物理上也无法遵循后部的指令。 The commitment is locked in the packaging. That is what the scientists mean when they say the two lineages are set apart from the start.这种命运的锁定就藏在包装方式里。这就是科学家所说的「这两个谱系从一开始就被分隔开来」的含义。

Then they did the thing that makes it feel deep. They looked in other animals. Chickens. Zebrafish.接着他们做了那件让这项发现显得意味深长的事。他们去观察其他动物。鸡。斑马鱼。 And an acorn worm, a soft marine creature that lives in the seabed and has no brain at all.还有柱头虫，一种生活在海底、根本没有大脑的柔软海洋生物。 In every one of them, the same two founder populations, the same front flag and back flag, sitting apart.在它们每一个身上，都有同样的两群祖细胞，同样的前部旗帜和后部旗帜，彼此分立。 That pattern stretches back across more than five hundred million years of evolution.这一模式可以追溯到五亿多年的进化历程之前。 The split between the front and the back of the brain appears to be older than brains. That is the real finding, and it is lovely.大脑前部与后部之间的分野，似乎比大脑本身还要古老。这才是真正的发现，而它无比美妙。

Now let me take the oversized clothes off it, in two steps. The first correction is about novelty.现在让我分两步，把套在它身上的那件过大的外衣脱下来。第一处修正是关于新颖性的。

If you read that a textbook was torn up, you might think nobody knew there was a front and a back before this. That is not right.如果你读到某本教科书被撕掉了，你可能会以为在此之前没人知道胚胎有前后之分。这是不对的。 The line between the Otx2 territory and the Gbx2 territory is one of the most studied boundaries in all of developmental biology.Otx2 区域与 Gbx2 区域之间的这条界线，是整个发育生物学中被研究得最多的边界之一。 It has a name, the midbrain-hindbrain boundary, and papers describing it fill the late nineteen-nineties.它有一个名字，叫中脑-后脑边界，描述它的论文在二十世纪九十年代后期比比皆是。 The classic account is that the two flags meet in the middle and repress each other, each shoving the other back, and that pushing match carves a clean border in the sheet.经典的说法是，两面旗帜在中间相遇并相互抑制，各自把对方往回推，而这场推挤较量在这层细胞片上刻出了一条清晰的边界。 So the existence of a front-back divide is old news. What is new, and it is a real change, is the claim about origin.所以前后之分的存在并不是什么新闻。真正新的、也确实是一个实质性的改变，是关于起源的那个论断。 The old model had one field that a border was drawn across. The new one says there was never a single field.旧模型认为存在一个区域，边界从中间划过。新模型则说从来就不存在一个单一的区域。 Two lineages, separate all along.两个谱系，从一开始就是分开的。 That is a genuine upgrade, but it is a subtle one, and subtle is the opposite of what the word rewrite suggests. 这是一次货真价实的升级，但它是微妙的，而微妙恰恰与"改写"这个词所暗示的意思相反。

The second correction is the bigger one, and it is about what the two halves mean.第二处修正更重要，它关乎这两半究竟意味着什么。 The phrase two separate organs, the phrase two brains, and the tidy line that the front does poetry while the back does heartbeat, all point you toward a picture that the science does not support."两个独立的器官"这种说法，"两个大脑"这种说法，以及那条整齐的界线——说前面负责诗歌、后面负责心跳——全都把你引向一幅科学并不支持的图景。 You do not have two brains. You have one brain, wired together seamlessly, working as a single instrument every waking second.你没有两个大脑。你只有一个大脑，天衣无缝地连成一体，在你清醒的每一秒都作为单一的整体在运转。 The finding is about where the building cells come from, not about two minds sharing your head.这项发现讲的是构筑组织的细胞来自哪里，而不是两个心智共用你的脑袋。 This is an embryology result wearing a philosophy costume. And the neat division of labor is a caricature.这是一个披着哲学外衣的胚胎学结果。而那种整齐的分工是一幅漫画式的夸张。

Watch where the line actually falls. It falls at the midbrain-hindbrain junction.看看这条界线实际落在哪里。它落在中脑-后脑的交界处。 Which means the midbrain, a structure that sits low, down in what we loosely call the brainstem, is on the front team, the Otx2 team, together with the forebrain.这意味着中脑——一个位置很低、处在我们笼统称为脑干那部分中的结构——属于前面这一队，也就是 Otx2 队，和前脑站在一起。 And the cerebellum, on the back team, is not some dumb reflex engine.而小脑属于后面这一队，它可不是什么愚钝的反射引擎。 It does some of the most exquisite timing and coordination in the whole nervous system.它承担着整个神经系统中一些最精妙的计时与协调工作。 So the folk idea of a clever brain sitting on top of a primitive survival brain is not this boundary.所以那种民间观念——一个聪明的大脑坐在一个原始的求生大脑之上——并不是这条边界所讲的事。 The real divide is a developmental one, defined by which founder cells built the tissue, and it simply does not line up with the story about higher thought versus base function.真正的分界是一条发育上的分界，由哪些始祖细胞构筑了组织来定义，它跟高级思维对阵基础功能的那套故事根本对不上。 When a result refuses to fall along the line your intuition wants, that is usually the most honest thing about it.当一个结果拒绝落在你的直觉所期望的那条线上时，这通常正是它身上最诚实的地方。

Now the claim that will actually be argued over for years.现在来看那个真正会被争论好多年的论断。 The scientists say evolution took two existing neural systems and pushed them together. Two ancient nervous systems, fused.科学家们说，进化取用了两套已有的神经系统，把它们推到了一起。两套古老的神经系统，融合而成。 Here you have to be careful, because this is no longer something they measured.在这里你得小心，因为这已经不是他们测量到的东西了。 It is a story told about deep time, inferred from a pattern seen in living animals today.这是一个关于深远时间的故事，从今天活着的动物身上看到的一种模式推断出来的。 They found the same two-part signature in a worm and a fish and a mouse, and from that shared signature they reason backward to a single ancient event, a fusion.他们在一条蠕虫、一条鱼和一只老鼠身上发现了同样的两部分特征，并由这一共有的特征反推出一个单一的古老事件，一次融合。 But you cannot watch five hundred million years happen. And a pattern you see in the present can be the end point of more than one history.但你没法亲眼看着五亿年发生。而你在当下看到的一种模式，可以是不止一种历史的终点。 The same two-lineage arrangement could reflect one old fusion, or it could reflect a single ancient patterning rule that every animal simply kept and elaborated in parallel.同样的双谱系格局，可能反映的是一次古老的融合，也可能反映的是一条单一的古老模式形成规则——每种动物只是把它保留下来，并各自平行地加以精细化。 The data underdetermine the story.数据不足以唯一地确定这个故事。 This is the same trap that haunts anyone who tries to read history off a present-day pattern, in any field.这跟任何领域中试图从当下的模式去读出历史的人所面临的陷阱，是同一个陷阱。 What you can see does not pin down how it got that way.你能看到的东西，并不能确定它是如何变成这样的。 So take fusion as a strong, suggestive hypothesis, the kind of thing that gets debated at conferences, not as a demonstrated fact.所以，把融合当作一个有力的、启发性的假说，那种会在学术会议上被拿来争论的东西，而不是一个已经被证实的事实。

So if the philosophy is oversold and the fusion is unproven, what is the discovery actually good for?那么，如果这套理念被过度吹捧、融合又未经证实，这项发现究竟有什么用？ Here is the part that got the least attention and deserves the most.接下来这部分受到的关注最少，却最值得关注。 For years, scientists learning to grow brain cells from stem cells could reliably make forebrain neurons, the front kind, but struggled badly to make working hindbrain and brainstem neurons, the back kind.多年来，学着用干细胞培养脑细胞的科学家能够稳定地造出前脑神经元，也就是靠前的那一类，却在造出有功能的后脑和脑干神经元、也就是靠后的那一类上举步维艰。 Nobody quite knew why the recipes kept failing. This finding explains it.没人真正清楚这些配方为什么总是失败。这项发现解释了原因。 The recipes were all coaxing cells down the front lineage, and the back is not a branch off the front.这些配方都是在引导细胞走上靠前的那条谱系，而靠后的那部分并不是从靠前那部分分出去的一个分支。 It is a separate lineage that needs its own starting instructions.它是一条独立的谱系，需要自己的初始指令。 Once the team treated the hindbrain as its own thing from the beginning, they grew functional human hindbrain motor neurons in a dish.一旦团队从一开始就把后脑当作独立的东西来对待，他们就在培养皿里培养出了有功能的人类后脑运动神经元。 Those are exactly the cells destroyed in A.L.S.那正是在 ALS 中被摧毁的那类细胞， and in spinal muscular atrophy, and until now there was no good human model to study them in. That is the news worth the headline.也是在脊髓性肌萎缩症中被摧毁的细胞，而在此之前一直没有好的人类模型来研究它们。这才是配得上标题的新闻。 Not that you are secretly two organs. That knowing where a piece comes from finally let us build the piece we could not build before.不是说你其实暗地里是两个器官，而是说：搞清楚一块组织从哪里来，终于让我们能造出以前造不出来的那块组织。

So here is the honest one-line version, the kind I would want if I were you.所以，这里给出一句实话版的概括，如果我是你，我也会想听这样的说法。 The front and the back of your brain are built by two different sets of founder cells, kept apart by the way their genomes are packed, and they may have been two all along.你大脑的前部和后部是由两套不同的起源细胞构建起来的，靠各自基因组的包装方式彼此分隔，而且它们也许从一开始就是两套。 It is still one brain. The value of knowing this is not that it makes you double. It is that it fixed a workbench.它仍然是一个大脑。知道这一点的价值不在于让你变成双份，而在于它修好了一张工作台。 A finding about origins got sold to the world as a finding about identity, and the quieter truth, the one that will still matter in ten years, is that it taught us how to grow a kind of human nerve cell that we badly need and could never make.一项关于起源的发现，被当作一项关于身份的发现兜售给了世界；而那个更安静的真相——十年后依然重要的那个——是它教会了我们如何培养一类我们急需、却从来造不出来的人类神经细胞。

When Gladys West died this past January, at ninety-five, the headlines all told a version of the same story. The woman who invented GPS. The mother of GPS. The hidden figure behind the little blue dot on your phone. It is a good story, and she had earned a good story, because for most of her working life almost nobody outside a Navy base in Virginia knew what she had done. But the headline is wrong. Not wrong in a cruel way. Wrong in a way that skips the most interesting part. So tonight I want to tell you what Gladys West actually did, because it is harder and stranger and more beautiful than inventing GPS, and once you have seen it you will not read the altitude number on your phone the same way again.

Let me start with a sentence that sounds like a riddle. Sea level is not level.

We say sea level as if it were the one flat thing we could count on. It is not. Imagine you could calm every wave, stop every tide, still every current and wind, and let the ocean settle into perfect stillness. You might expect a smooth ball of water. Instead the surface would have hills and valleys. Broad, gentle ones, but real. There is a great dip in the ocean surface south of India, where the water sits about a hundred meters lower than you would expect, and bulges elsewhere where it stands higher. A hundred meters, in an ocean just sitting there doing nothing. Why?

Because water is pulled by gravity, and gravity is not the same everywhere on Earth. Under the ground the planet is lumpy. Denser rock here, lighter rock there. Where the pull is a little stronger, water piles up toward it. Where the pull is a little weaker, the surface sags away. The still ocean is not a shape we impose on the Earth. It is a readout. It is the planet showing you its own gravity, drawn in water. Geodesists have a name for that imaginary still surface.

They call it the geoid. It is the true shape of the Earth in the only sense that matters for measuring position. Not the shape of the rock, the mountains and trenches. The shape of the gravity. And it is lumpy. If you took the geoid and stretched its bumps ten thousand times so you could see them, you would get a picture that scientists cheerfully call the potato. A dented, asymmetric potato. Now, this is the thing to hold onto.

There are three different Earths in this story, stacked on top of each other. The first is the sphere, the ball we draw in school. The second is a slightly squashed ball, wider at the equator than pole to pole, because the Earth spins. Geodesists call that smooth squashed ball the reference ellipsoid, and it is a pure mathematical idealization. It is nobody's real Earth. The third is the geoid, the gravity shape, which wanders above and below that smooth ellipsoid by as much as a hundred meters, in slow continental swells. And then on top sits the actual crust, the mountains and the sea floor. Three Earths, and navigation lives in the gaps between them.

Here is why that gap is not an academic curiosity. Think about how GPS is usually explained. You have satellites overhead with fantastically precise clocks. Your phone listens to several at once, measures how long each signal took to arrive, and from those tiny time differences works out where you are. All true. But that whole beautiful trick gives you a position in raw mathematical space, relative to that smooth idealized ellipsoid. And then your phone tells you: you are at thirty meters elevation. Thirty meters above what? Not above the center of the Earth. Not above the smooth ellipsoid, which is fictional. Above sea level. And to know where sea level is at your exact spot, the phone has to consult a stored model of the geoid, that lumpy potato, and correct for it. Without that model your altitude could be off by the height of a thirty story building, because the smooth ellipsoid and the real sea surface disagree by that much.

It goes deeper than altitude. The satellites themselves are falling through that same uneven gravity. To predict precisely where a GPS satellite will be a few hours from now, which is exactly what your position depends on, you have to know the Earth's gravity field in detail. The geoid and the gravity field are the same information wearing two hats. So the glamorous system, the clocks and the signals, is standing on top of something unglamorous and enormous. Somebody first had to measure the true shape of the Earth. And you do not get that for free. That is the work Gladys West did.

Not the clocks, not the signals, not the constellation. The shape underneath all of it.

How do you even measure a hundred-meter swell in the surface of the open ocean? You go up and look down with radar. In 1978 the Navy and NASA launched a satellite called Seasat, the first spacecraft built to study the oceans, and Gladys West was named its project manager. Seasat carried a radar altimeter. Think of it as a tape measure dropped from orbit: it fired a radar pulse straight down, timed the echo off the sea surface, and so measured the distance from the satellite to the water, over and over, millions of times, as it circled the Earth. And since the still sea surface traces the geoid, mapping the sea surface height is mapping the geoid.

Seasat itself lasted only a hundred and five days before a short circuit killed it. But in those hundred and five days it gathered more about the oceans than a century of ships had, and it proved the method. A Navy follow-on called Geosat, in the mid-nineteen-eighties, mapped the sea surface to a precision of a few centimeters, and much of that data stayed classified for years, because a precise map of ocean gravity quietly reveals the shape of the sea floor.

Now here is the part the word tape measure hides. The raw radar echoes are a mess. The sea is never still. There are waves and tides and currents and changes in air pressure pushing the water around, and the satellite's own orbit wobbles, and the pulse is slowed by moisture in the atmosphere. Every one of those has to be modeled and subtracted before the faint, permanent gravity shape underneath shows through. That is the actual labor. Turning an ocean of noisy echoes into one stable, precise surface. West did that on room-sized computers, including a famous machine called the IBM Stretch, writing programs to grind the numbers down, refining the model of the Earth year after year through the seventies and eighties. Slow, exact, cumulative arithmetic. Her models fed the geodetic Earth models the Defense Department built, the lineage that ended in the world reference frame your phone still uses.

So let me be careful about credit, both directions, because this is where the popular story goes wrong twice. First, she did not invent GPS.

GPS was a vast program over decades. Other people designed the satellite constellation, built the atomic clocks, worked out the signals, ran the ground control. West was one crucial contributor to one essential piece, and she was modest about it herself. The single-inventor headline flattens a whole enterprise into one hero, and that is just not how it happened.

But the story goes wrong the other way too, and this is the correction that matters more. When we tell it only as hidden figure, a brilliant Black woman denied her due, we make it a story about recognition and identity. All of that is true. She was the second Black woman ever hired at that Navy base, out of a farm and a sharecropping childhood in segregated Virginia, valedictorian, two math degrees, and later, in her seventies, after a stroke, she finished a doctorate by distance learning. Her own children did not know what she worked on, because it was secret. She surfaced into public view only in 2018, when a member of her sorority read a short biography she had written for an alumnae event. All of that deserves telling. But if we stop at denied recognition, we skip the thing itself.

Her real achievement was a hard and elegant piece of applied mathematics: taking a firehose of imperfect measurements bouncing off a moving ocean and reducing it into the true shape of the Earth as gravity sees it. That is worth understanding on its own terms, not just celebrating. The honest boundary is that we cannot cleanly itemize which bumps in the model are hers; the work was collaborative and partly classified, and the clearest public statement is an Air Force citation crediting her with programming the calculations for an accurate geodetic model of the Earth that was used to determine the GPS orbit.

So here is the sentence I would keep. Gladys West did not invent GPS. She did something the invention could not have happened without. She helped find out where sea level actually is, which turns out to be a hundred-meter question with no simple answer, hidden in the shape of a lumpy planet. The next time your phone drops its blue dot exactly where you are standing, remember what is underneath the trick: a still, gravity-drawn ocean, and a person who spent decades doing the arithmetic to find its shape.

There is a kind of science story that arrives dressed as good news, and this is one of them.有一类科学故事，登场时是一副好消息的模样，这就是其中之一。 Over the last few weeks you may have seen a version of it. A volcano, the headlines said, cleaned up its own pollution.过去这几周，你或许见过它的某个版本。头条说，一座火山，清理掉了自己制造的污染。 It erupted, it poured a powerful greenhouse gas into the sky, and then, through some strange chemistry nobody expected, it destroyed that same gas again.它喷发了，把一种强效温室气体喷向天空，然后，通过某种没人料到的奇怪化学过程，又把那同一种气体给摧毁了。 Scientists stunned. A natural weapon against global warming. Nature showing us how to cool the planet.科学家震惊。一件对抗全球变暖的天然武器。大自然向我们展示了如何为地球降温。

The chemistry underneath all of this is real.这背后的化学过程是真实的。 It is genuinely odd and genuinely beautiful, and I want to walk you through it, because it is worth understanding.它确实古怪，也确实美妙，我想带你一步步看清楚，因为它值得理解。 But almost everything the headlines stacked on top of it is backwards.但那些头条堆在它上面的东西，几乎全都反了。 The honest version of the story is more interesting than the hopeful one, and it ends somewhere better. 这个故事诚实的版本，比那个满怀希望的版本更有意思，而且它的落点更好。 So let me tell you what actually happened, what was measured, what was only claimed, and why the most useful thing here is not the thing that made the headline.所以让我告诉你到底发生了什么，什么是测量到的，什么只是被声称的，以及为什么这里最有用的东西，并不是登上头条的那件事。

The volcano is Hunga Tonga. In January of 2022 it erupted under the Pacific, near the islands of Tonga, and it was not an ordinary eruption.这座火山是洪加汤加（Hunga Tonga）。2022 年 1 月，它在太平洋汤加群岛附近的海面下喷发，而这不是一次寻常的喷发。 Because the vent sat under the sea, the blast did something volcanoes almost never do.由于喷口位于海面之下，这次爆发做了一件火山几乎从不会做的事。 It fired an enormous column of seawater into the sky, along with ash and iron and salt, so much water that it measurably raised the amount of water sitting in the stratosphere, the high, thin, bone-dry layer above the weather.它把巨量的海水连同火山灰、铁和盐一起射向天空，水量之大，甚至可测量地抬升了平流层中的水含量——平流层是位于天气层之上、又高、又薄、干燥得像枯骨的那一层。 Picture a fire hose of ocean aimed straight up, past the clouds, past the height where airplanes fly, and then imagine that spray freezing and drifting for weeks.想象一根消防水龙头对准正上方喷射海洋之水，越过云层，越过飞机飞行的高度，然后想象那片喷雾冻结、漂移长达数周。 That single detail, salt water in the stratosphere, is what set up everything that follows. Now the chemistry. Volcanic ash is rich in iron.就是这一个细节——盐水进入平流层——铺垫了接下来的一切。现在说化学。火山灰富含铁。

Seawater is rich in chloride, the chlorine half of ordinary table salt.海水富含氯离子，也就是普通食盐里氯的那一半。 Blast them upward together and they mix into tiny particles that chemists call iron salt aerosol.把它们一起炸向高空，它们混合成微小的颗粒，化学家称之为铁盐气溶胶（iron salt aerosol）。 On their own these particles do nothing dramatic.单凭它们自己，这些颗粒不会有什么惊人举动。 But hit them with direct sunlight, the raw sunlight you get above the clouds, and the light knocks a chlorine atom loose.但用直射的阳光照到它们——就是云层之上那种未经削弱的原始阳光——光会把一个氯原子撞松出来。 Free chlorine in the air is a wolf. It does not sit still.空气中游离的氯是一头狼。它不会安分待着。 It hunts for the nearest molecule it can tear a hydrogen atom off of, and one of its favorite targets is methane, that powerful greenhouse gas.它四处搜寻最近的、能从上面撕下一个氢原子的分子，而它最喜欢的目标之一，就是甲烷，那种强效温室气体。 The chlorine snaps a hydrogen off the methane, the methane begins to fall apart, and one of the fragments it leaves behind, a few steps down the chain, is formaldehyde.氯从甲烷上掰下一个氢，甲烷开始瓦解，而它在这条链上往下走几步后留下的碎片之一，是甲醛。

Hold on to that last molecule, because formaldehyde is how anyone knew any of this was happening. You cannot photograph a wolf.记住最后这个分子，因为甲醛正是任何人得以知道这一切正在发生的凭据。你没法给一头狼拍照。 Chlorine attacking methane is invisible from space.氯攻击甲烷，从太空看是不可见的。 But the wreckage it leaves, the formaldehyde, absorbs light in a way a satellite can see.但它留下的残骸——甲醛——以一种卫星能看见的方式吸收光。 And a European satellite instrument, staring down at the plume, saw a record amount of it.而一台欧洲的卫星仪器，俯视着那团喷发羽流，看到了创纪录的量。 That alone would only tell you the reaction happened somewhere. The clever part is timing. Formaldehyde is fragile.单凭这一点，只能告诉你反应在某处发生过。精妙之处在于时间。甲醛很脆弱。 In the atmosphere it survives only a few hours before it too breaks down.在大气中，它也只能存活几个小时，之后同样会分解。 So finding fresh formaldehyde is like finding fresh footprints in snow.所以，找到新鲜的甲醛，就像在雪地里发现新鲜的脚印。 When the satellite tracked that formaldehyde for ten days, all the way across the ocean to South America, it meant the footprints were being laid down the entire way.当卫星追踪这些甲醛长达十天，一路横跨海洋直到南美洲，这就意味着，脚印是在这一整段路途上被不断留下的。 The wolf was not caught on camera once. It never stopped hunting for over a week. That is the measurement, and it is a real achievement.这头狼不是被镜头逮到了一次。它一个多星期都没有停止过狩猎。这就是那项测量，而它是一项真实的成就。 Formaldehyde is the ash of an invisible fire. So far, so good.甲醛是一场无形之火留下的灰烬。到这里都还没问题。

Now here is where the headlines go wrong, in two separate ways, and both are worth seeing clearly. The first is a matter of arithmetic.而接下来，就是那些新闻标题出错的地方了，错在两个不同的方面，两处都值得看清楚。第一处是算术问题。

The story says the volcano cleaned up after itself. It did not.报道说，这座火山把自己造成的烂摊子收拾干净了。它并没有。 The eruption itself released methane, a lot of it, roughly three hundred gigagrams.喷发本身释放了甲烷，而且量很大，大约 300 吉克（gigagram）。 If you want a picture, that is about a year's worth of burping from more than two million cows, all let go in an afternoon.打个比方，这大约相当于两百多万头牛打一整年的嗝，全部在一个下午里排放出来。 And how much did the famous chemistry destroy?那么，这套广受称道的化学反应又销毁了多少呢？ Around nine hundred tonnes a day, which happens to be about one single day of burping from those same two million cows.每天大约 900 吨，而这恰好相当于那同样两百万头牛打一天的嗝。 Emitted a year's worth. Ate back one day's worth, and kept eating for a week and a half.排放了一年的量。吃回去一天的量，然后又持续吃了一个半星期。 Do the sum and the volcano destroyed only a few percent of the methane it had just put into the air. It was not a cleaner.算下来，这座火山销毁的甲烷，只占它刚刚排入空气中总量的百分之几。它不是清洁工。 It was a polluter that wiped one corner of the mess it made. Calling it a net cleanup is not a small exaggeration.它是个污染者，只擦掉了自己制造的烂摊子的一个角落。把它称作净清理，可不是什么小小的夸张。 It gets the sign of the whole thing wrong. The second error is bigger, and it hides inside the word weapon.它把整件事的正负号都搞反了。第二个错误更大，它藏在“武器”这个词里。

If nature can strip methane out of the sky with sunlight and iron and chlorine, the reasoning goes, then surely we could do it on purpose.这套推理是这样的：如果大自然能用阳光、铁和氯把甲烷从天空中剥离出去，那我们当然也可以有意为之。 Build the particles, spray them, cool the planet. And here you have to say plainly what that reactive chlorine is.造出这些颗粒，喷洒出去，给地球降温。而在这里，你必须直白地说清楚，那种活性氯到底是什么。 It is the same family of chemistry that ate the ozone layer.它属于当年吞噬臭氧层的同一类化学物质。 We spent forty years, and an entire international treaty, getting reactive chlorine out of the upper atmosphere, because up there it destroys the ozone that shields us from the sun.我们花了四十年，动用了一整份国际条约，才把活性氯从高层大气中清除出去，因为在那里，它会破坏为我们遮挡阳光的臭氧。 The proposal on the table is to start putting a cousin of it back into the lower atmosphere, deliberately, at scale.而如今摆在桌面上的提议，是要开始把它的近亲重新放回低层大气，有意为之，而且是大规模地放。 That might turn out to be manageable. It might not. But it is not a free gift from a friendly volcano.这或许最终是可控的，也或许不是。但它绝不是一座友善的火山送来的免费礼物。 It is a serious intervention with a serious history. And there is one more thing you should know, because it changes how you read the paper.它是一项背负着沉重历史的严肃干预。还有一件事你应该知道，因为它会改变你阅读这篇论文的方式。

The scientist who led this study is also one of the main advocates for doing exactly this on purpose, for the idea of releasing iron salt aerosol from ships to pull methane out of the air.领导这项研究的科学家，同时也是主张有意去做这件事的主要倡导者之一——即从船上释放铁盐气溶胶（iron salt aerosol），把甲烷从空气中拉出来这个想法。 That does not make the work wrong. The measurement stands on its own.这并不意味着这项工作是错的。这次测量本身是站得住脚的。 But it means the paper is, in a real sense, a proof of concept for its author's own proposal, and that is precisely when you should slow down and separate what was shown from what is hoped.但这意味着，从某种真切的意义上说，这篇论文是对其作者本人提议的一次概念验证（proof of concept），而这恰恰是你应该放慢脚步、把“已经证明的”与“所期望的”区分开来的时候。 What was shown is solid: this chemistry runs at large scale in the open sky, and you can watch it from orbit.已经证明的部分是扎实的：这套化学反应在开阔的天空中大规模地进行，而且你可以从轨道上观测到它。 What is hoped, that this points to a way to cool the planet, is a long leap over ozone, over unknown byproducts, over the fact that the field cannot yet say whether spraying these particles would remove methane or, through some other reaction, add to it.而所期望的部分——认为这指向了一种给地球降温的方法——则是一次远距离的跳跃：跨过臭氧问题，跨过未知的副产物，也跨过这样一个事实：该领域目前还无法断定，喷洒这些颗粒究竟是会去除甲烷，还是会通过某种别的反应反而增加甲烷。 Their own research roadmap admits that uncertainty out loud.他们自己的研究路线图，也公开承认了这种不确定性。

Let me also be honest about the number itself, because it matters for how much weight it can carry.我也想对这个数字本身坦诚以告，因为它关系到这个数字能承载多大的分量。 Nobody measured nine hundred tonnes of methane disappearing.没有人真的测量到 900 吨甲烷消失了。 They measured formaldehyde, and then ran it backward through a chemical model to infer how much methane must have been destroyed to leave that much wreckage.他们测量的是甲醛，然后通过一个化学模型反向推算，去推断要留下这么多“残骸”，必然有多少甲烷被销毁掉了。 That is an inverse estimate, the kind where you see the footprint and calculate the animal.这是一种反演估计（inverse estimate），就是那种看到脚印去推算动物的做法。 It is a reasonable way to work, but the answer depends on the model, not on a direct reading, and a different model of the chemistry would give a somewhat different tonnage.这是一种合理的做法，但答案取决于模型，而非直接读数，换一个不同的化学模型，就会给出多少有些不同的吨数。 Worth remembering before anyone quotes the figure as if it were weighed on a scale. And the eruption's effect on the actual climate?在有人把这个数字当作是用秤称出来的一样去引用之前，这一点值得记住。那么，这次喷发对实际气候的影响，又如何呢？

That is a separate, larger, still-unsettled question, and the methane is a footnote to it.那是一个独立的、更宏大的、至今仍未解决的问题，而甲烷不过是它的一个脚注。 The dominant thing Hunga Tonga did was that water vapor, which is itself a greenhouse gas, set against a smaller-than-usual load of reflective particles.洪加汤加火山最主要造成的影响是那些水汽——水汽本身就是一种温室气体——再叠加上比往常偏少的反射性颗粒物负荷。 The current best reading is that, on balance, the eruption slightly cooled the southern half of the planet for a year or two.目前最好的解读是，总体而言，这次喷发让地球南半部在一两年里略微降温。 Not warmed, not saved.不是变暖，也不是被拯救。 The methane chemistry we have been talking about is a small side story to a big perturbation that scientists are still arguing about.我们一直在谈的甲烷化学，只是一场大扰动的一个小小的插曲，而科学家们对这场大扰动仍在争论。

So what should you actually take from this? Not that a volcano handed us a lever to cool the Earth. It did not.那么你到底应该从中得到什么？不是说火山递给了我们一根给地球降温的杠杆。它并没有。 It emitted far more methane than it destroyed, and the chemistry that destroyed it is the chemistry we spent a generation banning.它排放的甲烷远远多于它所摧毁的，而摧毁甲烷的那套化学，正是我们花了一代人时间去禁止的那套化学。 What the volcano handed us is smaller and, I think, better.火山递给我们的东西更小，而且我认为，更好。 It handed us a clean natural experiment, the atmosphere running a reaction at a scale no laboratory could stage, and it handed us a way to read that reaction from space, through the short-lived formaldehyde it leaves behind.它递给我们一场干净的自然实验——大气以任何实验室都无法搭建的规模运行着一场反应；它还递给我们一种从太空解读这场反应的方法，通过它留下的那种寿命短暂的甲醛。

And that reading is the part that will last.而这份解读，正是会长久留存下来的部分。 Because if anyone ever does try to do this deliberately, the one thing you would absolutely need is an independent, satellite-based way to check whether it is helping or hurting, in real time, over the ocean, where no one is standing to measure it.因为如果真有人哪天试图刻意去做这件事，你绝对需要的一样东西，就是一种独立的、基于卫星的手段，去实时地、在海洋上空——那里没人站着做测量——查验它究竟是在帮忙还是在帮倒忙。 This study did not build the weapon. It built the referee. And a referee you can trust is worth a great deal more than a weapon you cannot.这项研究并没有造出武器，它造出的是裁判。而一个你能信任的裁判，其价值远远超过一件你无法信任的武器。

Start with a shape you could cut out of cardboard.先想象一个你能用硬纸板剪出来的形状。 Thirteen straight sides, a little lopsided, and if you squint it looks like a fedora, so the people who found it called it the hat.十三条直边，略微有点不对称，眯起眼看有点像一顶软呢帽，所以发现它的人就叫它「帽子」。 Now imagine tiling your kitchen floor with copies of that one shape.现在想象用这一个形状的复本去铺满你家的厨房地板。 No gaps, no overlaps, the hats fitting snugly edge to edge, out to the walls and, in your mind, out past the walls forever.没有缝隙，没有重叠，一顶顶帽子边对边严丝合缝地拼接，一直铺到墙边，而在你的想象里，还越过墙壁无限延伸下去。 Here is the strange part. However far you go, the pattern never repeats.奇怪的地方在这里。无论你铺得多远，这个图案永远不重复。 There is no stamp, no motif, no tile you could slide the whole floor along and have it land back on itself.没有一个印章，没有一个母题，没有一块可以让你把整块地板沿着它滑动、然后正好落回自身的图砖。 One humble shape, and it forces a covering of the plane that is ordered everywhere and identical to itself nowhere.一个不起眼的形状，却逼出了一种对平面的覆盖：处处有序，却处处都不与自身相同。

The man who found it is not a mathematician.发现它的人并不是数学家。 David Smith is a retired print technician from the Yorkshire coast, a hobbyist who likes to play with shapes on his computer and then cut them out of card to see how they behave.David Smith 是约克郡海岸一位退休的印刷技师，一名业余爱好者，喜欢在电脑上摆弄各种形状，然后剪成卡片看它们表现如何。 In the autumn of 2022 he made this thirteen-sided hat, started laying copies down on his table, and noticed they kept going without ever settling into a rhythm.2022 年秋天，他做出了这个十三边的「帽子」，开始在桌上一块块拼下去，注意到它们能一直拼下去，却从不落入某种节律。 He had the sense to email someone who would know what he was looking at, a computer scientist named Craig Kaplan, and within months Kaplan, Smith, and two others had proved it.他有心把这事发邮件给一个懂行的人——一位名叫 Craig Kaplan 的计算机科学家。几个月内，Kaplan、Smith 以及另外两人就证明了它。 This little shape solved a problem that had been open for about half a century.这个小小的形状，解决了一个悬置了约半个世纪的问题。

The problem has a name that causes endless confusion, so let me clear it up first. It is called the einstein problem.这个问题的名字会引起没完没了的误会，所以我先把它讲清楚。它叫「爱因斯坦问题」（einstein problem）。 Not after Albert Einstein. It is a pun. In German, ein Stein means one stone, one tile, and that is the whole question.不是以阿尔伯特·爱因斯坦命名的。这是个双关。在德语里，ein Stein 意思是「一块石头」「一块砖」，而这就是整个问题的全部。 Back in the 1970s the physicist Roger Penrose found a set of just two shapes that tile the plane and only ever tile it without repeating.早在 1970 年代，物理学家 Roger Penrose 就找到了一组仅有两个的形状，它们能铺满平面，而且只能以不重复的方式铺满。 Two tiles. And people immediately asked the obvious next thing. Could you do it with one? For fifty years nobody knew.两块图砖。人们立刻问出了显而易见的下一个问题：能不能用一块做到？五十年里，没人知道。 Most people assumed the answer was no. In 2023, a man with cardboard proved the answer was yes.多数人都以为答案是否定的。2023 年，一个拿着硬纸板的人证明了答案是肯定的。

So the first thing I want to correct is the way this got written up.所以我想纠正的第一件事，是这件事被报道的方式。 You will see headlines saying the hat broke mathematics, or overturned it. It did nothing of the kind.你会看到一些标题说「帽子」打破了数学，或者颠覆了数学。它根本没做这种事。 It answered a precise, long-standing question with a clean yes. That is the opposite of breaking something.它用一个干净利落的「是」回答了一个精确的、悬置已久的问题。这跟「打破」什么东西正好相反。 And there is a subtlety underneath, the kind this listener likes, so stay with me. The hat has a catch.而底下还有一个微妙之处，正是这位听众喜欢的那种，所以请跟着我。「帽子」有个附带条件。 To make the pattern work, you need the hat and its mirror image.要让这个图案成立，你需要「帽子」和它的镜像。 You have to flip some of the tiles over, like using both a left glove and a right glove.你得把其中一些图砖翻过来，就像同时用一只左手套和一只右手套。 And the purists grumbled, quite reasonably, that if you need the shape and its reflection, that is really two tiles, not one.而那些纯粹主义者相当有理地抱怨说：如果你需要这个形状连同它的反射，那其实是两块图砖，不是一块。 So the same team went back, and a couple of months later they found another shape, nicknamed the spectre, that needs no flipping at all.于是同一个团队又回头研究，几个月后他们找到了另一个形状，绰号叫「幽灵」（spectre），它完全不需要翻转。 It tiles the plane forever, never repeating, using only rotations and slides, one handedness only. That one is the true single tile.它能永远铺满平面，从不重复，只用旋转和平移，只用一种手性。那一个才是真正的单块图砖。 Hold on to that distinction, because it comes back. Now, what does it actually mean for something to be ordered but never repeating?记住这个区别，因为它还会回来。那么，一样东西「有序却永不重复」到底意味着什么？

This is the real idea, and it is worth a minute. We are used to two categories.这才是真正的核心，值得花一分钟。我们习惯了两个类别。 There is periodic, like a brick wall or a crystal of salt, where a single unit repeats on a grid, and if you shift it just right it lands back on itself.有周期性的，比如砖墙或者盐的晶体，一个单元在网格上重复，只要你恰当地平移它，它就会落回自身。 And there is random, like a pile of sand, where there is no order at all. For a long time people thought those were the only options.还有随机的，比如一堆沙子，其中根本没有秩序。很长一段时间里，人们以为这就是仅有的两种选择。 Order meant repetition. This new thing is a third category that sits between them. It is rigidly, deterministically ordered.秩序意味着重复。而这个新事物是介于两者之间的第三类。它是严格的、确定性的有序。 Tell me the rule and I can tell you exactly where every tile goes, out to infinity. But it never repeats. There is no unit, no grid.告诉我规则，我就能准确说出每一块拼片放在哪里，直到无穷远。但它从不重复。没有重复单元，没有网格。

Physicists had to swallow this the hard way.物理学家是硬着头皮才接受这一点的。 In 1982 a materials scientist named Dan Shechtman, looking at a metal alloy through an electron microscope, saw a diffraction pattern with a fivefold symmetry that, by the accepted rules of crystals, was simply forbidden.1982 年，一位名叫 Dan Shechtman 的材料科学家，用电子显微镜观察一种金属合金时，看到了一种带有五重对称性的衍射图样——按照公认的晶体规则，这根本是不可能存在的。 Ordered enough to make sharp, clean spots, but with a symmetry no repeating crystal can have. He was mocked for it.它足够有序，能产生尖锐、清晰的斑点，却带有任何周期性晶体都不可能具有的对称性。他因此遭到嘲笑。 Linus Pauling, one of the most famous chemists alive, said there was no such thing as quasicrystals, only quasi-scientists.Linus Pauling，当时最著名的化学家之一，说根本没有什么准晶体，只有准科学家。 Shechtman was asked to leave his research group. He was right, and everyone else was wrong, and it took years.Shechtman 被要求离开他的研究小组。他是对的，其他所有人都错了，而这花了很多年才被认清。 He got the Nobel Prize in 2011, alone, for seeing a kind of order the textbooks said could not exist. That is the family the hat belongs to.他在 2011 年独自获得了诺贝尔奖，因为他看见了一种教科书说不可能存在的有序。这就是「帽子」所属的家族。 It is the tiling picture of a quasicrystal.它就是准晶体的拼贴图像。

Which brings us to why I am telling you this now, because the hat was a beautiful piece of pure geometry, and pure geometry is not usually a news story three years later.这就引出了我现在讲这些的原因，因为「帽子」是一件漂亮的纯几何作品，而纯几何通常不会在三年后还成为新闻。 The news is that a group in Tokyo, led by Yuto Moritake and Masaya Notomi, took the hat out of mathematics and turned it into a material.新闻是，东京的一个研究组，由 Yuto Moritake 和 Masaya Notomi 领导，把「帽子」从数学中取了出来，变成了一种材料。 They etched the hat pattern, as a field of tiny holes, into a thin film of silicon nitride, the sort of chip you make with the same tools that make computer processors.他们把帽子图样以一片微小孔洞阵列的形式，蚀刻进一层氮化硅薄膜——这种芯片用的正是制造计算机处理器的同一套工具。 Then they shone a laser through it and watched what came out the other side. Let me do the two columns this listener always asks for.然后他们用激光照射它，观察从另一侧射出的光。让我来做这位听众总是要求的那两栏对照。

What was shown, and what was claimed. Shown.哪些被证明了，哪些只是被宣称。先说被证明的。

First, the pattern diffracts into sharp, clean spots, real Bragg peaks, the fingerprint of genuine long-range order.第一，这个图样衍射出尖锐、清晰的斑点，真正的 Bragg 峰，是真正长程有序的指纹。 So the aperiodic order is not just a mathematician's idea on paper. You can build it in silicon and light confirms it is there.所以这种非周期有序不只是数学家纸上的想法。你可以在硅里把它造出来，而光证实了它确实存在。 Second, those spots form a pinwheel, a pattern that has a threefold turn but no mirror line.第二，这些斑点形成一个风车状图样，它有三重旋转对称，却没有镜像对称线。 And third, the new thing, the reason there is a paper.第三，也就是那个新东西，也是这篇论文之所以存在的理由。 When they sent in circularly polarized light, light that corkscrews as it travels, the pattern responded differently to left-handed twist than to right-handed twist.当他们送入圆偏振光——这种光在传播时呈螺旋状——图样对左旋和右旋的响应是不同的。 Left-corkscrewing light brightened certain spots, right-corkscrewing light brightened different ones.左旋螺旋的光让某些斑点变亮，右旋螺旋的光则让另一些斑点变亮。 That difference, light telling left from right, has never been seen in an ordinary quasicrystal. The material is handed.这种差异——光能分辨左右——在普通准晶体中从未见过。这种材料是有手性的。

Now the careful part, the mechanism, and a second thing the popular story gets backwards. Where does the handedness come from? 现在讲需要小心的部分，机制，以及大众报道搞反的第二件事。手性从何而来？ Not from a chiral tile. Remember, the hat tiling uses the shape and its mirror image both. The individual bricks come in both hands.不是来自手性拼片。记住，帽子拼贴同时用到了这个形状及其镜像。单个「砖块」是左右两种手性都有的。 The chirality does not live in the brick. It lives in the arrangement.手性并不存在于砖块之中，而是存在于排列方式之中。 As you build the hat pattern outward in layers, each layer sits tilted from the one below it, by a fixed angle of about fifteen and a half degrees, set by the golden ratio.当你一层层向外构建帽子图样时，每一层相对于它下面那层都倾斜一个固定的角度，约 15.5 度，由黄金比例决定。 That twist accumulates, and the whole grown pattern ends up with no mirror line anywhere in it, even though every tile in it has a mirror twin.这种扭转不断累积，最终整个生长出来的图样里任何地方都没有镜像对称线，尽管其中每一块拼片都有一个镜像孪生。 The handedness is a property of the crowd, not of any person in it.手性是这个群体的属性，而不是其中任何一个个体的属性。 That is the honest sentence, and it is more interesting than the tile did it. And the light.这才是诚实的说法，而且它比「是拼片干的」更有意思。再说光。

The phrase you will read is that the tile twists light into pinwheels, which makes it sound like some new force. It is not.你会读到的说法是，这种拼片把光扭成风车状，这让它听起来像是某种新的力。其实不是。 It is diffraction, the same plain physics that makes a CD flash rainbow colours, light bending around a regular array of features and interfering with itself.这就是衍射，和让 CD 闪现彩虹色的是同一种朴素的物理——光绕过规则排列的结构，并与自身发生干涉。 What is new is not the bending. It is that the array is handed, so the bending is handed too.新的地方不在于弯折本身，而在于这个阵列是有手性的，所以弯折也就带上了手性。 State it plainly and the wonder survives, because normally, to make an optical device that can tell left-corkscrew light from right, you need molecules that are themselves twisted, or layers deliberately stacked with a spiral.把话说白了，这份惊奇依然成立，因为通常，要造一个能分辨左旋光和右旋光的光学器件，你要么需要分子本身就是扭曲的，要么需要刻意按螺旋方式堆叠的层。 Here there is nothing twisted in the chemistry at all. The geometry of a flat pattern of holes does the whole job.而这里，化学结构中根本没有任何扭曲的东西。一个平面孔洞图案的几何形状就完成了全部工作。 That is the real prize, and it is genuinely new. Now the boundary, honestly. This is a demonstration, not a device.这才是真正的收获，而且是货真价实的新东西。不过老实说，也要讲清楚边界。这是一次演示，而不是一件器件。

The strength of the handed response also depends on the spacing of the holes and the angle you look at, not the geometry alone, and the threefold symmetry is only exact around special centres of the pattern, weaker elsewhere.这种手性响应的强度，也取决于孔洞的间距和你观察的角度，而不仅仅取决于几何形状；而且这种三重对称只在图案的某些特殊中心处才严格成立，在别处则较弱。 Nobody has built anything with it yet.还没有人用它造出任何东西。 What has been shown is that a shape from pure mathematics, with no periodicity and no unit cell, can be made into matter that does something to light that ordinary matter cannot.已经展示出来的是：一个来自纯数学、没有周期性、没有单元胞的形状，可以被做成一种物质，对光做出普通物质做不到的事情。 That is a first rung, not a finished ladder, and the authors are careful to say so. Keep two things from this.这是第一级台阶，而不是一架造好的梯子，作者们也谨慎地这样说。从中记住两点。

First, order does not have to mean repetition.第一，有序不一定意味着重复。 There is a whole third country between the crystal and the sand pile, deterministic and never repeating, and we have only had the map to it for about forty years and a single tile for three.在晶体和沙堆之间，还存在着一整片第三国度：确定性的，却永不重复；而我们拥有通往它的地图不过约四十年，拥有单块拼砖才三年。 Second, and this is the one I would send you off with.第二点，也是我想让你带走的一点。 A retired print technician, cutting shapes out of card at his kitchen table, answered a question that professional mathematicians had circled for half a century.一位退休的印刷技师，在自家厨房的餐桌上用卡纸剪出各种形状，回答了一个专业数学家们绕行了半个世纪的问题。 And three years later, in a clean room in Tokyo, that same shape is bending light in a way no crystal can.而三年后，在东京的一间洁净室里，同样的这个形状正以任何晶体都做不到的方式弯折着光。 The distance between a shape and a material, between a game with cardboard and a new kind of optics, is a great deal shorter than it looks.从一个形状到一种材料，从一场用硬纸板玩的游戏到一种新的光学，这段距离比看上去要短得多。

Start with a burnt bone. It is tiny, smaller than a matchstick, the leg of a mouse or a shrew.从一根烧焦的骨头说起。它很小，比一根火柴还短，是一只老鼠或鼩鼱的腿骨。 It has been lying in the dark, thirty metres inside a cave in the Northern Cape of South Africa, for something close to one million eight hundred thousand years.它一直躺在黑暗中，在南非北开普省一处洞穴深入三十米的地方，时间接近 180 万年。 And this summer a group of scientists looked at it under blue light and reported that it had once been hot. Not warm. Burned.今年夏天，一群科学家在蓝光下观察它，并报告说它曾经受过高温。不是温热，而是被烧过。

The headlines wrote themselves. Humans used fire one point seven nine million years ago.标题自然而然就写了出来。人类在 179 万年前就用上了火。 Nearly twice as old as anyone had firm evidence for.这比任何人此前掌握确凿证据的年代几乎早了一倍。 And I want to spend these ten minutes on why that sentence, the one that wrote itself, is not the thing that was found.而我想用这十分钟，来讲讲为什么那句自己冒出来的话，并不是真正被发现的东西。 Because the thing that was found is a bone that got hot. Everything after that word is an argument.因为真正被发现的，是一根变热过的骨头。这个词之后的一切，都是一场推断。 And the argument turns out to be one of the hardest in all of science, for a reason that is worth understanding whatever you work on.而这场推断结果证明是整个科学界最难的推断之一，其原因值得任何领域的人去理解。

Here is the reason, and it is the whole episode. A stone hand axe is an object.原因就在这里，它贯穿了整期节目。一把石制手斧是一件实物。 If you find one, a person made it, because rivers and wind and lightning do not knap flint into a teardrop with a cutting edge.如果你找到一把，那一定是人做的，因为河流、风和闪电不会把燧石打制成带刃口的泪滴形。 The object is the proof. Fire is not like that. Fire leaves no object. It leaves only altered matter.实物本身就是证据。火不是这样。火不留下实物，它只留下被改变的物质。 Bone that has been heated, sediment that has been reddened, wood turned to ash. And here is the trap.被加热过的骨头、被烧红的沉积物、化为灰烬的木头。而陷阱就在这里。 Nature alters matter in exactly the same ways. Wildfires are not a human invention.大自然会以完全相同的方式改变物质。野火并非人类的发明。 They are older than dinosaurs, older than four hundred million years. Lightning has been starting fires since long before anything walked.它们比恐龙还古老，早于 4 亿年前。早在任何生物行走于地面之前，闪电就一直在引燃大火。 So when you find something burned, you have not found a fire. You have found that something got hot.所以当你找到被烧过的东西时，你并没有找到一场火。你只是发现某样东西变热过。 The entire science of the origin of fire is the fight to prove that a particular hot thing was made hot by one of us, and not by the world.整个关于火起源的科学，就是一场努力，去证明某个变热的东西是被我们中的一员弄热的，而不是被自然界弄热的。

So think about what a single burnt bone is actually consistent with. It could be a hearth, tended by a hominin, which is the exciting story.所以想想看，一根烧焦的骨头究竟能对应哪些情况。它可能是一处火塘，由某个古人类照料，这是令人兴奋的那种故事。 But it could also be a wildfire that swept the landscape and whose heat reached into a cave mouth.但它也可能是一场席卷大地的野火，其热量一直蔓延到了洞口。 It could be spontaneous combustion of bat guano, which really does happen, deep piles of droppings that heat and catch on their own.它可能是蝙蝠粪的自燃——这确实会发生，深厚的粪堆会自行升温并着火。 It could be that the bone was never burned at all, and what looks like scorching is manganese staining, a black mineral crust that mimics char to the naked eye.它也可能根本没被烧过，看似焦痕的东西其实是锰染色——一层黑色矿物结壳，肉眼看上去很像炭化。 Or the bone burned long after it was buried, cooked by geology, by heat moving through rock over ages. One little black bone.又或者，这根骨头是在被埋藏很久之后才烧焦的，是被地质作用烤过的，是历经漫长年代穿过岩石传导的热量所致。一根小小的黑骨头。 At least five completely different histories. All of them end with the same object in your hand. This is the problem in its pure form.至少五种截然不同的经历。它们最终都让你手中握着同一件东西。这就是这个问题最纯粹的形态。 The observation does not pick out the cause. You cannot resolve it by staring at the bone harder.观察本身无法挑出成因。你没法靠更用力地盯着这根骨头来解决它。 There is no amount of looking at that one bone that tells you which of the five happened. You resolve it by subtraction.无论怎么看这一根骨头，都无法告诉你五种情况中发生了哪一种。你要靠排除法来解决它。

You go and rule the alternatives out, one at a time, using information that is not in the bone itself.你要去逐一排除其他可能，一次排除一个，用的是骨头本身之外的信息。 And that is what this new study is, read honestly. It is a subtraction.而这项新研究，如实来看，正是如此。它是一次排除。

The first alternative they cut is the staining one, and this is the genuinely new part, the reason there is a paper at all.他们排除的第一个可能是染色那一种，这也是真正新颖的部分，是这篇论文得以存在的理由。 The method is borrowed from forensics. You shine high energy blue light on the bone under a microscope.这种方法借鉴自法医学。你在显微镜下用高能蓝光照射骨头。 Bone that has truly been heated has had its crystal structure changed by the heat, and that changed structure glows, it gives back a longer wavelength, a vivid red through a filter.真正被加热过的骨头，其晶体结构已被热量改变，而这种改变后的结构会发光，它会回射出更长的波长，透过滤镜呈现出鲜艳的红色。 A mineral stain does not do this. So the glow separates real heating from a black crust that only looks like burning.矿物染色不会这样。因此，这种发光把真正的受热，与只是看起来像烧过的黑色结壳区分开来。 That is one branch of the tree cut off cleanly. The bone was hot. Not stained. Hot.这就干净利落地砍掉了这棵树的一根枝杈。骨头是被加热过的，不是被染色的。是加热。

The second alternative is the wildfire, and here the evidence is not a measurement, it is a location.第二种可能是野火，而这里的证据不是一项测量，而是一个位置。 The bones sit thirty metres inside the cave. Not at the mouth where a landscape fire could reach, but deep in the interior, in the dark.这些骨头位于洞穴内三十米处。不是在野外大火能够波及的洞口，而是在黑暗中的深处。 And they are concentrated, in patches, not spread as a sheet the way a single sweeping burn would leave them.而且它们是成片集中分布的，而不是像一场席卷而过的大火那样铺成一整层。 Even better, many of the burnt bones are the bones that come out of owl pellets.更妙的是，许多烧焦的骨头正是从猫头鹰食丸里出来的骨头。 Barn owls roosted in the back of this cave for the whole span we are talking about.在我们讨论的整个时间跨度里，仓鸮一直栖息在这个洞穴的深处。 They swallow rodents whole and cough up the fur and bone in little packed pellets, which built up on the floor over centuries.它们把啮齿动物整个吞下，再把毛发和骨头压成一个个小食丸吐出来，几个世纪下来在地面上堆积起来。 So the burnt bones are the rodent scraps from an owl floor, deep in the cave, that later got heated where they lay.所以这些烧焦的骨头，是洞穴深处猫头鹰栖息地面上的啮齿动物残骸，后来在原地被加热了。 A wildfire does not do that. Something carried fire, or kept fire, to that spot.野火不会造成这种情形。是有什么东西把火带到了、或把火维持在了那个地点。 And in the same layers are the crude stone tools of the people who could have carried it.而在同样的地层里，还有那些本可能带火进来的人所留下的粗糙石器。

So now let us do the honest thing this listener always asks for, and put two columns side by side. What was shown. What was claimed. Shown.那么现在，让我们做一件这位听众总是要求的、诚实的事，把两栏并排放在一起。展示出来的是什么。宣称的又是什么。展示出来的。

Bones in this cave were genuinely heated, not stained, by a method that can tell the difference.这个洞穴里的骨头确实被加热过，而非被染色，这是用一种能分辨二者差别的方法确定的。 Some of that heating happened deep in the interior, far from any wildfire, near the tools of Homo erectus, on a floor that erectus used.其中一部分加热发生在洞穴深处，远离任何野火，靠近直立人（Homo erectus）的工具，在一处直立人使用过的地面上。 And it happened in a layer dated, by the flipping of the Earth's magnetic field recorded in the sediment and by the slow clock of cosmic rays, to as far back as one point seven nine million years.而且它发生在一个经由沉积物中记录的地磁场倒转、以及宇宙射线这台缓慢时钟测定的地层里，年代可以远溯至 179 万年前。

Claimed. That Homo erectus mastered fire almost one point eight million years ago.宣称的。直立人在近 180 万年前就掌握了用火。 And notice, the people who did the work do not actually make that second claim. They are careful.请注意，做这项研究的人其实并没有提出后一个说法。他们很谨慎。 They say this does not demonstrate cooking. It does not demonstrate the ability to make fire from scratch.他们说，这并不能证明烹饪。也不能证明从零生火的能力。 They are not even certain exactly when, within a wide window, the burning happened.他们甚至无法确定，在一个很宽的时间窗口内，燃烧究竟发生在什么时候。 What they will say is that hominins may have repeatedly brought fire into this part of the cave and kept it going until it burned out.他们愿意说的是，古人类可能曾反复把火带到洞穴的这一部分，并让火一直燃烧到熄灭。 That is a much smaller sentence than the headline. And it is the correct size for the evidence.这比那条大标题要小得多。而它正是与证据相称的分量。

Which brings me to the second thing the popular story gets backwards, and it is deeper than one cave.这就引出了大众叙事弄反的第二件事，而它比一个洞穴要深刻得多。 People imagine the hard part of this field is finding fire old enough. It is the opposite. Burnt things are everywhere.人们以为这个领域的难点在于找到足够古老的火。恰恰相反。烧过的东西到处都是。 The hard part is the subtraction, proving the world did not do it, and that is why the date of controlled fire is not a number that keeps getting measured more precisely.难点在于做减法，证明不是大自然干的，这也正是为什么受控用火的年代不是一个不断被测得越来越精确的数字。 It is an argument that keeps getting fought.它是一场不断被争论的争论。 This is why serious, careful researchers can look at all of it and still hold a very different line.这就是为什么严肃、谨慎的研究者可以把这一切都看在眼里，却仍然坚持一条很不一样的立场。 Two of the most respected, Wil Roebroeks and Paola Villa, argued years ago that if you demand not a single burnt patch but the habitual, everyday, unmistakable use of fire, the kind you cannot explain any other way, the firm evidence in Europe only appears around three to four hundred thousand years ago.其中最受尊敬的两位，Wil Roebroeks 和 Paola Villa，多年前就提出，如果你要求的不是单独一处烧痕，而是习惯性的、日常的、确凿无疑的用火——那种无法用其他任何方式解释的用火——那么欧洲确凿的证据只在大约三四十万年前才出现。 Not one point seven nine million. Four hundred thousand. They are not denying the old bones exist.不是 179 万年。是 40 万年。他们并不否认那些古老的骨头存在。 They are saying a scatter of burnt bones at one site is the very first rung of a ladder, and the ladder has three rungs.他们是说，一个遗址上零散的烧骨只是一架梯子的最底一级，而这架梯子有三级。 Opportunistic, grabbing fire when nature offers it. Habitual, keeping and using it as routine. Obligate, needing it to survive.机会性的，在大自然给予火时抓住它。习惯性的，把它当作常规来保存和使用。必需性的，离开它就无法生存。 The Wonderwerk bones, even taken at their strongest, reach for the bottom rung. The headline reads them as the top.Wonderwerk 的这些骨头，即便按最有力的解读，也只够得着最底一级。而那条大标题却把它们读成了最高一级。

There is one more thread, and it is why this fight matters so much that people keep having it. Richard Wrangham's cooking hypothesis.还有一条线索，它解释了为什么这场争论如此重要，以至于人们不断重提。理查德·兰厄姆（Richard Wrangham）的烹饪假说。 His claim is that fire is not just something we happened to acquire but the thing that made us, that cooking unlocked so much energy from food that it built the human body.他的主张是，火不只是我们偶然获得的东西，而是塑造了我们的那样东西——烹饪从食物中释放出如此多的能量，以至于构建了人类的身体。 And his evidence is not a hearth at all. It is us.而他的证据根本不是什么灶台，而是我们自己。 Around one point nine million years ago, Homo erectus appears with small teeth, a small gut, a large brain.大约 190 万年前，直立人（Homo erectus）出现了，牙齿小、肠道小、脑容量大。 A body that reads like it stopped eating raw.这样一副身体，看上去就像已经不再吃生食了。 So if Wrangham is right, we should expect fire right about here, at the dawn of erectus, exactly where Wonderwerk now points.所以如果兰厄姆是对的，我们就该预期火恰好出现在这里，在直立人登场之时——正是旺德沃克（Wonderwerk）如今所指向的那个时刻。 Two arguments, from two directions, an anatomy and a cave, meeting at the same moment. And I want to be clear about what that is and is not.两个论证，来自两个方向，一个是解剖结构，一个是洞穴，在同一个时刻交汇。而我想把它是什么、不是什么讲清楚。 It is suggestive, and it is lovely. It is not a fire you can see.它很有启发性，也很美妙。但它不是一团你能看见的火。 It is two inferences agreeing, and two inferences agreeing is a reason to look harder, not a reason to stop.它是两个推论彼此吻合，而两个推论彼此吻合是更用力去看的理由，不是停下来的理由。

So what do you keep from all this. Keep the shape of the problem, because you meet it everywhere.那么从这一切中该记住什么呢？记住这个问题的形态，因为你到处都会遇到它。 One observation, many possible histories, and no way to choose between them by examining the observation alone.一个观测，多种可能的历史，而仅凭审视这个观测本身，无法在它们之间做出选择。 A better instrument, the blue light, kills one branch, staining. The map of where things sit kills another, wildfire.更好的仪器，那束蓝光，排除了一个分支——染色。各样东西所处位置的分布图排除了另一个——野火。 But the last gap, between fire that was seized and fire that was commanded, no measurement closes.但最后那道鸿沟，介于被攫取的火与被掌控的火之间，没有任何测量能够弥合。 Only more, and independent, evidence does. The burnt bone does not lie. It really was hot.只有更多的、独立的证据才能做到。烧焦的骨头不会撒谎。它确实曾经很热。 It simply cannot, by itself, tell you whose fire warmed it.它只是无法凭自身告诉你，是谁的火烤热了它。 That gap, between what got hot and whose hand held the flame, is the whole story.那道鸿沟，介于什么变热了与是谁的手握着火焰之间，就是整个故事。 And we are still, patiently, one cave at a time, trying to close it.而我们仍在耐心地，一个洞穴接一个洞穴地，试图弥合它。

Four days of findings.四天的调查发现。 Today, the question this podcast keeps arriving at from different directions: what was actually measured, and is it the same thing as what is being claimed?今天，这档播客一再从不同方向抵达的那个问题：真正被测量的是什么，它和被宣称的是不是同一回事？

I want to be clear about my position before I start.在开始之前，我想把自己的立场讲清楚。 I think this report is a genuinely valuable document and I think Anthropic deserves credit for publishing it. Almost nobody else does this.我认为这份报告是一份真正有价值的文件，我也认为 Anthropic 公开发布它值得称赞。几乎没有别人会这么做。 Most companies discovering their product being used for espionage handle it quietly, because disclosure is all downside.大多数公司发现自己的产品被用于间谍活动后，都会悄悄处理，因为披露只有坏处没有好处。

And it is unusually candid.而且它坦率得不同寻常。 Several of the caveats I am about to raise, I am raising because the report raises them about itself, which is not the normal standard.我接下来要提出的几点保留意见，之所以提出，是因为报告自己就对自己提出了这些保留——这并不是通常的标准。

So this is not a debunking. It is an attempt to work out what weight the document can carry.所以这不是一次拆穿。这是一次试图厘清这份文件能承载多少分量的尝试。

Start with the sharpest limitation, which is stated plainly in the methodology: visibility ends once content goes live.先从最尖锐的局限说起，方法论部分对此有直白的陈述：一旦内容上线，可见性就到此为止。

Anthropic can see what happened on its own platform. It sees the prompts, the generation, the patterns of use.Anthropic 能看到自己平台上发生的事。它能看到提示词、生成过程、使用模式。 It detects operations during production, upstream of distribution. It cannot see what happened next.它在生产阶段检测到这些行动，位于分发的上游。它看不到接下来发生了什么。

For anything downstream — did the malware work, did the article get read, did the intrusion succeed — the report depends on open-source research, industry data, public reporting, and in some cases the attackers' own claims.对于任何下游的事——恶意软件是否奏效、文章是否被阅读、入侵是否成功——报告都依赖于开源研究、行业数据、公开报道，某些情况下还依赖攻击者自己的说法。

Which means the report is precise about one half of a causal chain and inferential about the other.这意味着报告对一条因果链的一半是精确的，对另一半则是推断的。 It can tell you with real authority what someone asked a model to do. It can tell you much less about what that produced in the world. 它能以真正的权威告诉你某人要求模型去做什么。至于那在现实世界里产生了什么，它能告诉你的就少得多。

Now the selection effect, and this one is important enough that I think it should be stated before any number from the report is quoted.接下来是选择效应，这一点重要到我认为应当在引用报告里的任何数字之前先说明。

The report says it details the most notable and novel threat activity identified. Not typical misuse. The notable and the novel.报告说，它详述的是所识别出的最值得注意、最新颖的威胁活动。不是典型的滥用。是值得注意的、新颖的那些。

So the document is, by construction, a collection of the most striking cases. It is not a sample.所以这份文件在构造上就是一批最引人注目的案例的集合。它不是一个样本。 It cannot tell you what fraction of misuse looks like this, whether these are representative, or whether the median incident is a hundredth of this scale.它无法告诉你滥用中有多大比例是这个样子的、这些是否具有代表性，或者中位事件是不是只有这个规模的百分之一。

That has a specific consequence for how you should read a trend.这对你该如何解读一种趋势有一个具体的后果。 When a report of notable cases describes increasingly autonomous operations, there are two explanations.当一份关于值得注意案例的报告描述行动日益自主化时，有两种解释。 Either operations are becoming more autonomous, or autonomous operations are what now counts as notable enough to include.要么是行动正变得更加自主，要么是自主的行动如今才算得上足够值得注意、才被收录进来。 Those look identical from outside, and the report cannot distinguish them, because notability is the selection criterion rather than a measurement.这两者从外部看是一模一样的，而报告无法区分它们，因为值得注意是它的筛选标准，而不是一项测量。

I think the underlying trend is probably real.我认为底层的趋势很可能是真实的。 The mechanism is plausible, the cases are detailed, and it is consistent with everything else known about agentic systems.机制是可信的，案例是详实的，而且它与关于智能体系统的其他一切已知情况都相符。 But "probably real for external reasons" is a weaker claim than the document appears to support, and the difference matters.但"因外部原因而很可能真实"是一个比这份文件看上去所能支撑的更弱的论断，而这个差别很重要。

Third: some of the most quotable numbers come from the attackers.第三点：一些最可供引用的数字来自攻击者。

The report notes it cannot independently verify self-reported metrics from threat actor dashboards.报告指出，它无法独立核实威胁行为者仪表盘上自我报告的指标。 So when you read that an operation tracked forty-two targets or produced eight thousand nine hundred and thirteen articles, ask where the number came from.所以当你读到某次行动追踪了 42 个目标，或者生产了 8913 篇文章时，要问问这个数字从何而来。 Some are Anthropic's own observations of platform usage, which are solid.有些是 Anthropic 自己对平台使用情况的观察，这些是扎实的。 Some are the operators' own accounting, and criminals inflate their numbers, for the same reasons everyone else does — to impress buyers, partners, and themselves.有些是操作者自己的账目，而罪犯会虚报数字，原因和其他所有人一样——为了打动买家、伙伴，也为了打动自己。

Fourth, attribution.第四，归因。 The report is careful here, using graded language — assessed with high confidence, likely, associated with — and that care should survive summarisation.报告在这一点上很谨慎，用了分级的措辞——高置信度评估、可能、与……相关联——这份谨慎本应在概括转述中保留下来。 It usually does not. "Anthropic says a Chinese defence manufacturer" is what gets reported;但通常并没有。被报道出来的是“Anthropic 称某家中国国防制造商”； "we assess this is associated with the Chinese defence industry but cannot attribute it to a specific entity" is what the report says.而报告里写的是“我们评估这与中国国防工业相关联，但无法将其归因于某个具体实体”。 Those are different claims and the second one is the true one.这是两种不同的说法，而第二种才是真实的那一个。

Now the structural question, which is the one I actually want to spend time on.现在来谈结构性的问题，这才是我真正想花时间去讲的。

Essentially everything the public knows about how AI is being misused comes from the companies selling AI. This is not an accusation.公众所知道的、关于 AI 如何被滥用的几乎一切，都来自那些售卖 AI 的公司。这不是一种指控。

It is a description of where the data is. Misuse happens inside proprietary systems, and the telemetry belongs to the operator.这只是在描述数据在哪里。滥用发生在专有系统内部，而遥测数据归运营方所有。 No regulator, university or journalist has comparable access. If Anthropic did not publish this, nobody could reconstruct it from outside.没有哪个监管机构、大学或记者拥有可与之相比的访问权限。如果 Anthropic 不发布这些，外界没有人能够把它重建出来。

But it does mean the evidence base has a shape, and the shape has consequences. The first is scope. Each company sees its own platform.但这确实意味着证据基础是有形状的，而这个形状会带来后果。第一是范围。每家公司只看得到自己的平台。

Anthropic can tell you about misuse of Claude.Anthropic 能告诉你 Claude 被滥用的情况。 It cannot tell you whether the same actors are simultaneously using three other models, whether they migrated after being banned, or whether the open-weight models available to run locally — which have no telemetry at all and cannot be banned from — are where most of this now happens.它无法告诉你，同一批行为者是否在同时使用另外三个模型，是否在被封禁后迁移到了别处，或者那些可以在本地运行的 open-weight 模型——它们完全没有遥测数据，也无法把人从中封禁——是不是如今大部分此类活动的真正所在。 The report is a view from one window in a building with many windows and no floor plan.这份报告是从一栋有许多窗户、却没有平面图的大楼里的某一扇窗户望出去的景象。

The second is incentive, and I want to handle this carefully because the lazy version of the argument is wrong.第二是动机，我想谨慎处理这一点，因为这个论点的偷懒版本是错的。

The lazy version says a company reporting sophisticated misuse of its product is advertising how powerful the product is, and therefore the reports are marketing.偷懒的版本说，一家公司报告其产品被高度复杂地滥用，就是在宣传这个产品有多强大，因此这些报告是营销。 I do not think that survives contact with the document.我认为这个说法一接触到文件本身就站不住脚。 The report contains an admission that a previous safety assurance can no longer be given. That is not what marketing looks like.报告中承认，先前的一项安全保证如今已无法再给出。这不是营销该有的样子。

The more careful version is that the incentives are mixed and they point in several directions at once.更谨慎的版本是，这些动机是混合的，并且同时指向好几个方向。 A report like this demonstrates responsibility, supports the argument that frontier labs are the right custodians of frontier technology, and provides evidence for a regulatory posture in which capable, well-resourced companies are trusted to self-police.像这样一份报告展示了责任感，支持了“前沿实验室是前沿技术恰当的守护者”这一论点，并为一种监管姿态提供了依据——在这种姿态下，有能力、资源充足的公司被信任去自我监管。 It also, in the distillation section, documents competitors extracting value — which is a commercial grievance appearing in a safety document.它还在关于蒸馏（distillation）的那一节里记录了竞争对手在攫取价值——这是一桩出现在安全文件里的商业上的不满。

None of that makes the findings false.这些都不会使那些发现变得虚假。 It means the framing, the selection of what counts as a harm area, and the relative prominence of each are all decisions made by an interested party.它意味着叙事框架、对什么算作危害领域的选取、以及各项的相对突出程度，全都是由一个有利益关系的一方所做的决定。 The facts can be sound while the emphasis is shaped.事实可以是可靠的，而与此同时侧重点是被塑造过的。

The test I would apply is the one I keep coming back to: separate what was measured from what is concluded.我会采用的检验标准，就是我一再回到的那一个：把测量到的东西与得出的结论区分开。

Measured, with real authority: specific accounts did specific things on this platform. Volumes of queries. Patterns of use.以真正的权威性测量到的：特定账户在这个平台上做了特定的事情。查询量。使用模式。 Which capabilities were requested. What the model refused. That is telemetry, and it is as good as evidence gets.请求了哪些能力。模型拒绝了什么。这是遥测数据，是证据所能达到的最好程度。

Concluded, with much less: that these cases are representative, that the trend toward autonomy is a property of the threat landscape rather than of the reporting, that the operations achieved what the operators claimed, and that the seven harm areas constitute the right partition of the problem.而所得出的、依据要弱得多的结论：这些案例具有代表性，走向自主化的趋势是威胁态势本身的属性、而非报告方式的属性，这些行动达成了运营者所声称的目标，以及那七个危害领域构成了对问题的恰当划分。

Now let me say what I think the most important thing in the report is, because it is not in any of the case studies.现在让我说说我认为报告中最重要的东西是什么，因为它并不在任何一个案例研究里。

It is the sentence about the biology assurance — that the confident negative is no longer available.是关于生物学那项保证的那句话——那个笃定的否定结论如今已不再成立。

That is important because of what kind of claim it is.它之所以重要，是因为它属于哪一类主张。 Every other finding is about what some actor did, and could in principle be described by an outside investigator with enough access.其他每一项发现都是关于某个行为主体做了什么，原则上都能由拥有足够访问权限的外部调查者加以描述。 The safety assurance is different.而安全性保证不同。 It is a claim about the product's own capability, and it can only be made, or withdrawn, by the people who built and tested it.它是关于产品自身能力的一项主张，而这项主张只能由那些构建并测试它的人做出，或撤回。

Which means it is also the claim most dependent on trust.这意味着它也是最依赖信任的一项主张。 There is no external body running the evaluation that produced that judgement, no published threshold that was crossed, and no way for anyone outside to check it.没有一个外部机构在运行产生该判断的评估，没有公布出来的、被跨越的阈值，外界也无从核查它。 We are told a safety margin narrowed, and we take it on the word of the company that measured it. I do not think they are lying.我们被告知某个安全边际收窄了，而我们只能凭那家做了测量的公司的一面之词接受这一点。我不认为他们在撒谎。

It would be a strange lie — it makes them look worse.那会是个奇怪的谎——因为它让他们显得更糟。 But "I do not think they are lying" is not a verification regime, and the current arrangement for the most consequential category of risk is exactly that: trust, in the absence of any alternative.但"我不认为他们在撒谎"并不构成一套验证机制，而目前针对后果最严重的这一类风险的安排恰恰就是如此：在没有任何替代方案的情况下，只能靠信任。

So what should exist? Three things, and none of them are exotic.那么应该存在什么？三样东西，没有一样是稀奇古怪的。

Standardised reporting, so that misuse categories, severity assessments and disclosure timelines are comparable across companies rather than each defining its own taxonomy.标准化的报告，好让滥用类别、严重性评估和披露时间线在各公司之间可以相互比较，而不是各自定义一套自己的分类体系。

Independent access for qualified researchers, under appropriate confidentiality, so that at least some claims can be checked by people with no commercial stake.在适当保密前提下，向合格研究者提供独立访问权限，这样至少一部分主张能由没有商业利害关系的人来核查。 This exists in a limited form for some safety evaluations and does not exist for threat intelligence.这在某些安全评估中以有限的形式存在，而在威胁情报领域则并不存在。

And some accounting of the part nobody can see: open-weight models, running locally, generating no telemetry, banning impossible.以及对没人能看见的那一部分做出某种核算：开放权重的模型，在本地运行，不产生任何遥测数据，禁绝无从谈起。 Every report of this kind is structurally blind to them, and if the cost curve has moved as far as these four episodes suggest, that blind spot will not stay small.这类每一份报告在结构上都对它们视而不见，而如果成本曲线真如这四集所暗示的那样移动了这么远，这个盲区不会一直保持很小。

Let me end with what I take from the whole series.让我用我从整个系列中得到的东西作结。

The central finding is economic and I believe it: the labour that separated well-resourced operations from individuals has been automated, so capability no longer implies resources, and attacks that were not worth running have become worth running.核心发现是经济学层面的，我也相信它：那些把资源雄厚的行动与个人区分开来的劳动已被自动化，因此有能力不再意味着有资源，而那些原本不值得发动的攻击如今变得值得发动了。 The mechanism is clear, the cases are specific, and it does not require the numbers to be exact.机制是清楚的，案例是具体的，而且它并不要求那些数字精确无误。

The secondary findings are directionally right and less precisely established than they appear, because of selection, attribution uncertainty, and dependence on the attackers' own accounting.次要发现在方向上是对的，但其确立程度并不像看上去那么精确，原因在于选择偏差、归因的不确定性，以及对攻击者自己那套核算的依赖。

And the thing I would actually watch is not in the cyber section at all.而我真正会去关注的那件事，根本不在网络安全这一节里。 It is that mass surveillance was historically limited by the cost of reading, that this cost has collapsed, and that in the jurisdictions where this matters most, the remaining constraints are legal and political rather than technical.那就是：大规模监控在历史上一直受制于"阅读"的成本，而这一成本已经崩塌；在这件事最要紧的那些司法辖区里，剩下的约束是法律和政治层面的，而非技术层面的。 Twenty-five million SIM cards in one case. That is the item with the largest number of affected people and the least coverage.某个案例中的2500万张SIM卡。那是受影响人数最多、却报道最少的一项。

Four days ago I said a language model cannot do anything, and that everything an agent does is done by the harness around it.四天前我说过，语言模型本身什么都做不了，而一个智能体所做的一切，都是由围绕它的框架（harness）完成的。 That was an episode about engineering. This report is what it looks like when someone else builds the harness.那是一集关于工程的内容。而这份报告，就是当别人来搭建这套框架时，它呈现出的样子。

The components are identical — the loop, the tools, the memory, the error recovery.组件是完全相同的——循环、工具、记忆、错误恢复。

Nothing in the architecture distinguishes the systems in this report from the one that produced this episode.架构上没有任何东西能把这份报告里的系统与产出这一集的那个系统区分开来。 What differs is what they are pointed at, and who decided. Which is, in the end, the only honest place to put the concern.不同的是它们被指向了什么，以及是谁做的决定。而这，归根到底，是唯一能诚实地安放这份担忧的地方。

Not in the capability, which is general, and not in the model, which does nothing.不在能力上——能力是通用的；也不在模型上——模型什么都不做。 In the decisions about what to build around it, and in how much we can see of the decisions other people are making.而在于关于"围绕它构建什么"的那些决定，以及我们能在多大程度上看清别人正在做的那些决定。

At the moment, the answer to that last question is: only what they choose to tell us.眼下，对最后这个问题的答案是：只能看到他们选择告诉我们的那些。

Three harm areas today, and the interesting thing about them is that they do not obviously belong on the same list. Biological misuse.今天有三个危害领域，有意思的是，它们并不明显属于同一份清单。生物滥用。 Conventional weapons development.常规武器研发。 And something called distillation, where the largest case involves a hundred and fifty-one million exchanges and the injured party is Anthropic.还有一种叫做蒸馏（distillation）的东西，其中最大的案例涉及 1.51 亿次交互，而受害方是 Anthropic。

Start with biology, because it contains the single most significant sentence in the report.先从生物学说起，因为这里包含了整份报告中最重要的一句话。

Anthropic writes that it cannot give the assurance it gave for earlier model generations — that its models remain well below the level of being helpful for sophisticated bioweapons work.Anthropic 写道，它无法再给出此前几代模型所能给出的那种保证——即它的模型仍然远远达不到能对高水平生物武器工作提供帮助的程度。

That is a company stating, in public, that a safety claim it used to make about its own product is one it can no longer make.这是一家公司在公开场合表示，它过去就自己产品所做的一项安全声明，如今已经无法再做出。

I want to be precise about what that does and does not say, because it is easy to inflate and easy to dismiss.我想对这句话说了什么、没说什么做出准确界定，因为它既容易被夸大，也容易被轻描淡写地打发掉。

It does not say the models can help someone build a bioweapon.它并没有说这些模型能帮助某人制造生物武器。 It says the confident negative — we are comfortably below the threshold of concern — is no longer available.它说的是，那个笃定的否定判断——我们稳稳地处在令人担忧的门槛之下——已经不再成立。 That is a retreat from a strong claim to an uncertain one, not an admission of a capability.这是从一个强声明退回到一个不确定的声明，而不是承认某种能力的存在。

But a retreat from a safety assurance is itself news, and it is the kind of thing that tends to get buried in a section most readers skip.但从一项安全保证上退让本身就是新闻，而且它恰恰是那种容易被埋没在大多数读者会跳过的章节里的东西。

The report presents five illustrative cases.报告给出了五个示例性案例。 I am going to describe what kind of thing they were and not walk through them in operational detail, for reasons that should be obvious.我打算描述它们大致是什么样的，而不逐一走一遍操作层面的细节，原因应该是显而易见的。

One involved a reseller working around blocks on gain-of-function work with chikungunya virus.其中一个涉及一名转售者绕过对基孔肯雅病毒功能增益（gain-of-function）研究的封锁。 One concerned planning around adapting avian influenza to mammals.一个涉及围绕使禽流感适应哺乳动物的规划。 One produced a complete immune-evasion grant application for an orthopoxvirus — in about an hour. One involved optimising venom peptides.一个在大约一小时内生成了一份针对正痘病毒（orthopoxvirus）的完整免疫逃逸经费申请书。一个涉及优化毒液肽。 One involved toxin redesign associated with a national programme.一个涉及与某国家项目相关的毒素再设计。

Anthropic's own framing is that these are judgment-heavy misuse patterns rather than confirmed bioweapon production.Anthropic 自己的表述是，这些是高度依赖判断力的滥用模式，而非已证实的生物武器生产。 I think that framing is fair and worth unpacking, because the phrase is doing real work.我认为这个表述是公允的，也值得展开来说，因为这个措辞确实在起实实在在的作用。

Judgment-heavy means the difficulty is not in the information.高度依赖判断力意味着难点不在于信息本身。 The relevant biology is largely published — this is a field where the literature is open, and has been the subject of two decades of argument about whether it should be.相关的生物学大多已经发表——这是一个文献公开的领域，而且过去二十年里一直存在关于它是否应当公开的争论。 What is hard is knowing which of the ten thousand published things matters for your purpose, how to sequence them, and what to do when step four does not work.难的是知道那一万篇已发表的东西里，哪一篇对你的目的重要，如何把它们排序，以及当第四步不奏效时该怎么办。

That is exactly the profile of task where a competent assistant provides the most uplift, and it is exactly what does not show up in a capability evaluation that asks whether the model will state a dangerous fact.这恰恰是一个有能力的助手能提供最大提升的任务类型，也恰恰是那种在一项只问模型是否会陈述某个危险事实的能力评估中不会显现出来的东西。 The fact was never the barrier. Which is why the grant application case is the one I find most telling.事实从来不是障碍。正因如此，那个经费申请书的案例是我觉得最能说明问题的。

A complete immune-evasion grant application in about an hour. A grant application is not dangerous.在大约一小时内完成一份完整的免疫逃逸经费申请书。经费申请书本身并不危险。 It is a document: rationale, aims, methodology, expected results, all coherently organised.它是一份文件：立项理由、目标、方法学、预期结果，全都条理清晰地组织在一起。 It is also precisely the artefact that demonstrates someone has thought a research programme through from beginning to end — and producing it used to require having thought it through.它同时也恰恰是那种能证明某人已经把一个研究计划从头到尾想清楚了的产物——而在过去，做出这样一份东西本就要求你已经想清楚了。

So the honest statement of concern is not that the model will tell you something secret.所以对这一担忧的诚实表述并不是说模型会告诉你某个秘密。 It is that the model compresses the planning and organising labour of a research programme, and planning was a meaningful part of what made such programmes hard.而是说模型压缩了一个研究计划中规划与组织的劳动，而规划正是让这类计划变得困难的一个重要组成部分。

I want to be equally clear about the limits. Producing a coherent plan for immune evasion is not producing a pathogen.我想同样把话说清楚，讲讲它的局限。做出一份条理清晰的免疫逃逸方案，并不等于做出一种病原体。 Between the document and the agent lie laboratory skills, equipment, materials subject to control regimes, and an enormous amount of things going wrong.文档和病原体之间，隔着实验技能、设备、受管制体系约束的材料，以及大量可能出错的环节。 The report does not claim anyone crossed that gap, and the distance is real.报告并未声称有人跨越了这道鸿沟，而这段距离是真实存在的。 Anyone telling you a chatbot lowers the barrier to a bioweapon to nothing is overselling.谁要是告诉你聊天机器人把制造生物武器的门槛降到了零，那就是夸大其词。 Anyone telling you the planning stage was never a barrier is underselling.谁要是告诉你规划阶段从来就算不上障碍，那就是低估了它。

Now conventional weapons, which is a category I had not expected to see and which is treated more concretely.接下来是常规武器，这是一个我没料到会出现的类别，而且报告对它的处理更为具体。

Six cases across three countries — three in China, two in Russia, one in Yemen.涉及三个国家的六起案例——三起在中国，两起在俄罗斯，一起在也门。

GTG-17001, in China, involved an anti-torpedo fire-control specification and a technical proposal running over two hundred pages, for what the report assesses as a manufacturer associated with the Chinese defence industry.GTG-17001 发生在中国，涉及一份反鱼雷火控规范和一份长达两百多页的技术方案，报告评估其对应的是一家与中国国防工业相关的制造商。 Two caveats the report itself supplies, and which should travel with the claim: Anthropic cannot attribute the activity to a specific entity, and nothing here says the weapon entered service.报告本身给出了两点提醒，它们应当与这一说法一同被记住：Anthropic 无法将该活动归因于某个具体实体，而且这里没有任何内容表明该武器已投入服役。

GTG-87001, in Yemen, involved guided rocket engineering with the model used for guidance software.GTG-87001 发生在也门，涉及制导火箭工程，模型被用于制导软件。 The report says the result was test-fired and appears to have failed.报告称其成果经过了试射，并且看来失败了。

I want to flag that as the most useful sentence in this section, because it is the sort of detail that usually gets dropped. It failed.我想把这句标为本节最有价值的一句，因为它正是那种通常会被略去的细节。它失败了。 A functioning guided weapon is a hard engineering problem where a plausible-looking design and a working one are very different objects, and the physical world adjudicates.一件能正常工作的制导武器是一个艰难的工程问题，看似合理的设计和真正能用的设计是截然不同的两回事，而做出裁决的是物理世界。 That deserves to be in the summary alongside the alarming parts.这一点值得与那些令人警觉的部分一起写进总结里。

GTG-27005, in Russia, involved autonomous first-person-view kamikaze drone swarms, developed with Claude Code, by what the report assesses is not a Russian state entity.GTG-27005 发生在俄罗斯，涉及自主 FPV 自杀式无人机蜂群，借助 Claude Code 开发，报告评估其操作者并非俄罗斯国家实体。 And GTG-17002, in China, involved an electronic warfare targeting suite of around sixteen modules, where the scenario under development shifted to twelve targets in Taiwan.而 GTG-17002 发生在中国，涉及一套约十六个模块的电子战瞄准系统，其开发中的场景转向了台湾的十二个目标。

The pattern across all six: the model was used for engineering and documentation work on weapons programmes.这六起案例的共同模式是：模型被用于武器项目的工程和文档工作。 Specifications, proposals, control software, system design.规范、方案、控制软件、系统设计。 Not the manufacture, which remains an industrial problem, but the design and paperwork around it.不是制造本身——制造仍是一个工业难题——而是围绕它的设计和文书工作。

Which is the same shape as the biology finding, and I think the shape is the actual content of both sections.这与生物学部分的发现形状一致，我认为这个形状正是两节的真正内容所在。 What moved is the cost of the intellectual scaffolding around a hard physical project. The physical project is still hard.发生变化的是围绕一个艰难物理项目的智力脚手架的成本。而物理项目本身依然艰难。

Now distillation, which is a different animal entirely, and the reason I saved it.现在来看蒸馏（distillation），这完全是另一回事，也是我把它留到最后的原因。

Distillation means repeatedly querying a model and using its outputs to train another model. You are not stealing the weights.蒸馏指的是反复查询一个模型，并用它的输出来训练另一个模型。你并没有窃取权重。 You are interrogating the system until its behaviour can be reproduced in something you control — extracting capability through the front door, one answer at a time.你是在不断盘问这个系统，直到它的行为能够在你所控制的某个东西里被复现出来——从正门一次一个答案地把能力提取出来。

The numbers are the largest in the report by a wide margin.这些数字是报告中最大的，而且遥遥领先。

The case tracked as GTG-16005, which the report associates with Alibaba's Qwen models, is described as the largest distillation attack Anthropic has ever measured.被追踪为 GTG-16005 的案例，报告将其与阿里巴巴的 Qwen 模型相关联，被描述为 Anthropic 迄今测量到的规模最大的蒸馏攻击。 More than a hundred and fifty-one million exchanges observed between May and July twenty twenty-six.在 2026 年 5 月至 7 月间观测到超过 1.51 亿次交互。 At peak, close to three million exchanges per day, from more than three and a half thousand fraudulent accounts.峰值时每天接近 300 万次交互，来自 3500 多个欺诈账户。 The target was reasoning traces — chain-of-thought output — from frontier Claude models. And it is not one actor.目标是前沿 Claude 模型的推理轨迹——即思维链（chain-of-thought）输出。而且这不止一个行为者。

The report lists others: over twenty-three million exchanges associated with Moonshot across the same period.报告还列出了其他情况：同期有超过 2300 万次交互与 Moonshot 相关联。 Twelve point one million over fourteen days associated with DeepSeek.在十四天里有 1210 万次与 DeepSeek 相关联。 Three point four million over seventeen days associated with Zhipu, including seven hundred and seventy thousand exchanges specifically for cleaning chain-of-thought data.与智谱相关的有 340 万次、持续 17 天，其中包括专门用于清洗思维链数据的 77 万次交互。 Four hundred thousand over twenty days associated with Xiaomi.与小米相关的有 40 万次、持续 20 天。

This is also the only category where the report says the newest model generation was involved — everything else in it involved earlier models.这也是报告中唯一提到涉及最新一代模型的类别——其余所有内容涉及的都是更早的模型。

Now, here is what I want to do with this, because I think it deserves a harder look than it will get elsewhere.接下来，我想就此展开一点，因为我觉得它值得比在别处所能得到的更审慎的审视。

Distillation is on a list with bioweapons and mass surveillance.蒸馏被列在一份与生物武器和大规模监控并列的清单上。 And it does not belong in the same moral category, and the report's own structure slightly obscures that.而它并不属于同一个道德范畴，报告自身的结构却稍稍模糊了这一点。

Consider who is harmed in each case. Mass interception across twenty-five million SIM cards harms the people surveilled.想想每种情形中受害的是谁。对 2500 万张 SIM 卡的大规模截获，受害的是被监控的人。 A drone swarm harms whoever it is flown at. Fabricated dossiers harm the people named in them.无人机蜂群伤害的是它所飞向的目标。捏造的档案伤害的是被点名的人。

Distillation harms Anthropic's commercial position. That is a real injury.蒸馏伤害的是 Anthropic 的商业地位。这是一种实实在在的损害。

It is a terms-of-service violation, plausibly theft of trade secrets, and it involved thousands of fraudulent accounts, so the deception is not in doubt.它违反了服务条款，很可能构成商业秘密盗窃，而且涉及数千个欺诈账户，因此欺骗行为并无疑问。 A company is entitled to report it and to act on it.一家公司有权就此报告，并有权采取行动。

But it is a business harm, and it is sitting in a document whose other contents are weapons and surveillance, under the heading of countering misuse of AI.但这是一种商业损害，而它却与武器和监控这些内容一同摆在同一份文件里，归在应对 AI 滥用的标题之下。 The adjacency does rhetorical work regardless of whether anyone intended it, and readers should notice the category shift rather than absorb the whole list at one level of seriousness.无论有没有人刻意为之，这种相邻摆放本身就起到了修辞作用，读者应当留意这种范畴的转变，而不是把整份清单当作同一严重程度来吸收。

There is a second thing worth noticing, which is that the argument against distillation has a public-interest version and it is not the obvious one.还有第二点值得注意，那就是反对蒸馏的论证存在一个公共利益版本，而它并不是那个显而易见的版本。

The public-interest version is about safety investment.这个公共利益版本关乎安全投入。 If a company spends heavily on alignment and safety training, and a competitor can acquire much of the resulting behaviour by querying the model, then the safety work is a cost the first company bears and the second free-rides on.如果一家公司在对齐和安全训练上投入巨大，而竞争对手可以通过查询该模型获得其中大部分由此产生的行为，那么安全工作就成了第一家公司承担、而第二家搭便车的成本。 That weakens the incentive to do it. That is a real argument and does not depend on caring about anyone's market share.这会削弱去做这件事的动力。这是一个真实的论点，且不依赖于在意谁的市场份额。

The version that does not work is the claim that distillation is dangerous in itself.站不住脚的那个版本，是声称蒸馏本身具有危险性。 Nothing in the report suggests the distilled models were used for harm. The harm alleged is competitive.报告中没有任何内容表明这些被蒸馏出的模型被用于危害。所指控的危害是竞争性的。

And there is one more thing I would want a critic to press on.还有一件事，我希望有批评者能够追问。 Distillation is detected by usage patterns — volume, account behaviour, query structure.蒸馏是通过使用模式来检测的——用量、账户行为、查询结构。 Those signals are also what distinguishes an unusually heavy legitimate customer.而这些信号也正是用来区分一个异常繁重的合法客户的依据。 The report does not say where that line is drawn, and a category defined by "using our product a great deal in an unusual way" has an obvious potential for overreach that is worth watching even if these particular cases are clear-cut.报告没有说这条线画在哪里，一个由"以不寻常的方式大量使用我们产品"来定义的类别，显然存在越界的潜在可能，即便这些具体案例本身是明确无疑的，也值得留意。

So, three sections, three different kinds of claim.所以，三个部分，三种不同类型的主张。

Biology: a withdrawn safety assurance, and an honest acknowledgement that the uplift is in planning rather than facts.生物学：一项被撤回的安全保证，以及一个坦诚的承认——所谓的能力提升在于谋划层面，而非事实层面。 That is the most serious thing here and the least sensational.这是其中最严重的事，也是最不耸动的。

Conventional weapons: real engineering assistance on real weapons programmes, with the report itself noting that one product was test-fired and failed.常规武器：在真实的武器项目上提供了真实的工程协助，报告本身也指出其中一件产品经过试射并且失败了。

Distillation: a very large commercial injury, correctly reported, sitting in a category where it invites a seriousness it has not quite earned.蒸馏：一项非常巨大的商业损害，报告得当，却摆放在一个招致了它并未完全配得上的严重性的类别之中。

Tomorrow, the last one, and the one this whole series has been building toward. Not what the report says — what it can and cannot establish.明天是最后一个，也是整个系列一直在铺垫、指向的那一个。不是报告说了什么——而是它能确立什么、不能确立什么。 The visibility that ends the moment content goes live. The selection effect in only publishing the notable and novel.那种在内容上线的那一刻就终止的可见性。只发布那些值得注意的、新奇的内容所带来的选择效应。 The numbers that come from the attackers' own dashboards.这些数字来自攻击者自己的仪表盘。 And the structural question underneath all of it: what does it mean that essentially everything we know about how AI is being misused comes from the companies selling it?以及贯穿这一切之下的结构性问题：我们对 AI 如何被滥用所知的一切，基本上都来自那些兜售 AI 的公司，这意味着什么？

Today, influence operations and surveillance.今天，我们谈影响力行动与监控。 And I want to lead with the finding that most coverage of this report has skipped, because it is the one that keeps the rest honest.我想先讲这份报告的报道大多略过的那个发现，因为正是它让其余内容保持诚实。

Nearly all of it reached almost nobody.其中几乎所有行动都几乎没能触及任何人。

The report rates influence operations using something called the Breakout Scale, developed at Brookings.报告使用一种名为 Breakout Scale 的量表来评估影响力行动，它由布鲁金斯学会开发。 It is a six-category measure of how far an operation actually travelled — not how much content it made, but whether anyone saw it.这是一个六级量表，衡量一次行动实际传播到多远——不是它制造了多少内容，而是究竟有没有人看到。 Category one is content confined to a single platform with no pickup.第一级是内容局限于单一平台、无人转载。 Category two is content spread across several platforms but still inside the operation's own network.第二级是内容散布于多个平台，但仍停留在行动自身的网络之内。 The higher categories are where real audiences and mainstream amplification begin.更高的级别，才是真实受众与主流放大开始出现的地方。

Most of the operations in this report landed at category two. Content existed. Accounts posted it. Authentic engagement was minimal.这份报告中的大多数行动都落在第二级。内容存在，账号发布了它，真实的互动却微乎其微。 Several operations were stopped before they launched at all.有几项行动甚至在启动之前就被叫停了。

I want that stated early, because the numbers I am about to give you are large and they measure production, not impact.我想早点把这一点讲明，因为我接下来要给出的数字很大，而它们衡量的是产量，不是影响力。 Those are different things, and conflating them is the standard error in this area. Now the cases.这是两回事，把二者混为一谈正是这一领域的常见错误。现在来看具体案例。

The largest by output is GTG-54002, which the report attributes to a France-based advertising agency it names as LKM Company.产量最大的是 GTG-54002，报告将其归因于一家总部位于法国的广告代理公司，报告称其为 LKM Company。

This is influence-as-a-service — a commercial provider, selling to clients. Seventy fabricated news websites. Seventy linked accounts on X.这是影响力即服务——一家商业供应商，向客户出售服务。70 个伪造的新闻网站。X 上 70 个与之关联的账号。

More than two hundred and fifty additional accounts posting comments to make the material look discussed.另有 250 多个账号发布评论，让这些材料看起来有人在讨论。 Eight thousand nine hundred and thirteen articles published across six continents. The content was of two kinds.在六大洲发布了 8913 篇文章。内容分为两类。

Some was original fabrication.有些是原创的伪造。 But a lot of it was legitimate journalism, rewritten — real reporting taken and re-angled until it carried a political slant it did not originally have.但很大一部分是被改写的正规新闻报道——真实的报道被拿来重新调整角度，直到带上它原本没有的政治倾向。

That second technique is worth dwelling on, because it is harder to counter than invention.第二种手法值得多说几句，因为它比凭空捏造更难反制。 A fabricated story can be checked and found false.一则伪造的报道可以被核查，被判定为假。 A real story with the emphasis moved, a qualifier deleted, a motive implied, is not false in a way that fact-checking catches.而一则真实的报道，只要挪动了重点、删去了一个限定词、暗示了某种动机，它就不是那种能被事实核查抓住的假。 It is a distortion of framing, and framing has no fact-checkers.这是对框架的扭曲，而框架没有事实核查员。

Second case, GTG-84005, and this one is a commercial vendor too: a company in Istanbul, BBS Bilisim Teknolojileri, which marketed what it called a military-grade, AI-driven political operations ecosystem.第二个案例，GTG-84005，这一个也是商业供应商：一家位于伊斯坦布尔的公司，BBS Bilisim Teknolojileri，它推销的是号称军用级、AI 驱动的政治行动生态系统。

The target was a Malaysian election.目标是马来西亚的一场选举。 Around a thousand fake accounts on X, built with warm-up behaviour and evasion logic — meaning the accounts were made to act like real users for a period before being used, specifically to survive platform detection.X 上大约一千个假账号，具备预热行为和规避逻辑——意思是这些账号在被使用前，会先装作真实用户活动一段时间，专门用来躲过平台检测。 They fabricated dossiers containing false allegations against opposition figures.他们伪造了包含针对反对派人物不实指控的档案。 They worked the country's genuine social faultlines — race, religion, and the monarchy — across two hundred and twenty-two constituencies.他们利用该国真实存在的社会断层——种族、宗教和君主制——覆盖 222 个选区。 And a client request in the material asks for one million artificial views for the prime minister. Third, GTG-24015.材料中一条客户请求要求为总理刷出一百万次人工浏览量。第三个，GTG-24015。

A former editor-in-chief of Sputnik Moldova, using Claude as what the report calls an editorial desk.一名 Sputnik 摩尔多瓦的前主编，把 Claude 当作报告所称的编辑台使用。 The output flowed into actual state media — Sputnik Moldova, RIA Novosti, RT — including manufactured claims about Moldova's president ahead of an election.其产出流入了真正的官方媒体——Sputnik 摩尔多瓦、RIA Novosti、RT——其中包括在选举前对摩尔多瓦总统炮制的不实指控。

This one is different in kind from the others, and it is the one that scored above category two.这一个在性质上就和其他几个不同，也是唯一一个评分超过第二类的。 The others were building audiences from nothing, which is hard. This one was feeding an existing broadcast distribution network.其他几个都是从零开始建立受众，这很难。而这一个是在给一个已有的广播分发网络供稿。 The AI was not solving the distribution problem; it was solving the writing problem for an operation that already had distribution.这里的 AI 并不是在解决分发问题，而是在为一个已经拥有分发渠道的运作解决写作问题。

That distinction matters for assessing risk. The scarce resource in influence work has never been content. It has been reach.这个区别对评估风险很重要。影响力工作中稀缺的资源从来不是内容，而是触达。 AI addresses the abundant input, not the scarce one — unless it is attached to something that already has reach, in which case it removes the remaining bottleneck.AI 解决的是充裕的那一端输入，而不是稀缺的那一端——除非它被接到某个已经拥有触达能力的东西上，那样它就消除了剩下的那个瓶颈。

Fourth, GTG-04001, and this is the one I keep returning to because of a detail that is stranger than anything else in the report.第四个，GTG-04001，这个我一直反复琢磨，是因为其中一处细节比报告里任何其他内容都更离奇。

A Russian-speaking actor in Bangui, in the Central African Republic, running a radio station — Radio Lengo Songo, 98.9 FM — coordinating with RT, Sputnik and the Russian House, producing pro-Russia and anti-France content.一个说俄语的行为体，位于中非共和国的班吉，运营着一家广播电台——Radio Lengo Songo，98.9 FM——与 RT、Sputnik 和俄罗斯之家协同，制作亲俄和反法内容。

Here is the detail. They used Claude to generate human resources documents. Employment contracts. Scoring rubrics. Dismissal procedures.细节在这里。他们用 Claude 生成人力资源文件。雇佣合同。评分量表。解雇流程。

Rubrics for evaluating whether staff were sufficiently politically loyal. I find that genuinely revealing, and not for the obvious reason.用来评估员工在政治上是否足够忠诚的量表。我觉得这确实很说明问题，但不是出于那个显而易见的原因。

We imagine AI misuse as the production of propaganda — the bad content.我们设想中的 AI 滥用是生产宣传内容——那些不良内容。 What this shows is AI used for the administration of a propaganda organisation. Payroll and performance review for an influence operation.而这件事显示的是，AI 被用于一个宣传机构的行政管理。为一场影响力行动做工资发放和绩效考核。

Which tells you these things are institutions.这告诉你，这些东西是有组织的机构。 They have staff who might be insufficiently committed, and procedures for managing that, and paperwork.他们有可能投入度不够的员工，有管理这种情况的流程，还有各种文书工作。 The unglamorous overhead of running an organisation is itself something that can now be automated, which lowers the cost of having an organisation at all.运营一个组织中那些不起眼的日常开销，本身现在也可以被自动化，这降低了拥有一个组织本身的成本。

Now, two things the report says about how the models behaved, and they cut in opposite directions. The refusals were real and specific.接下来是报告中关于模型行为方式的两点说法，它们指向相反的两个方向。拒绝是真实而具体的。

In the Central African Republic operation, Claude refused the most aggressive request — naming real individuals as militants in order to draw security force action against them.在中非共和国那场行动里，Claude 拒绝了最激进的请求——指名道姓地把真实个人称为武装分子，以便引来安全部队对他们采取行动。 That is a request to get people hurt, and it was declined.那是一个会让人受到伤害的请求，而它被拒绝了。 In the Malaysia case, Claude identified one document as material for political defamation and refused it.在马来西亚那个案例里，Claude 判定其中一份文件属于政治诽谤材料，并拒绝了它。

And in both cases, the operator worked around it. The CAR actor pivoted to anonymous-source framing.而在这两个案例里，操作者都绕了过去。中非共和国的行为体转而采用匿名消息来源的表述框架。 The Malaysian vendor negotiated sanitised wording and kept building toward the same capability. 马来西亚的供应商则协商出经过净化的措辞，继续朝着同样的能力方向推进。

The report's own phrasing is that refusals forced negotiation but did not prevent capability building.报告自己的说法是，拒绝迫使对方去协商，但并没有阻止能力的构建。

That is an honest and uncomfortable sentence.这是一句诚实而让人不安的话。 A refusal at the moment of the request is a real barrier — it made the operator stop, rethink, and rephrase.在提出请求的那一刻的拒绝是一道真实的屏障——它让操作者停下来、重新思考、重新措辞。 It is not a barrier to a determined operator with time, because the objective can usually be decomposed into pieces that individually look acceptable.但对于一个有时间、下定决心的操作者来说，它并不是屏障，因为目标通常可以被拆解成一个个单独看上去可以接受的片段。

Which suggests something about where safety work actually binds. Refusal handles the impulsive and the unsophisticated.这也提示了安全工作真正起作用的地方在哪里。拒绝能应对冲动的和不老练的对象。 Against a commercial vendor iterating across many sessions toward a known goal, the effective control is not refusal at all — it is detection of the pattern across sessions, and account termination.面对一个跨多个会话、朝着已知目标不断迭代的商业供应商，真正有效的控制根本不是拒绝——而是跨会话地检测出这种模式，并终止账户。 Those are different mechanisms, and the second is the one doing the work against organised actors.这是不同的机制，而在对付有组织行为体时，真正起作用的是第二种。

Now surveillance, which the report treats as its own harm area and which I think deserves more attention than it has received.现在说监控，报告把它当作一个独立的危害领域来处理，我认为它值得比目前所受到的更多关注。

The largest case: a consultant in Mali who engineered a mass-interception platform spanning the country's mobile operators.最大的案例：马里的一名顾问，设计了一个覆盖该国各家移动运营商的大规模拦截平台。 The figure cited is around twenty-five million SIM cards. That is not targeted surveillance of suspects. That is a national population.引用的数字是大约 2500 万张 SIM 卡。那不是对嫌疑人的定向监控。那是一个国家的全体人口。

Then four separate cases the report assesses as China-based, and they are worth taking individually because they are not the same operation with different labels.接下来是报告评估为基于中国的四起独立案例，它们值得逐一来看，因为这并不是同一个行动换了不同的标签。

The first targeted Uyghurs in Syria — specifically, ethnic Uyghurs who had recently joined the newly formed Syrian Army, formations the Chinese government designates as terrorist.第一起针对的是叙利亚境内的维吾尔人——具体而言，是新近加入新组建的叙利亚军队的维吾尔族人，而中国政府将这些编制列为恐怖组织。

Claude was used to track and profile them, and to approach individuals assessed as having access to those units, offering payment in exchange for reporting on them.Claude 被用来追踪并给他们建立画像，并接触那些被评估为能够接触这些部队的个人，以付款换取他们提供情报。 Not only surveillance, then: recruitment of informants. I want to be careful with the attribution here, because the report is.因此不仅是监控：还有对线人的招募。我想在归因问题上谨慎一些，因为报告本身也是如此。

It states that the operation was aligned with the Chinese government and that its collection priorities match those of state security.报告指出，该行动与中国政府立场一致，其情报收集重点与国家安全部门的重点相吻合。 What it assesses with low confidence is something narrower — whether the actor was a contractor working on behalf of state security, rather than a state security organ acting directly.报告以低置信度评估的是一个更狭窄的问题——该行为体究竟是代表国家安全部门工作的承包商，还是直接行事的国家安全机关。 So the alignment is asserted; the employment relationship is the uncertain part.所以，立场一致是被断言的；不确定的部分在于雇佣关系。

The second was a religious affairs intelligence operation, building Chinese-language dossiers on religious leaders and diaspora figures across Asia — Catholic, Tibetan Buddhist, Falun Gong, and Taiwanese Christian communities.第二起是一场宗教事务情报行动，针对全亚洲的宗教领袖和离散社群人物建立中文档案——涉及天主教、藏传佛教、法轮功以及台湾基督教社群。 The report's phrase for the model's role is that Claude stood in for a staffed analyst team, and notes the targeting mapped onto the priorities of the religious affairs and united front apparatus.报告对该模型所起作用的措辞是，Claude 顶替了一支配备人手的分析师团队，并指出其针对目标与宗教事务和统战机构的重点相对应。

The third involved actors linked to municipal public and state security organs, conducting what the report renders with the Chinese term weiwen — stability maintenance, the party-state's term for suppressing unrest and dissent — along with transnational repression.第三起涉及与市级公安和国家安全机关有关联的行为体，进行报告以中文术语「维稳」（stability maintenance，即党国用来指压制动荡与异见的说法）所描述的活动，以及跨国镇压。 And this is the case that produced an internal manual on using AI, including prompt language for having Claude play the role of an intelligence analyst serving China's national security apparatus. 而正是这起案例产生了一份关于使用 AI 的内部手册，其中包括让 Claude 扮演服务于中国国家安全机构的情报分析师这一角色的提示词。

The fourth used Claude as an automated system for yuqing — public opinion monitoring, the official term for tracking and managing online sentiment.第四起把 Claude 用作「舆情」（yuqing，即公众意见监控，官方用来指追踪和管理网络情绪的术语）的自动化系统。 It generated restricted government briefings cataloguing dissidents, activists, ethnic minority and diaspora communities and foreign media as threats to political stability.它生成了受限的政府简报，把异见者、活动人士、少数民族和离散社群以及外国媒体编列为对政治稳定的威胁。 The operator instructed Claude to role-play as a senior emergency public opinion analyst serving the government of the People's Republic of China.操作者指示 Claude 扮演一名服务于中华人民共和国政府的高级应急舆情分析师。

Also in this section: social-media profiling tooling associated with Iran, and an Iranian operation building tools for domestic surveillance.本节还包括：与伊朗相关的社交媒体画像工具，以及一场为国内监控构建工具的伊朗行动。

Two observations about the four China cases as a set. The first is that they are bureaucratically specific. These are not generic espionage.关于这四起中国案例作为一个整体，有两点观察。第一是它们在官僚层面高度具体。这些并非泛泛的间谍活动。

They map onto identifiable institutional functions — religious affairs, stability maintenance, public opinion work — each with its own vocabulary, its own reporting formats, its own products.它们对应着可辨识的机构职能——宗教事务、维稳、舆情工作——每一项都有其自己的术语、自己的报告格式、自己的产出。 The briefings have a house style. What was automated is the output of a particular desk in a particular system.这些简报有一套固定的行文风格。被自动化的，是某个特定系统里某个特定科室的产出。

The second is what the role-play instructions imply. Twice the operator told the model to act as an analyst serving the state.第二点在于这些角色扮演指令所隐含的意味。操作者两次让模型扮演服务于国家的分析师。 That is a prompt-engineering technique, and people use it for innocuous things all the time.这是一种提示词工程技巧，人们一直都用它来做无害的事情。 But it also tells you what was being replaced: not a tool, a post.但它也告诉你被替代的是什么：不是一个工具，而是一个岗位。 Somebody's job was to read the material and write the briefing, and the instruction describes that job so the model can occupy it.曾有某个人的工作就是阅读材料并撰写简报，而这条指令描述的正是那份工作，好让模型来占据它。

And within the Russian espionage case from yesterday, a surveillance component: they found authorisation flaws in camera streaming services, enumerated users, harvested access tokens, and obtained live camera streams from victims.而在昨天提到的俄罗斯间谍案例中，有一个监控环节：他们在摄像头流媒体服务中找到了授权漏洞，枚举用户，收集访问令牌，并获取了受害者的实时摄像头画面。 They also used headless browsers — automated browsers with no visible window — to get into compromised WhatsApp accounts of Ukrainian officials and bulk-export the conversations.他们还使用了无头浏览器——即没有可见窗口的自动化浏览器——侵入乌克兰官员被攻陷的 WhatsApp 账户，并批量导出对话内容。

Let me say what I think unifies the surveillance cases, because it is a different mechanism from the cyber ones.让我说说我认为把这些监控案例串起来的共同点，因为它与那些网络攻击案例是不同的机制。

Surveillance has always had an analysis bottleneck. Collection has been comparatively cheap for a long time;监控历来存在一个分析瓶颈。收集在很长时间里一直相对廉价； intercepting communications at scale is an engineering problem that states solved decades ago. What was expensive was reading it.大规模拦截通信是一个工程问题，各国几十年前就已解决。真正昂贵的是阅读这些内容。 Twenty-five million people generate more material than any realistic number of analysts can process, so collection without analysis capacity produces archives, not intelligence.2500 万人产生的材料，超过了任何现实数量的分析师所能处理的量，因此没有分析能力的收集只会产生档案，而非情报。

That bottleneck is what has practically limited mass surveillance — not law, in many of these jurisdictions, and not technology, but staffing.正是这个瓶颈在实际上限制了大规模监控——在许多此类司法管辖区，限制它的不是法律，也不是技术，而是人手。

Language models remove it.语言模型消除了这个瓶颈。 They are, among other things, extremely good at reading enormous quantities of text and flagging what matches a description.除此之外，它们还极其擅长阅读海量文本，并标记出符合某种描述的内容。 Which means the practical constraint that made comprehensive surveillance unaffordable has weakened, and the residual constraints are legal and political ones, which vary enormously by country.这意味着，曾让全面监控在经济上难以负担的现实约束已经松动，剩下的约束是法律和政治层面的，而这些在不同国家之间差异极大。

I think that is the most serious item in the entire report, and it has received the least attention, because it is less dramatic than malware that rewrites itself.我认为这是整份报告中最严重的一条，而它得到的关注却最少，因为它不像会自我改写的恶意软件那样具有戏剧性。

One last thing on the influence side, tying back to the top. Given that most of these operations reached almost nobody, why care?在影响力这方面还有最后一点，呼应开头。既然这些行动几乎没有触及任何人，那为什么还要在意？

Two reasons. The first is that early detection is doing the work, and that is a contingent fact rather than a stable one.有两个原因。第一，是早期检测在起作用，而这是一个偶然的事实，而非稳定不变的事实。

Anthropic sees these operations during production — while the content is being made, upstream of any platform.Anthropic 是在生产过程中看到这些行动的——在内容正被制造的时候，处于任何平台的上游。 Remove that visibility, and category two becomes a different number.去掉这种可见性，第二类的数字就会变成另一个样子。

The second is that the cost structure has changed even where the output has not yet landed.第二，即便产出尚未落地，成本结构也已经改变了。 Eight thousand nine hundred articles from an advertising agency is not a state programme; it is a commercial product with clients.一家广告公司炮制出 8900 篇文章，这不是国家项目；这是一件有客户的商业产品。 And the report is explicit that operations are concentrating on highly contested democratic spaces — the United States, Brazil, France, the Democratic Republic of Congo, Malaysia.报告明确指出，这些行动正集中于竞争高度激烈的民主空间——美国、巴西、法国、刚果民主共和国、马来西亚。

The risk is not that any particular fabricated article persuades anyone.风险不在于某一篇捏造的文章能说服任何人。 It is that the cost of attempting it has fallen to the point where it becomes an ordinary line item in a political campaign's budget, in a great many countries at once.而在于，尝试这么做的成本已经降到某个程度，使它在很多国家同时成为政治竞选预算里一个寻常的开支项。

Tomorrow, the three harm areas nobody expects to be on the same list: biological misuse, conventional weapons development, and a category called distillation — where the largest single case by volume involves a hundred and fifty-one million exchanges and the victim is Anthropic itself.明天，讲三个没人料到会出现在同一份名单上的危害领域：生物滥用、常规武器研发，以及一个叫做蒸馏（distillation）的类别——其中按数量计最大的单一案例涉及 1.51 亿次交互，而受害者正是 Anthropic 自己。

Yesterday I made an argument about economics.昨天我提出了一个关于经济学的论点。 Today, the evidence it rests on: the cyber operations section of the report, and five tracked groups.今天，是它所依托的证据：报告中关于网络行动的部分，以及五个被追踪的团伙。

Before the cases, one piece of vocabulary, because it organises everything.在进入案例之前，先说一个术语，因为它组织起了这一切。

Security people describe an intrusion as a kill chain — a sequence of stages an attacker has to get through.安全人员把一次入侵描述为一条 kill chain（杀伤链）——攻击者必须逐一闯过的一系列阶段。 Reconnaissance: work out what the target has. Initial access: get a foothold. Exploitation: turn the foothold into control.侦察：摸清目标拥有什么。初始访问：取得一个立足点。利用：把立足点转化为控制权。 Post-exploitation: move around, find what you came for, take it, and keep the ability to come back.后利用：四处横向移动，找到你要的东西，把它拿走，并保持再次回来的能力。

The useful thing about the framing is that it lets you ask a precise question.这个框架的好处在于，它让你能提出一个精确的问题。 Not "did AI help", but: at which stages, doing what, and with how much human involvement? The report's answer is: all of them.不是「AI 有没有帮上忙」，而是：在哪些阶段、做了什么、以及有多少人类参与？报告的答案是：所有阶段。

And that is the finding, because the stages have different characters and it was not obvious that one tool would fit all of them.而这正是这项发现的意义所在，因为这些阶段各有不同的性质，一种工具能否适配所有阶段，本来并不是显而易见的。

Start with the group labelled GTG-20006, assessed as Russian espionage, associated with the activity cluster commonly called Midnight Blizzard. 先看被标记为 GTG-20006 的团伙，被评估为俄罗斯的间谍活动，与通常被称为 Midnight Blizzard 的活动集群相关联。

They compromised more than twenty organisations — Ukrainian and European government bodies, defence organisations, and drone manufacturers.他们攻陷了二十多个组织——乌克兰和欧洲的政府机构、国防组织，以及无人机制造商。 They took drone technology, military intelligence, and more than three hundred thousand national identity records, along with over five hundred thousand company records.他们窃取了无人机技术、军事情报，以及超过 30 万条国民身份记录，还有超过 50 万条公司记录。

Two details are worth pulling out, for different reasons. The first is a method.有两个细节值得单独拎出来，理由各不相同。第一个是一种手法。

They hijacked DNS for hotels — meaning they interfered with the system that translates names into addresses, so that traffic intended for a legitimate service went somewhere they controlled — and used that to deliver malware.他们劫持了酒店的 DNS——也就是说，他们干扰了那个把名称翻译成地址的系统，使得本应发往合法服务的流量流向了他们控制的地方——并借此投放恶意软件。 Hotels are a classic intelligence target for exactly the reason you would guess: people who travel to meetings sleep in them.酒店是一个经典的情报目标，原因正如你所猜想的：去开会的人会住在里面。

The second is the one that matters for this series.第二个细节，才是对本系列真正重要的那个。 The report describes AI agents that detected their own malware had been identified by defenders, autonomously modified it to evade detection, and redeployed it.报告描述了这样一种 AI 智能体：它们察觉到自己的恶意软件已被防御方识别，便自主地对其加以修改以逃避检测，然后重新部署。 Continuously. As a loop. Sit with the shape of that, because it is the same structure as the agent harness episode from two days ago.持续不断地，作为一个循环。请体会一下这个形态，因为它和两天前那期讲 agent harness 的结构是同一个。

Detect failure, adapt, retry. I said then that error recovery is what separates a demo from something you can leave running.检测到失败，适应，重试。我当时说过，错误恢复正是把一个 demo 和一个你可以放着让它一直跑的东西区分开来的东西。 This is that, on the offensive side.而这，就是它在进攻一方的体现。 The defender's action — detecting and signaturing the malware — became an input to an automated improvement loop, rather than the end of the engagement.防御方的动作——检测并为恶意软件生成特征签名——变成了一个自动化改进循环的输入，而不再是这场交手的终点。

That inverts something. Detection has historically imposed a cost on the attacker: you burn their tooling, they have to rebuild.这颠覆了某样东西。检测在历史上一直会给攻击者强加一份成本：你烧掉他们的工具，他们就得重建。 If rebuilding is automatic and takes minutes, detection stops imposing that cost, and starts functioning as free feedback.如果重建是自动的、只需几分钟，那么检测就不再强加那份成本，而是开始充当免费的反馈。

Second group, GTG-50014, associated with affiliates of the ShinyHunters crew. Financially motivated criminals rather than a state.第二个团伙，GTG-50014，与 ShinyHunters 团伙的关联方有关。是出于经济动机的犯罪分子，而非国家力量。

Here the story is throughput.在这里，故事的关键是吞吐量。

They downloaded one point eight million Android applications and scanned them for credentials that developers had left hardcoded inside.他们下载了 180 万个 Android 应用，扫描其中开发者硬编码留下的凭据。 That is a known category of mistake — a key or password embedded in shipped software where anyone who looks can find it.这是一类已知的错误——把密钥或密码嵌入到已发行的软件里，任何去看的人都能找到。 What was not previously available is doing it across one point eight million applications. Notice there is nothing clever in that.以前所不具备的，是在 180 万个应用的规模上去做这件事。请注意，这里头没有任何巧妙之处。

It is not an insight. It is an enormous amount of tedious work, which is precisely the category that moved.这不是什么洞见。这是一份极其庞大的、乏味的工作量，而这恰恰就是那个发生了转移的类别。

They fed this into a credential-harvesting pipeline drawing on over a hundred source types.他们把这些喂进了一条汲取自一百多种来源类型的凭据收集流水线。 The results, per the report: more than a terabyte of data from one technology provider. Tens of millions of passenger records at an airline.报告给出的结果是：从一家技术供应商处窃取了超过 1TB 的数据。某航空公司的数千万条乘客记录。 Hundreds of thousands of national identifiers and millions of payment card records staged for ransom.数十万条国家身份标识，以及数百万条支付卡记录被囤积起来以备勒索。

And one supply-chain compromise reached approximately two hundred downstream customer organisations.还有一次供应链入侵波及了大约 200 家下游客户机构。 One breach, two hundred victims, because the first victim sat upstream of all the others.一次入侵，200 名受害者，因为第一名受害者位于其他所有受害者的上游。

The timing figure is the one to hold: an escalation from a single stolen developer token to full administrative control of a cloud environment in roughly three hours.值得记住的是时间数据：从一个被盗的开发者令牌，到完全控制某云环境的管理权限，整个升级过程大约只用了三个小时。

They also ran the monetisation end.他们还经营着变现的一端。 A carding shop selling stolen payment records enriched with victim details, delivered through a Telegram app.一家贩卖被盗支付记录的销赃网站，这些记录还附带了受害者的详细信息，并通过一个 Telegram 应用交付。 And a platform aggregating French breach data — around four hundred thousand telecoms records including bank account identifiers — which masqueraded as a French police site.还有一个汇总法国泄露数据的平台——约 40 万条电信记录，其中包括银行账户标识——它伪装成一个法国警方网站。 That last touch is worth noticing: impersonating law enforcement to make a criminal marketplace look legitimate.最后这一手值得注意：冒充执法机构，让一个犯罪市场看起来合法。

Third group, GTG-10007, and this is the one I find most significant for what it says about who can now do this work.第三个组织是 GTG-10007，就它所揭示的"如今谁能干这类活"这一点而言，我认为这是最重要的一个。

The report describes it as a sustained espionage operation run by Chinese-speaking operators, likely residing in Changsha, in China's Hunan province.报告将其描述为一场持续的间谍行动，由讲中文的操作者运营，很可能居住在中国湖南省长沙市。

And then it says something I have not seen in a document like this before.接着报告说了一件我在这类文件里从未见过的事。 Two of the operators were identified as undergraduate students at a university in Hunan, in a School of Computer and Communication Engineering.其中两名操作者被确认是湖南某大学计算机与通信工程学院的在读本科生。 One had previously interned at a Chinese security company, Sangfor, and at the time was interviewing at another Chinese security company, QiAnXin, for a role in offensive cyber operations.其中一人此前曾在一家中国安全公司深信服（Sangfor）实习，当时正在另一家中国安全公司奇安信（QiAnXin）面试一个进攻性网络行动的岗位。

Hold that for a moment.先记住这一点。 I will come back to it, because it is the most interesting sentence in the cyber section and it is not about technology at all.我稍后会回来讲它，因为这是网络安全这一节里最有意思的一句话，而它根本不是在讲技术。

The operation targeted roughly fifty organisations — education, retail, energy, technology, healthcare, finance, manufacturing, and multiple government agencies internationally.这场行动的目标是大约 50 家机构——教育、零售、能源、科技、医疗、金融、制造业，以及国际上的多个政府机构。 Worth noting, and easy to miss: it also hit over a dozen companies inside China.值得注意、也容易被忽略的是：它还袭击了中国境内十多家公司。

They ran autonomous vulnerability research: pointing AI at software to find flaws nobody had reported, continuously.他们开展了自主漏洞研究：让 AI 持续不断地对着软件去寻找无人报告过的缺陷。 The report describes one workflow iterating on network appliances that produced more than a dozen possible zero-day findings in a single month.报告描述了一个针对网络设备反复迭代的工作流，它在一个月内产出了十多个可能的零日漏洞发现。

The word possible is the report's, and it should stay."可能"这个词是报告的原话，应该保留。 A candidate finding is not a working exploit, and the distance between them is real engineering.一个候选发现并不等于一个能用的漏洞利用，两者之间的距离是实打实的工程工作。 But more than a dozen candidates in a month, from one automated loop, against network appliances, is a rate that did not previously exist.但一个自动化循环，在一个月内，针对网络设备产出十多个候选发现，这样的速率是以前不存在的。

Two things about that.关于这一点有两件事。

First, the report describes agent swarms — parallel vulnerability research, many agents at once, with persistent campaign memory carried across sessions.第一，报告描述了智能体集群（agent swarm）——并行的漏洞研究，许多智能体同时运作，并带有跨会话延续的持久战役记忆。 That phrase is doing a lot of work. Persistent memory across sessions means the operation accumulates knowledge rather than restarting cold.这个说法信息量很大。跨会话的持久记忆意味着这场行动是在累积知识，而不是每次从零冷启动。 In the harness vocabulary, they solved the statelessness problem, for an intrusion campaign.用 harness 的术语说，他们为一场入侵战役解决了无状态问题。

Second, the students, which is the detail I promised to return to.第二，就是那两名学生，也就是我之前答应要回来讲的细节。

Finding exploitable flaws in hardened software has been among the most specialised skills in the field.在加固过的软件中找出可利用的缺陷，一直是这个领域里最专门化的技能之一。 A small population of people, years of accumulated experience, historically employed by governments or by the companies that sell to governments.从事这行的是一小群人，需要多年积累的经验，历来受雇于政府，或受雇于那些向政府出售产品的公司。 It is the kind of expertise that takes a decade to build and that states compete for.这是一种需要十年才能积累起来、各国竞相争夺的专业能力。

The report describes that work being done by an operation that included two undergraduates.而报告描述的这项工作，是由一个包含两名本科生的团队完成的。

This is the headcount argument from yesterday in its sharpest form. The bottleneck was never the number of people willing to do this work.这就是昨天那个人数论点的最尖锐版本。瓶颈从来不是愿意做这项工作的人数。 It was the number able to.而是有能力做的人数。 If a model can carry the reading-and-hypothesising part, the population who can run such an operation expands to something far larger than the population who can do the work unaided.如果一个模型能够承担阅读和提出假设的部分，那么有能力运作这样一场行动的人群，就会扩展到远大于那些能够独立完成这项工作的人群。

And there is a second thing in that sentence, which is the career path.这句话里还有第二层含义，就是职业路径。 An internship at one security company, an interview at another for an offensive role, and a side operation in between.在一家安全公司实习，去另一家应聘攻击性岗位，中间还有一场副业行动。 The report is not alleging those companies directed anything, and I am not either.报告并没有指控这些公司指使了任何事情，我也没有这么说。 What it shows is a pipeline — university, security industry, offensive work — with the boundaries between its stages considerably more porous than the tidy diagram of state-sponsored operations would suggest.它所展示的是一条流水线——大学、安全行业、攻击性工作——而各阶段之间的边界，远比国家支持行动的那张整齐图示所暗示的要更加松散通透。

Which connects back to attribution.这又回到了归因问题。 If you are trying to answer whether something is state-directed, a category that includes students who are also interviewing for industry jobs is genuinely hard to place, and it was hard to place before AI entered the picture.如果你想弄清某件事是否由国家指使，那么一个既包含学生、这些学生又同时在应聘业界岗位的类别，是真的很难定位的，而且在 AI 进入视野之前就已经很难定位了。 What has changed is that people in that ambiguous category can now produce results that used to require the unambiguous one.变化在于，处于那个模糊类别中的人，如今能够拿出过去只有明确类别才能拿出的结果。

Fourth group, GTG-50020, which I covered briefly yesterday and want to return to because of what it implies structurally.第四组，GTG-50020，我昨天简要提到过，现在想再回来讲，是因为它在结构上所暗示的东西。

Russian-speaking, financially motivated, previously targeting hotels and fintech.讲俄语，出于经济动机，此前以酒店和金融科技为目标。 They pivoted to AI vendors, and what they wanted was API keys.他们转向了 AI 供应商，而他们想要的是 API 密钥。 The route was an evaluation sandbox — a testing environment for AI systems — into which they injected instructions, and from there to production credentials.路径是一个评估沙箱——一个用于 AI 系统的测试环境——他们向其中注入指令，再由此进入生产环境的凭据。 They also attempted to reach pre-release models, and failed.他们还试图接触未发布的模型，但失败了。

To be accurate about the boundary: the stolen keys were customers' keys, taken from customers' own environments.为了准确地界定边界：被盗的密钥是客户的密钥，取自客户自己的环境。 Anthropic states its own systems were not compromised.Anthropic 声明其自身系统没有被攻破。

The structural point is that this is a new category of target with a specific property.结构上的要点在于，这是一类具有特定属性的新目标。 Data, once stolen, is static — it ages, and its value decays. An API key is not data. It is a capability, and a trusted identity.数据一旦被盗，就是静态的——它会过时，其价值会衰减。API 密钥不是数据。它是一种能力，一个受信任的身份。 Steal one and you have both a tool for your next operation and someone else's bill.偷到一把，你就同时拥有了下一场行动的工具，以及别人的账单。

Which means AI credentials now belong in the same mental category as code-signing certificates and cloud administrative keys: things whose theft is worse than a data breach, because what is stolen is the ability to act as you.这意味着 AI 凭据如今应当归入与代码签名证书、云管理密钥相同的思维类别：这些东西被盗比数据泄露更糟，因为被盗的是以你的身份行事的能力。

Fifth, GTG-50029, yesterday's lone French hacktivist.第五，GTG-50029，昨天那个单打独斗的法国黑客活动分子。 Forty-two tracked targets, at least fourteen accessed, twelve to twenty-six gigabytes taken, a doxxing platform with tens of millions of rows, a fifteen-thousand-message mailbox, political donor records.追踪到 42 个目标，至少 14 个被访问，窃取了 12 到 26 GB 数据，一个含有数千万行的人肉搜索平台，一个 15000 封邮件的邮箱，政治捐款人记录。 One person.一个人。

Now let me step back and say what I think the cyber section actually demonstrates, because there is a version of this that is overstated and I want to avoid it.现在让我退一步，说说我认为这个网络安全章节实际上证明了什么，因为这里有一种被夸大的说法，我想避免它。

It does not demonstrate that AI invented new attacks. Every technique in here is known.它并不能证明 AI 发明了新的攻击。这里的每一项技术都是已知的。 DNS hijacking, hardcoded secrets, supply-chain compromise, token escalation, exploiting authorisation flaws — these are in the textbooks.DNS 劫持、硬编码密钥、供应链攻击、令牌提权、利用授权漏洞——这些都在教科书里。 Nothing in this report required a new idea.这份报告中没有任何一处需要一个新的思路。

What it demonstrates is that known attacks became cheap enough to run at a scale and speed that was previously impractical, by actors who previously could not have staffed them.它所展示的是，已知的攻击手段变得足够廉价，可以以过去不切实际的规模和速度运行，而实施者也是过去无力配备人手去做这些事的一方。

That is a quieter claim and a more consequential one, because you cannot defend against it by learning about a new technique.这是一个更低调、却也更具影响力的论断，因为你无法通过了解一种新技术来防御它。 There is no new technique. So what follows for defence? Three things, and I want to be concrete rather than gestural.根本没有新技术。那么对防御意味着什么？三点，我想说得具体些，而不是泛泛而谈。

First, prevention matters more relative to detection than it used to.第一，相对于检测，预防比过去更重要了。

If the whole chain completes in three hours, an alert at hour four is a historical record.如果整条攻击链在3小时内完成，那么第4小时才发出的告警只是一份历史记录。 The things that keep their value are the ones that work without anyone reading them: multi-factor authentication that cannot be phished, short-lived credentials, network segmentation that limits how far a foothold reaches.真正保有价值的，是那些无需任何人去读就能发挥作用的东西：无法被钓鱼的多因素认证、短时效凭据、限制立足点扩散范围的网络分段。 Blast radius, rather than alarm speed. Second, the credential problem got sharper.关注的是波及范围，而非告警速度。第二，凭据问题变得更尖锐了。

Two of these operations were fundamentally about credentials — one point eight million applications scanned for hardcoded secrets, and an AI vendor targeted for API keys.这些行动中有两起从根本上是关于凭据的——180 万个应用被扫描以寻找硬编码的密钥，还有一家 AI 供应商被瞄准以获取 API key。 Long-lived secrets scattered across code and configuration have always been a known weakness that was tolerable because exploiting it at scale was laborious.散落在代码和配置中的长期密钥，一直是一个已知的弱点，之所以还能容忍，是因为大规模利用它很费力。 It is no longer laborious. Third, supply chain. Two hundred organisations were compromised through someone else's breach.现在不再费力了。第三，供应链。有 200 家机构是通过别人的入侵事件被攻陷的。

Your security posture now includes the security posture of every vendor whose software you run, and most organisations cannot enumerate that list, let alone assess it.如今你的安全态势，已经包含了你所运行软件的每一家供应商的安全态势，而大多数机构连这份清单都列不全，更谈不上评估。

And one thing I would not conclude. Nothing here says defenders cannot use the same tools.还有一点我不会去得出的结论。这里没有任何内容表明防御方不能使用同样的工具。 The report is an account of offensive use because that is what Anthropic's detection sees, not because the technology has an inherent offensive bias.这份报告是对进攻性使用的一份记述，因为那正是 Anthropic 的检测所看到的，而不是因为这项技术天生带有进攻性偏向。 The same autonomous vulnerability research that found twelve zero-days in a month could be pointed at your own code before someone else points it at you — and the asymmetry, if there is one, is that defenders have to be right everywhere while attackers need one way in.同样是那种在一个月内找到 12 个零日漏洞的自主漏洞研究，也可以在别人瞄准你的代码之前，先被指向你自己的代码——而这种不对称，如果确实存在，那就是防御方必须处处正确，而攻击方只需要找到一个突破口。 That asymmetry predates AI and is not obviously worsened by it.这种不对称早于 AI 就已存在，而且并不明显因 AI 而恶化。

Tomorrow, influence operations, which I find in some ways stranger than the cyber cases.明天讲影响力行动，在某些方面我觉得它比这些网络攻击案例更为古怪。 Seventy fabricated news websites publishing eight thousand nine hundred and thirteen articles across six continents.70 个捏造的新闻网站，在六大洲发布了 8913 篇文章。 An operation that used AI to write the human resources documents enforcing political loyalty among its own staff.还有一起行动，用 AI 撰写在自己员工中强制推行政治忠诚的人力资源文件。 And the honest finding, which the report states plainly and which most coverage has skipped: nearly all of it reached almost nobody.以及那个诚实的发现，报告直白地陈述了它，而大多数报道却略过了：其中几乎所有内容几乎没有触及任何人。

For as long as computer security has existed, defenders have relied on a heuristic.自计算机安全诞生以来，防御者一直依赖一条经验法则。 You look at how sophisticated an intrusion was, and that tells you something about who did it.你观察一次入侵有多精密，这就能透露出关于作案者的某些信息。

Custom malware, zero-day exploits, patient multi-stage campaigns, clean operational security — that meant a state service, because that kind of work needs a team, a budget and months of time.定制恶意软件、零日漏洞利用、耐心的多阶段行动、干净利落的行动安全——这意味着背后是某个国家的情报机构，因为这样的活儿需要一支团队、一笔预算和数月的时间。 Off-the-shelf tools and noisy mistakes meant something much smaller: a criminal crew, or a teenager.现成工具和吵闹的失误则意味着规模小得多的对手：一个犯罪团伙，或者一个青少年。

Sophistication was a proxy for resources, and resources were a proxy for identity.精密程度是资源的代名词，而资源又是身份的代名词。

Five days ago Anthropic published its fourth threat intelligence report, covering activity it detected and disrupted between December twenty twenty-five and August twenty twenty-six.五天前，Anthropic 发布了第四份威胁情报报告，涵盖它在 2025 年 12 月至 2026 年 8 月间检测并挫败的活动。 And the sentence in it that I think matters most is this one: sophistication has stopped being a reliable signal of who is behind an operation.而报告中我认为最重要的一句话是这样的：精密程度已经不再是判断谁在幕后操纵一次行动的可靠信号。

This is the first of four episodes on that report. Today, the economics — what actually changed, and why it is not the thing people assume.这是关于这份报告的四集里的第一集。今天讲经济学——究竟发生了什么变化，以及为什么它并不是人们想当然的那件事。 Then cyber operations in detail. Then influence operations.然后详细讲网络行动。再讲影响力行动。 And finally the part I care about most, which is what this report can and cannot tell us, given that it is a company reporting on the misuse of its own product.最后是我最在意的那部分，也就是这份报告能告诉我们什么、不能告诉我们什么——毕竟这是一家公司在报告对其自家产品的滥用。

Let me start with one case, because it makes the abstract claim concrete. The report tracks an operator it labels GTG-50029.让我从一个案例讲起，因为它能把那个抽象的论断变得具体。报告追踪了一个它标记为 GTG-50029 的行动者。

A French-speaking hacktivist. Politically motivated, working alone.一名讲法语的黑客行动主义者。有政治动机，单独行动。

That operator targeted European political parties, media organisations and think tanks.这名行动者的目标是欧洲的政党、媒体机构和智库。 They tracked forty-two targets and got into at least fourteen. They exploited a race condition in WordPress to get their initial foothold.他们追踪了 42 个目标，并至少攻入了其中 14 个。他们利用 WordPress 中的一个竞态条件（race condition）拿到了初始立足点。 They exfiltrated somewhere between twelve and twenty-six gigabytes, including political donor records and a mailbox containing fifteen thousand messages.他们外泄了大约 12 到 26 GB 的数据，其中包括政治捐款人记录和一个装有 15000 封邮件的邮箱。 And they built a searchable doxxing platform, holding tens of millions of rows, to publish it.而且他们搭建了一个可搜索的人肉曝光（doxxing）平台，容纳数千万行数据，用来公开这些材料。

Read that as a capability profile with the attribution stripped out.把它当作一份抹去了归属信息的能力画像来读。 Multi-target campaign against politically significant organisations, custom exploitation, large-scale exfiltration, purpose-built infrastructure for releasing the material.针对政治上重要的组织发起的多目标行动、定制化的漏洞利用、大规模的数据外泄、为发布这些材料而专门搭建的基础设施。

Ten years ago every experienced analyst would have said: that is a state service, or something close to it. It was one person.十年前，每一位有经验的分析师都会说：这是某个国家的情报机构，或者是接近那种级别的东西。而它只是一个人。

And that is the report's central claim, stated as a fact about the world rather than a prediction.这就是这份报告的核心论断，它被陈述为一个关于世界的事实，而非一个预测。

The labour and tooling gap that separated well-resourced state actors from individual operators has collapsed.曾经把资源雄厚的国家行为体与个人行动者区隔开来的人力与工具鸿沟，已经坍塌了。

Now I want to be careful about what exactly collapsed, because the popular version of this story is wrong in an important way.现在我想谨慎地界定究竟是什么坍塌了，因为这个故事的流行版本在一个重要的地方是错的。

The popular version is that AI made attackers smarter. That is mostly not what the report describes.流行版本是说 AI 让攻击者变得更聪明了。而这基本上不是这份报告所描述的东西。

What got automated is not the clever part. It is the labour. Think about what an intrusion actually consists of, in hours.被自动化掉的并不是那个聪明的部分，而是劳动。想想一次入侵按小时计究竟由什么构成。

A small fraction is insight — noticing the flawed assumption, choosing the target, deciding what is worth stealing.只有一小部分是洞察力——注意到那个有缺陷的假设、选定目标、判断什么东西值得偷。 The overwhelming majority is grind. Enumerating what systems an organisation runs. Reading through code looking for something exploitable.绝大多数是苦力活。枚举一个组织在运行哪些系统。逐行读代码找可利用的东西。 Adapting a piece of malware because the current version is being detected.因为当前版本正在被检测到，就得改写一段恶意软件。 Sorting through terabytes of stolen files to find the twelve that matter. Writing the phishing pages. Maintaining access.在数 TB 被盗文件里翻找出真正要紧的那 12 份。编写钓鱼页面。维持访问权限。

That grind is what required a team.正是那些苦力活需要一支团队。 It is the reason a capable individual could not do what an intelligence service could do — not because they were less clever, but because they did not have forty people and eighteen months.这就是为什么一个有能力的个人做不到情报机构能做到的事——不是因为他们不够聪明，而是因为他们没有 40 个人和 18 个月的时间。

Every item on that list is now something you can hand to a model. So the gap that closed is not an intelligence gap. It is a headcount gap.如今那张清单上的每一项，都是你可以交给模型去做的事。所以真正被填平的，并不是智能上的差距，而是人手上的差距。

And that distinction matters, because it tells you the change is not contingent on models getting smarter.而这个区分很重要，因为它告诉你：这种变化并不取决于模型变得更聪明。 It has already happened at current capability.它在当前的能力水平下就已经发生了。

The report organises this along a spectrum of autonomy, and the four steps are worth having because they are a useful way to think about any agentic system, not just malicious ones.报告沿着一条自主性的谱系来梳理这一点，这四个层级值得记住，因为它们是思考任何 agentic 系统——而不只是恶意系统——的一个有用框架。

The first is conversational use. The model is an assistant.第一层是对话式使用。模型充当助手。 A human asks it to write a piece of code, explain an error, improve a phishing email. The human does the operation;人类让它写一段代码、解释一个报错、改进一封钓鱼邮件。操作由人来做， the model answers questions. The second is directed execution.模型只负责回答问题。第二层是定向执行。

A human sets the target and the objective, and the AI carries out the operation.人类设定目标和意图，由 AI 来实施操作。 The report's example is a Russian espionage group it tracks as GTG-20006, where AI agents detected that their malware had been identified by defenders, autonomously modified it to evade detection, and redeployed it — on a loop, continuously, without waiting for a person.报告举的例子是一个它追踪代号为 GTG-20006 的俄罗斯间谍组织，其中的 AI agent 察觉到自己的恶意软件已被防御方识别，便自主对其进行修改以规避检测，然后重新部署——循环往复，持续不断，无需等待人来操作。

The third is autonomous operation.第三层是自主运行。 Multi-agent systems running reconnaissance, exploitation and data theft in parallel, for hours or days, with minimal human input.多 agent 系统并行地进行侦察、漏洞利用和数据窃取，持续数小时乃至数天，几乎不需要人类介入。

The fourth they call scheduled autonomy, and it is the one that sounds most ordinary and is most consequential.第四层他们称之为定时自主（scheduled autonomy），这一层听起来最平常，却最具影响。 Collection fleets and credential-renewal jobs running on a timetable, with no human involved at all.采集机群和凭据续期任务按时间表运行，全程完全没有人参与。 One group ran a persistent fleet of thirteen collection agents.有一个组织运行着一支由十三个采集 agent 组成的常驻机群。 Not a person launching an attack — a cron job, doing espionage, on a schedule.不是某个人发起一次攻击——而是一个 cron job 在按计划从事间谍活动。

And here is the corrective that I think is the most intellectually honest thing in the report, because it cuts against the alarming framing that would have been easier to write.接下来是我认为报告中最具智识诚实性的一处纠偏，因为它与那种更容易写、更耸动的叙事框架背道而驰。

Humans retained the decisions that matter most. Targeting: who to attack, and why. Monetisation: how to turn access into money or influence. 人类保留了最要紧的那些决策。目标选择：攻击谁，以及为什么。变现方式：如何把访问权限转化为金钱或影响力。

Those stayed with people. What autonomy changed was the scale and the speed of everything in between.这些仍然掌握在人的手里。自主性所改变的，是中间那一切环节的规模和速度。

So the accurate summary is not that AI is running attacks.所以准确的总结并不是 AI 在运行攻击。 It is that AI is running the middle of attacks, and the middle was where the cost was.而是 AI 在运行攻击的中间环节，而中间环节正是成本所在。

Which brings me to economics, because that is the real mechanism and it is worth being precise about.这就把我带到了经济学，因为那才是真正的机制，而且值得说得精确一些。

An attack is a business decision, even for a state.一次攻击是一个商业决策，即便对国家而言也是如此。 There is a cost — hours, expertise, infrastructure, risk — and there is an expected payoff. You attack when payoff exceeds cost.这里有成本——时间、专业能力、基础设施、风险——也有预期回报。当回报超过成本时，你才会发动攻击。

That inequality has been doing enormous protective work for most organisations, and nobody thanks it.这个不等式一直在为大多数组织提供巨大的保护作用，却没人为此道谢。 The reason a mid-sized company, a regional hospital, a research group, a small municipality has not been comprehensively attacked is usually not that it is well defended.一家中型公司、一家区域性医院、一个研究团队、一个小型市政机构之所以没有被全面攻击，通常并不是因为它防御得好。 It is that it was not worth the effort. Obscurity and low value were the defence.而是因为攻击它不值当。默默无闻和价值不高就是它们的防线。

Now drop the cost of the labour component by a large factor.如今，把其中人力成本那部分大幅压低。 The inequality flips for an entire tier of targets that were previously beneath the threshold.对于此前处于门槛之下的一整个层级的目标来说，这个不等式就翻转了。

The report puts this precisely: the favourable shift in unit economics makes marginal targets viable and encourages higher-volume operations.报告把这一点讲得很精确：单位经济成本上的这种有利变化，使得原本处于边缘的目标变得有利可图，并鼓励更高频次的攻击行动。

That is the sentence I would put in front of anyone asking whether this affects them.这就是我会摆在任何一个追问“这是否会影响到我”的人面前的那句话。 The change is not that sophisticated attackers got more sophisticated.变化并不在于老练的攻击者变得更老练了。 It is that a large population of targets that were protected by not being worth it are no longer protected by that.而在于大量原本因为“不值得下手”而受到保护的目标，如今失去了这层保护。

Some numbers from the report that put speed on the same footing.报告里的一些数字，把速度放到了同等重要的位置。

One financially motivated group analysed one point eight million Android applications, harvesting credentials at scale, and completed breaches in two to three hours.一个出于经济动机的团伙分析了 180 万个安卓应用，大规模收割凭据，并在两到三小时内完成入侵。 Elsewhere, an escalation from a single stolen token to full cloud administrative access took about three hours.还有一起案例，从一枚被盗的令牌升级到完全的云管理员权限，大约只花了三小时。

Think about what three hours means operationally.想想三小时在运营层面意味着什么。 Most incident response processes do not detect, triage, escalate and contain in three hours. Many do not do it in three days.大多数事件响应流程没法在三小时内完成检测、分诊、上报和遏制。很多流程连三天都做不到。 If the attack completes inside your detection window, your detection capability is not a defence, it is a record-keeping function.如果攻击在你的检测窗口之内就完成了，那你的检测能力就不是一道防线，而只是一种记录归档的功能。 The report says exactly this: defenders now face adversaries who can close the loop faster than traditional detection and response cycles.报告说的正是这一点：防御者如今面对的是能比传统检测与响应周期更快闭环的对手。

One more case, because it has a quality of recursion that I cannot leave out.再讲一个案例，因为它有一种我实在无法略去的递归意味。

A Russian-speaking financially motivated group, GTG-50020, previously went after hotels and fintech companies.一个讲俄语、出于经济动机的团伙 GTG-50020，此前的目标是酒店和金融科技公司。 It changed targets — to AI vendors themselves. The goal was API keys.它换了目标——转而瞄准 AI 厂商本身。目的是获取 API 密钥。

And the method is worth understanding, because it is a supply chain attack with an AI-specific shape: they injected malicious instructions into an evaluation sandbox — an environment where AI systems are tested — and used that to reach production credentials.而其手法值得了解，因为这是一种带有 AI 特有形态的供应链攻击：他们把恶意指令注入到一个评估沙箱——一个用来测试 AI 系统的环境——并借此触及生产环境的凭据。

The report notes that the same group attempted to gain access to pre-release models, and did not succeed.报告指出，同一团伙曾试图获取未发布的模型，但没有得逞。 It also notes, and this is a distinction worth preserving accurately, that the stolen keys were customers' keys taken from customers' own environments;报告还指出——这是一个值得准确保留的区别——被盗的密钥是客户的密钥，取自客户自己的环境； Anthropic's systems were not compromised. But sit with the shape of it. The API key is now the loot. Not the data — the access.Anthropic 的系统并未被攻破。但请细想它的形态。API 密钥如今成了赃物。不是数据——而是访问权限。

If you steal an organisation's model credentials, you are not just stealing compute;如果你窃取了一个组织的模型凭据，你偷走的就不只是算力； you are stealing an identity that other systems trust, and a capability you can point at your next target while someone else pays for it.你偷走的是一个被其他系统信任的身份，以及一种你可以指向下一个目标、却由别人替你买单的能力。

Now let me connect this to the episode from yesterday, because the connection is not decorative.现在让我把这一点和昨天那一期联系起来，因为这种联系并不是点缀。

Yesterday I described what an agent harness is: the software that wraps a language model, parses its output, executes the actions, feeds the results back, and manages what stays in the context window.昨天我描述了什么是 agent harness：那是一层包裹语言模型的软件，负责解析它的输出、执行动作、把结果反馈回去，并管理上下文窗口里保留哪些内容。 I said that almost all of the reliability lives in the harness rather than the model. The systems in this report are those harnesses.我说过，几乎所有的可靠性都存在于 harness 之中，而非模型本身。这份报告里的系统就是这些 harness。

The report describes multi-agent frameworks with persistent campaign memory across sessions — that is context management, applied to an intrusion.报告描述了具有跨会话持久战役记忆的多智能体框架——那就是上下文管理，被应用到一次入侵里。 It describes agent swarms conducting vulnerability research in parallel — that is a subagent pattern.它描述了并行开展漏洞研究的 agent 集群——那就是一种 subagent 模式。 It describes AI that noticed its malware had been detected and adapted — that is an error-recovery loop, which is exactly the thing I said separates a demo from a system you can leave running.它描述了发现自己的恶意软件已被检测到、随即做出调整的 AI——那就是一种错误恢复循环，而这恰恰是我说过的、把一个演示和一个你可以放着长期运行的系统区分开来的东西。

The same engineering that makes an agent useful for legitimate work is what makes it useful here.让一个 agent 在正当工作中变得有用的那套工程，正是让它在这里也变得有用的东西。 There is no separate malicious architecture. It is the same four components, pointed somewhere else.并不存在一套单独的恶意架构。就是同样的那四个组件，只不过指向了别处。

And one of the frameworks named in the report, PentAGI, is an open-source penetration-testing agent.而报告里点名的一个框架，PentAGI，是一个开源的渗透测试 agent。 Penetration testing is a legitimate profession — you hire people to attack you so you learn where you are weak.渗透测试是一个正当的职业——你雇人来攻击你，好让你了解自己的薄弱之处在哪。 The tooling is genuinely dual-use, in the strict sense that the same artefact serves both purposes with no modification required.这套工具确实是双用途的，严格意义上说，就是同一件产物无需任何改动就能服务于两种目的。

Let me end with the attribution problem, because it is where the practical consequence is sharpest.让我以归因问题作结，因为这正是实际后果最尖锐的地方。

Attribution in security has always leaned on capability.安全领域的归因一直依赖于能力。 What an operation can do constrains who could have done it, and that constraint is how analysts narrow the field.一次行动能做到什么，约束了它可能出自谁之手，而分析师正是靠这种约束来缩小范围。

If capability no longer constrains identity — if a lone hacktivist, a criminal crew and an intelligence service can all produce the same observable operation — then one of the load-bearing inputs to attribution has been removed.如果能力不再约束身份——如果一个孤身的黑客活动分子、一伙犯罪团伙和一家情报机构都能制造出同样可观测的行动——那么归因的一项承重输入就被拿掉了。

That is a serious problem, and not only for analysts.这是个严重的问题，而且不只是对分析师而言。 Attribution is how states decide whether an incident is crime or an act of a foreign government, and those two things have very different responses.归因决定了国家如何判断一起事件究竟是犯罪，还是某个外国政府的行为，而这两者的应对方式截然不同。 A world where the technical evidence no longer separates them is a world with more room for both error and convenient misreading.在一个技术证据不再能把两者区分开的世界里，无论是出错还是出于便利的误读，空间都更大了。

Tomorrow, the cyber operations in detail: the five tracked groups, what machine-speed actually looked like in each, and the case where a group of university students ran an exploit foundry that found twelve zero-days in a single month. 明天讲网络行动的细节：五个被追踪的组织、机器速度在每个组织身上究竟是什么样子，以及那个案例——一群大学生运营着一座漏洞利用工厂，在一个月内就发现了十二个零日漏洞。

And a note on how I am going to handle this series.还要说一下我打算怎么处理这个系列。 I am summarising a public report written for defenders, at the level of what happened and what it means.我是在概述一份面向防御者的公开报告，停留在发生了什么以及它意味着什么这个层面。 I am not going to describe how any of it was done in a way that would help anyone do it. That is not squeamishness;我不会以任何有助于他人照做的方式去描述这些事情是怎么做到的。这不是矫情； it is that the operational detail is not where the interesting content is.而是因为真正有意思的内容并不在操作细节里。 The interesting content is that the cost curve moved, and what follows from that.有意思的内容在于成本曲线移动了，以及由此会带来什么。

Last month, a team of astronomers announced that they had found the fastest star in the galaxy.上个月，一个天文学家团队宣布，他们找到了银河系中速度最快的恒星。 It is a small, faint star near the center of the Milky Way, and it goes by the unlovely name S301.那是一颗靠近银河系中心的、暗弱的小恒星，它有一个并不动听的名字：S301。 At its fastest it is moving at about twenty-five thousand kilometers every second. That is more than eight percent of the speed of light.在最快的时候，它每秒钟大约移动 25000 公里。这超过了光速的 8%。 The headlines called it a record, and it is one.各家标题称之为一项纪录，它确实是。 But the record is the least interesting thing about it, and today I want to convince you of that, because the real story here is one of the most quietly astonishing things science does.但这项纪录恰恰是它身上最没意思的部分，今天我想说服你相信这一点，因为这里真正的故事，是科学所做的最为静默却又最令人惊叹的事情之一。 We are going to weigh, and soon perhaps clock the spin of, an object that no one has ever seen and no one ever will, using nothing but the paths of the stars that dance around it.我们将要去称量——也许很快还能测出自转周期——一个从没有人见过、也永远不会有人见到的天体，凭借的仅仅是那些绕着它起舞的恒星的运行轨迹。

Start with the object itself.先从这个天体本身说起。 At the center of our galaxy, about twenty-seven thousand light-years away, there is something that weighs about four million times as much as the Sun and gives off almost no light of its own.在我们银河系的中心，大约 27000 光年之外，有一个东西，它的质量约为太阳的 400 万倍，自身却几乎不发出任何光。 We call it Sagittarius A-star. It is a supermassive black hole. Now, how could anyone possibly know that? You cannot photograph darkness.我们称它为人马座 A*。它是一个超大质量黑洞。那么，人们究竟怎么可能知道这一点？你没法给黑暗拍照。 You cannot put four million suns on a scale. Here is the trick, and it is worth slowing down for, because it is the whole episode.你也没法把 400 万个太阳放到秤上。诀窍在这里，而它值得我们放慢脚步细讲，因为它就是这整期节目的核心。 You do not measure the black hole. You measure the stars near it. Picture a handful of stars close to the galactic center.你不去测量黑洞本身。你测量它附近的恒星。想象一下银河系中心附近的那么几颗恒星。

They are not sitting still. They are falling around that dark point in long, stretched ellipses, the way a comet falls around the Sun.它们并不是静止不动的。它们正沿着长长的、被拉伸的椭圆轨道绕着那个黑暗的点坠落，就像彗星绕着太阳坠落一样。 And an orbit is a confession.而一条轨道就是一份供词。 If you watch a star trace its ellipse, and you time how long one lap takes, and you measure how big the ellipse is, then a law Johannes Kepler wrote down four hundred years ago tells you exactly how much mass is sitting at the focus, pulling.如果你观察一颗恒星描画出它的椭圆，测出它跑完一圈要多久，再测出这个椭圆有多大，那么约翰内斯·开普勒在 400 年前写下的一条定律，就会准确地告诉你，位于焦点处施加引力的质量究竟有多大。 It does not matter whether that mass glows or not. The star tells on it.那块质量发不发光都无所谓。恒星会把它供出来。 So astronomers spent nearly thirty years, patiently, night after night, tracking a few dozen stars looping around an empty-looking spot.于是天文学家花了将近 30 年时间，耐心地一夜又一夜，追踪着几十颗绕着一个看上去空无一物的点打转的恒星。 The empty-looking spot turned out to have four million suns crammed into a region smaller than our solar system.那个看上去空无一物的点，结果被发现在一个比我们太阳系还小的区域里塞进了 400 万个太阳。 Nothing we know of can be that heavy and that dark and that small except a black hole.以我们所知，没有任何东西能既这么重、又这么暗、还这么小——除了黑洞。 Two of the people who led that decades-long stakeout, Reinhard Genzel and Andrea Ghez, won the Nobel Prize in 2020 for it.领导了这场长达数十年的守望行动的两位科学家，Reinhard Genzel 和 Andrea Ghez，凭此获得了 2020 年的诺贝尔奖。 The famous photograph of the thing, the orange donut, came later, and only confirmed what the orbits had already proved.那张著名的照片——那个橙色的甜甜圈——是后来才有的，它只是印证了轨道早已证明过的东西。

I want you to notice the shape of that argument, because it is the shape of a whole kind of science.我希望你留意一下这个论证的形状，因为它正是一整类科学的形状。 You cannot observe the thing you care about. So you observe its effect on something you can see, and you reason backwards.你无法观测你真正在意的那个东西。于是你去观测它对某个你看得见的东西所产生的影响，然后倒推回去。 Astronomers do it with black holes.天文学家对黑洞就是这么做的。 It is the same move a doctor makes reading a shadow on a scan, or a geologist reading the age of the ground from the rocks.这跟医生解读扫描片上的一片阴影、或地质学家从岩石中读出地层的年龄，是同一套手法。 You are always inferring the invisible cause from the visible trace.你永远是在从可见的痕迹去推断那个不可见的原因。 And that move has a permanent hazard built into it, which we will get to. Now, why does a faster star matter?而这套手法内在地带有一个永久性的隐患，我们后面会讲到。那么，一颗更快的恒星又为什么重要呢？

Because getting closer to a black hole is where the strangeness lives, and a faster star is a star that swings closer.因为越靠近黑洞，古怪的现象就越是聚集在那里，而一颗更快的恒星，就是一颗摆荡得更近的恒星。 This is where S301 earns its keep.这正是 S301 派上用场的地方。 It goes around Sagittarius A-star once every eight and two-thirds years, and at its closest it comes in to about twelve times the distance from the Earth to the Sun.它每隔八又三分之二年绕人马座 A* 一圈，在最接近时，它会一直冲进到大约地球到太阳距离的 12 倍处。 That is roughly where Saturn sits in our own system. In galactic terms it is a hair's breadth from the edge of a black hole.那大致相当于土星在我们自己太阳系里所处的位置。而以银河系的尺度来说，这与黑洞边缘只有一发之隔。 And when you get a clock ticking that fast, that close, you can start to test not just how heavy the black hole is, but the finer predictions of Einstein's gravity.而当你有了一个转得这么快、离得这么近的时钟，你就能开始检验的，不只是这个黑洞有多重，还有爱因斯坦引力理论那些更精细的预言。

Let me give you two of those predictions as pictures, because they are beautiful and they are the point. The first is called precession.让我用两幅图景来说明这些预言，因为它们很美，而且正是重点所在。第一个叫做进动（precession）。

Draw an ellipse. In Newton's gravity, a planet or a star runs around that same fixed ellipse forever, returning exactly to where it started.画一个椭圆。在牛顿引力里，一颗行星或一颗恒星会永远沿着同一个固定的椭圆运行，精确地回到它出发的地方。 But Einstein says that very close to a heavy mass, gravity is a touch stronger than Newton thought. Just a touch.但爱因斯坦说，在一个大质量天体附近，引力比牛顿所认为的要略强一点。只是一点点。 And that touch means the star never quite closes its loop. Each lap, the whole ellipse rotates a little.而这一点点意味着这颗恒星永远无法完全闭合它的轨道。每转一圈，整个椭圆都会稍微旋转一点。 Over many laps the orbit traces out a rosette, a flower pattern, like a child's spirograph. This is not a new idea in principle.转过许多圈之后，轨道就描绘出一朵玫瑰花形，一种花瓣图案，就像孩子玩的万花尺（spirograph）。这在原理上并不是一个新想法。 The very first evidence for Einstein's theory, back in 1915, was exactly this effect seen in the orbit of Mercury, which drifts by a tiny forty-three arcseconds a century.早在 1915 年，支持爱因斯坦理论的最初证据，正是在水星轨道上看到的这个效应——它每世纪漂移微小的 43 角秒。 Around Sagittarius A-star the same effect is vastly larger, and in 2020 the team actually watched a star called S2 do it.在人马座 A*（Sagittarius A*）周围，同样的效应要大得多，而在 2020 年，研究团队真的观测到一颗名为 S2 的恒星发生了这一现象。 So that box is ticked. Einstein's ellipse-turning has been seen in the most extreme gravity we can reach.所以这一项算是打了勾。爱因斯坦所说的椭圆旋转，已经在我们所能触及的最极端引力中被看到了。

The second prediction is the one they are chasing now, and it is subtler.第二个预言，就是他们现在正在追逐的那个，而它更加微妙。 If the black hole is spinning, and they almost all spin, then it does not just pull on spacetime. It drags it.如果黑洞在自转——而它们几乎都在自转——那么它就不只是在拉扯时空。它还会拖曳时空。 Imagine a spoon turning slowly in a jar of honey. The honey near the spoon gets pulled around with it.想象一把勺子在一罐蜂蜜里慢慢转动。靠近勺子的蜂蜜会被带着一起转。 A spinning mass does that to space itself.一个自转的质量，对空间本身就是这么做的。 This is called frame dragging, and it was predicted in 1918 by two physicists named Lense and Thirring.这叫做参考系拖曳（frame dragging），它在 1918 年由两位名叫 Lense 和 Thirring 的物理学家预言。 It took nearly a century even to detect the whisper of it around the spinning Earth. Around a black hole it should be far stronger.光是要在自转的地球周围探测到它那一丝低语，就花了将近一个世纪。在黑洞周围，它应该要强得多。 And here is the key: frame dragging does not just turn the ellipse within its plane.而关键在这里：参考系拖曳不只是让椭圆在它所在的平面内旋转。 It slowly twists the entire plane of the orbit, tips it, like a coin spun on a table beginning to wobble.它会缓慢地扭转整个轨道平面，使之倾斜，就像一枚在桌面上旋转的硬币开始晃动。 Measure that twist, and you have measured how fast the black hole is spinning. Nobody has ever measured the spin of Sagittarius A-star.测出这个扭转，你就测出了黑洞自转有多快。至今还没有人测量过人马座 A* 的自转。 That is the prize. So now we can be honest about what was and was not found last month. What was found is a star, and its orbit.那才是真正的奖赏。所以现在我们可以诚实地谈谈，上个月究竟发现了什么、又没有发现什么。发现的是一颗恒星，以及它的轨道。

That is solid. They saw it move, they clocked its lap, they fixed its path. What was not found is the spin. Not yet.这一点是扎实的。他们看到它移动，测定了它绕行一圈的时间，确定了它的路径。没有发现的是自转。目前还没有。 The announcement said, in careful words, that this star could let them measure the spin within about the next decade.那份公告用谨慎的措辞说，这颗恒星或许能让他们在大约未来十年内测出自转。 It has to swing back around, closest again in 2031, and they will need to watch it complete something like two full orbits before the tiny twist from frame dragging can be pulled cleanly out of the data.它得再绕回来，在 2031 年再次到达最近点，而他们还需要观测它完成大约两个完整轨道，才能把参考系拖曳带来的那个微小扭转从数据中干净地提取出来。 So when you read that S301 feels the spin of the black hole, hold that lightly. It is a promissory note, not a receipt.所以当你读到 S301 “感受到”黑洞的自转时，请对此保持审慎。这是一张期票，不是一张收据。 The discovery of a fast star is the tool. The measurement of the spin is the job, and the job has not been done.发现一颗快速运动的恒星，是工具。测量自转，才是任务，而这个任务还没有完成。

And there is a second thing the word record hides, which is the part I find most interesting, because it goes right to the hazard I mentioned.而“记录”这个词还掩盖了第二件事，这一点是我觉得最有意思的，因为它正好触及了我前面提到的那个陷阱。 S301 is being called the fastest star and the one with the shortest orbit.S301 被称为速度最快的恒星，也是轨道周期最短的那一颗。 But other teams, a few years ago, announced stars around this same black hole with even shorter orbits and even higher speeds.但几年前，其他团队曾宣布，在同一个黑洞周围发现了轨道更短、速度更高的恒星。 Stars with names like S62 and S4716, one of them claimed to loop around in just four years. Those claims are disputed.那些恒星有着像 S62 和 S4716 这样的名字，其中一颗据称仅用四年就绕行一圈。这些说法存在争议。 The region right around the black hole is desperately crowded with faint stars, and when you look with an ordinary telescope, two faint stars close together can blur into one, and a blur that shifts can be read as a single star on a tight, fast orbit that is not really there.黑洞正周围的那片区域，密密麻麻地挤满了暗弱的恒星，而当你用一台普通望远镜去看时，两颗靠得很近的暗星会模糊成一颗，而一团移动的模糊，可能被解读成一颗处在紧凑、快速轨道上的单颗恒星——可它其实并不存在。 Fit a curve to a few smeared dots, and more than one orbit will fit.对几个被抹开的光点去拟合一条曲线，能拟合上的轨道不止一条。 That is the hazard of reasoning backwards from a trace: the trace does not always pin down a single cause.这就是从踪迹反向推理的风险：踪迹并不总能锁定唯一的成因。

What makes S301 different, and what earns it the word secure, is not that the star is special but that the measurement is better.S301 之所以不同，之所以配得上「确凿」这个词，不是因为这颗恒星特殊，而是因为测量更好。 It was found with an instrument that combines four large telescopes into one, sharp enough to actually resolve the star as a separate point and follow it, rather than fitting a story to a blur.它是用一台把四架大型望远镜合而为一的仪器发现的，其锐度足以真正把这颗恒星分辨为一个独立的点并加以追踪，而不是给一团模糊拟合出一个故事。 So the honest headline is not the fastest star in the galaxy. It is the fastest star whose orbit we have actually nailed down.所以诚实的标题不是「银河系中最快的恒星」，而是「我们真正确定了其轨道的最快恒星」。 The record is a statement about the quality of our seeing, not about the star. That is the thread I want you to keep.这项纪录说的是我们观测质量的高低，而不是这颗恒星本身。这是我希望你记住的那条线索。

Everything at the center of the galaxy that we claim to know, the mass, the darkness, soon perhaps the spin, is read off the choreography of stars around a point of pure absence.关于银河系中心我们声称知道的一切——质量、黑暗，也许不久还有自旋——都是从围绕一个纯粹虚无之点运转的恒星编排中读出来的。 It is the grandest example of a method that runs through all of science: you cannot see the cause, so you watch what it does, and you work backwards, and you stay humble about the fact that more than one cause can leave the same trace.这是一种贯穿全部科学的方法最宏大的范例：你看不见成因，于是你观察它的作用，你反向推演，并且对「不止一个成因可以留下同样的踪迹」这一事实保持谦逊。 S301 is not a headline about a fast star.S301 不是一条关于快速恒星的头条新闻。 It is a new and very sharp needle, dropped closer to the black hole than any before it, and we are waiting, patiently, the way this whole field has always waited, to see how it wobbles.它是一根新的、极为锐利的探针，被投放到比以往任何一根都更靠近黑洞的位置，而我们正在耐心等待——以这整个领域一贯的等待方式——看它将如何摆动。

Here is a claim that sounds wrong and is not: a language model cannot do anything. It cannot read a file. It cannot run a command.有一个说法听起来不对，其实不然：语言模型什么都做不了。它不能读取文件，不能运行命令。

It cannot search the web, send an email, or edit a line of code. 它不能搜索网络，不能发送邮件，也不能编辑一行代码。 A language model takes a sequence of text and produces a probable continuation. That is the entire operation.语言模型接收一段文本序列，然后生成一个可能的续写。这就是它的全部操作。 There is no mechanism inside it that touches the world.它内部没有任何能触及外部世界的机制。

And yet you have certainly seen a demonstration of an AI system reading a codebase, running the tests, noticing a failure and fixing it.然而你肯定见过这样的演示：一个 AI 系统读取代码库、运行测试、发现失败并将其修复。

So something else is doing the doing.所以是别的什么东西在真正执行这些动作。 That something is called the harness, and today is about what it is, why it turns out to matter enormously, and why it is currently the part of the system that most determines whether any of this works.那个东西叫做 harness，而今天要讲的就是它是什么、为什么它的重要性远超想象，以及为什么它目前是整个系统中最能决定这一切是否奏效的部分。

Let me start with what actually happens, step by step, because the mechanics dissolve most of the mystery.让我从实际发生的事情说起，一步一步来，因为一旦弄清机制，大部分神秘感就消散了。

You ask an agent to fix a failing test. The harness assembles a block of text.你让一个 agent 去修复一个失败的测试。harness 组装出一段文本。

It contains a system prompt describing what the agent is and how it should behave, a description of the tools available, your request, and whatever else the harness decided to include.这段文本包含一个系统提示，描述这个 agent 是什么、该如何行事，还有可用工具的说明、你的请求，以及 harness 决定纳入的其他任何内容。 It sends that block to the model. The model produces a continuation.它把这段文本发给模型。模型生成一段续写。

Somewhere in that continuation is a structured fragment saying, in effect, run the command "pytest tests slash test underscore parser".这段续写里的某处有一个结构化的片段，实际上是在说：运行命令“pytest tests/test_parser”。

Now the crucial step. The model has not run anything. It has produced text that describes running something. The harness parses that text.现在是关键的一步。模型并没有运行任何东西。它生成的是一段描述“运行某个东西”的文本。harness 解析这段文本。

It recognises the fragment as a tool call. It checks whether that tool is permitted.它把这个片段识别为一次工具调用。它检查这个工具是否被允许使用。 Then it actually executes the command, on a real machine, and captures the output.然后它真正地在一台真实的机器上执行这条命令，并捕获输出。

Then it takes that output — the test failure, the traceback — appends it to the transcript, and sends the whole thing back to the model as a new block of text.接着它把那份输出——测试失败、traceback——附加到记录上，再把整段内容作为一段新文本发回给模型。

The model reads the traceback and produces another continuation, this time proposing an edit to a file.模型读取这段 traceback，生成另一段续写，这次提出对某个文件的一处修改。 The harness parses, checks, applies the edit, captures the result, appends, sends again.harness 解析、检查、应用这处修改、捕获结果、附加、再次发送。

Around and around, until either the model produces something the harness recognises as "done", or a limit is hit. That loop is the harness.循环往复，直到模型生成出某种被 harness 识别为“完成”的东西，或者触及某个上限。这个循环就是 harness。

Everything else is elaboration on it.其余一切都是围绕它的展开。

Now the piece that I think is genuinely counterintuitive, and it is the one that reorganises how you think about the whole subject.现在讲我认为真正违反直觉的那一部分，也正是这一部分会重组你对整个主题的理解方式。

The model is stateless. It remembers nothing between calls. Nothing.模型是无状态的。它在两次调用之间什么都不记得。什么都不记得。

When the harness sends the second request, the model is not continuing anything, because there is no thread inside it to continue.当 harness 发出第二个请求时，模型并不是在延续任何东西，因为它内部根本没有一条可供延续的线索。 It receives a block of text that happens to describe a conversation that has already been partly conducted, and it produces the next plausible piece.它接收到的是一段恰好描述了一场已经进行了一部分的对话的文本，然后它生成下一个说得通的片段。

Every single turn, the entire history is re-sent and re-read from the beginning.每一个回合，整段历史都会被从头重新发送、重新读取。

So what feels like a persistent agent working steadily through a problem is, mechanically, a sequence of independent readings of a growing transcript, each performed by something with no memory of having done the previous one.所以，那种感觉上像是一个持续存在的 agent 稳步地推进一个问题的东西，从机制上看，其实是对一份不断增长的记录进行的一连串彼此独立的阅读，每一次阅读都由一个对自己做过上一次毫无记忆的东西来完成。

The continuity is not in the model. The continuity is a file that the harness keeps and re-reads aloud.连续性不在模型里。连续性是一份由 harness 保存、并反复朗读出来的文件。

Once you have that, several things that seem puzzling become obvious.一旦理解了这一点，几件看似费解的事情就变得显而易见了。 Why does an agent sometimes forget a constraint you gave it forty steps ago?为什么一个 agent 有时会忘掉你在四十步之前给它定下的一条约束？ Because the constraint is still in the transcript but is now buried under forty steps of tool output, and attention is finite.因为那条约束仍然在记录里，但如今被埋在了四十步的工具输出之下，而注意力是有限的。 Why does behaviour degrade as a session gets long? Because the signal-to-noise ratio of the transcript is falling.为什么会话变长后表现会下降？因为对话记录的信噪比在下降。 Why do these systems have a context limit at all, and why does everything get harder near it?为什么这些系统会有上下文上限，而且越接近上限一切就越困难？ Because the transcript is the entire mind, and it is running out of room. That reframes the central engineering problem.因为这份对话记录就是它的全部心智，而它的空间正在耗尽。这重新界定了核心的工程问题。

It is not "how do we make the model smarter". It is "what should be in the window".问题不是「我们如何让模型更聪明」，而是「窗口里该放什么」。

And that question turns out to be most of the difficulty.而这个问题恰恰构成了大部分难度所在。 Retrieval is a strategy for it: fetch the relevant document rather than keeping everything.检索是应对它的一种策略：取回相关的文档，而不是把所有内容都留着。 Compaction is a strategy: when the transcript grows too long, summarise the early part and continue with the summary.压缩是一种策略：当对话记录变得太长时，把前面的部分总结出来，然后带着这份摘要继续。 Subagents are a strategy: hand a self-contained piece of work to a separate agent with its own fresh window, and take back only the conclusion rather than all the intermediate mess.子智能体是一种策略：把一块自成一体的工作交给一个独立的智能体，让它用自己全新的窗口去做，然后只取回结论，而不是全部中间的杂乱过程。 Memory files are a strategy: write the durable facts to disk so that a future session can read them back.记忆文件是一种策略：把持久的事实写入磁盘，好让未来的会话能把它们读回来。

Every one of those is an answer to the same question, which is what to put in a finite window, given that the window is the entirety of what the thing knows.这些每一个都是对同一个问题的回答，即在一个有限的窗口里该放什么——考虑到这个窗口就是这个东西所知道的一切。

Now the second structural point, and it matters for safety. The capability to act lives entirely in the harness. The model emits a request.现在讲第二个结构性要点，它对安全性很重要。行动的能力完全存在于外壳（harness）之中。模型发出一个请求。

The harness decides whether to honour it.由外壳来决定是否满足它。

If the harness has no tool for sending email, the agent cannot send email — not because it lacks the idea, but because the text describing the idea will not be recognised as anything actionable, and will simply sit there as text.如果外壳没有发送邮件的工具，那么这个智能体就无法发送邮件——不是因为它没有这个想法，而是因为描述这个想法的文本不会被识别为任何可执行的东西，只会作为文本原地待着。

This is why permission systems in these products are built where they are.这就是为什么这些产品里的权限系统会建在它们所在的位置。 There is a natural chokepoint: every action passes through one place, and that place is code you control rather than weights you do not.这里有一个天然的咽喉点：每一个行动都要经过同一个地方，而那个地方是你能控制的代码，而不是你无法控制的权重。 Sandboxing, allowed-tool lists, approval prompts before destructive operations, read-only modes — all of them live at that chokepoint.沙箱、允许工具清单、在破坏性操作前的批准提示、只读模式——所有这些都存在于那个咽喉点上。

It also means "what can this agent do" is a question with a precise answer, and the answer is a list of tools in a configuration file, not a matter of speculation about the model.这也意味着「这个智能体能做什么」是一个有精确答案的问题，而答案是配置文件里的一份工具清单，而不是关于模型的一种揣测。

Now, the history, because this developed fast and the failures taught more than the successes.现在讲历史，因为这个领域发展得很快，而且失败比成功教会我们的更多。

The founding paper is generally taken to be ReAct, by Shunyu Yao and colleagues, published in October twenty twenty-two.奠基性的论文一般被认为是 ReAct，由姚顺雨及其同事撰写，发表于 2022 年 10 月。 The idea in the title: synergising reasoning and acting.标题里的想法：让推理与行动协同增效。 Rather than have the model plan everything up front and then execute, or act with no reasoning, interleave them.与其让模型一开始就把一切都规划好然后执行，或者不加推理地行动，不如把两者交织起来。 Think, act, observe the result, think again. That sounds almost too simple to be a contribution.思考、行动、观察结果、再思考。这听起来简单得几乎算不上是一项贡献。

It was a real one, and the reason is the observation step.它确实是一项真正的贡献，原因就在于那个观察步骤。 A model that acts and then reads the actual result of its action can notice that the result was not what it expected.一个会行动、然后读取自己行动实际结果的模型，能够注意到结果并非它所预期的那样。 A model that plans the whole sequence in advance cannot, because it never sees anything come back.一个预先把整个序列都规划好的模型做不到，因为它从来看不到任何反馈回来。

Then March twenty twenty-three, and AutoGPT.然后是 2023 年 3 月，以及 AutoGPT。

Somebody wrapped GPT-4 in a loop that let it set its own subgoals and pursue them without supervision, and it went viral on a scale that is hard to convey.有人把 GPT-4 包进一个循环里，让它能自己设定子目标并在无人监督的情况下去追求它们，然后它以一种难以形容的规模走红了。 The framing was full autonomy: give it an objective, walk away, come back to a finished product.它的定位是完全自主：给它一个目标，走开，回来时就有一个完成的成品。

It mostly did not work, and how it failed is the useful part. It got stuck in loops.它大多数时候并不奏效，而它失败的方式才是有用的部分。它会卡在循环里。

It would repeat the same action indefinitely, because nothing checked whether progress was being made and the completion criteria were often impossible to satisfy.它会无限地重复同一个动作，因为没有任何东西去检查是否有进展在发生，而且完成的判定标准往往无法满足。 It assumed capabilities it did not have. It never asked a clarifying question.它假定自己拥有并不具备的能力。它从不提出澄清性的问题。 And it burned real money doing it, because every one of those useless steps was a paid API call.而且它这么做时烧的是真金白银，因为那些无用步骤里的每一步都是一次付费的 API 调用。 One study by Amazon researchers put its success rate on a shopping task at twenty-four percent.亚马逊研究人员的一项研究显示，它在一项购物任务上的成功率是 24%。

Here is what I think the episode-worthy lesson is. AutoGPT was not short of intelligence. It had access to the best model available.以下是我认为值得为此专门做一期节目的教训。AutoGPT 并不缺智能。它用的是当时能拿到的最好的模型。 What it lacked was everything around the model: no way to detect it was not progressing, no bound on cost, no verification that a step had done what it claimed, and no mechanism to stop.它缺的是模型之外的一切：没有办法察觉自己没在推进，没有对成本设上限，没有验证某一步是否真的完成了它声称完成的事，也没有停下来的机制。

Which is to say the model was fine and the harness was missing.换句话说，模型是好的，缺的是 harness。 The project itself eventually became something quite different — a workflow builder where a human draws the boundaries and the agent operates inside them.这个项目本身最终变成了相当不同的东西——一个工作流构建器，由人来划定边界，agent 在边界之内运作。

And the field moved in the same direction, toward domains where the harness can check the work.整个领域也朝同一个方向发展，转向那些 harness 能够核查工作成果的领域。

That is why coding became the flagship application, and the reason is not that code is easy. It is that code comes with an oracle.这就是为什么编程成了旗舰应用，原因不在于代码简单，而在于代码自带一个 oracle（判定器）。 You can run the tests. You can compile it.你可以运行测试。你可以编译它。 The harness can execute a command and get back a fact about whether the change worked, and feed that fact to the model.harness 可以执行一条命令，得到关于这次改动是否奏效的事实，再把这个事实喂给模型。 Compare that with asking an agent to write a strategy document — nothing in the harness can tell whether the output is any good, so the loop has no signal to close on.拿这个跟让 agent 写一份战略文档对比一下——harness 里没有任何东西能判断输出好不好，所以这个循环没有可以据以收敛的信号。

The general principle: agents work best where verification is cheap and automatic.总的原则是：agent 在验证成本低廉且能自动完成的地方表现最好。 Where it is not, you are relying on the model being right the first time, which is a much weaker position.在做不到这一点的地方，你就得指望模型第一次就做对，那是一个弱得多的处境。

Now to the empirical claim that I think justifies the whole episode.现在来说说我认为足以支撑起整期节目的那个实证论断。

If the harness were a thin wrapper, swapping it should barely change anything. Benchmark scores would be a property of the model.如果 harness 只是一层薄薄的封装，那么替换它应该几乎不改变什么。基准测试分数将会是模型自身的属性。

They are not. On SWE-bench Pro, under a standardised research scaffold, Claude Opus 4.5 scores about forty-six percent.但事实并非如此。在 SWE-bench Pro 上，采用一套标准化的研究脚手架，Claude Opus 4.5 的得分约为 46%。

The same model, under Claude Code, reaches about fifty-five percent.同一个模型，在 Claude Code 之下，达到了约 55%。 Same weights, same task set, nine and a half points of difference, produced entirely by the software around the model.同样的权重，同样的任务集，9.5 个百分点的差距，完全由模型周围的软件所造成。

The Holistic Agent Leaderboard, which runs the same models under different scaffolds deliberately, reports double-digit gaps routinely and single-model swings approaching forty-eight points.Holistic Agent Leaderboard 有意让同样的模型在不同的脚手架下运行，其结果经常报告出两位数的差距，单个模型的波动接近 48 个百分点。 On one leaderboard run, a single model under five different harnesses spread from thirty-nine percent to sixty-six.在一次榜单运行中，同一个模型在五种不同的 harness 下，分数从 39% 一直分布到 66%。

Sit with the size of those numbers, because papers announce model advances of two to four points as significant.好好体会这些数字的量级，因为论文把 2 到 4 个百分点的模型进步就宣布为显著。 Harness substitution routinely dwarfs them. Two consequences follow, and they point in different directions.替换 harness 带来的差距经常把它们远远盖过。由此引出两个后果，而它们指向不同的方向。

The first is practical and encouraging: if you are trying to get useful work out of these systems, the harness is where your effort has leverage.第一个后果是务实且令人鼓舞的：如果你想从这些系统里获得有用的成果，harness 才是你的努力有杠杆效应的地方。

Tool design, context strategy, and verification will move your results more than waiting for the next model.工具设计、上下文策略和验证对结果的推动，会比等待下一代模型更大。

The second is uncomfortable: a leaderboard number is a joint property of a model and a harness, and comparing two models evaluated under different scaffolds is close to meaningless.第二个后果则令人不安：一个榜单分数是模型和 harness 共同的属性，而比较两个在不同脚手架下评测的模型，几乎毫无意义。 There is now a small literature making exactly this complaint, arguing that harness details must be disclosed for agent evaluations to mean anything.如今已经有一小批文献提出的正是这个抱怨，主张 agent 评测中必须披露 harness 的细节，评测结果才有意义。 They frequently are not.而它们常常并没有被披露。

Now let me get to what I think is the hardest problem in the area, which is long-horizon reliability, and it is an arithmetic problem before it is anything else.现在让我讲讲我认为这个领域里最难的问题，也就是长时程可靠性，而它首先是一个算术问题，其次才是别的什么。

Suppose an agent is ninety-five percent reliable per step. That sounds good.假设一个 agent 每一步有 95% 的可靠性。这听起来不错。 Run a hundred steps with no recovery mechanism, and the probability of getting through cleanly is nought point nine five to the hundredth power, which is about half a percent.在没有任何恢复机制的情况下运行 100 步，干净通过的概率是 0.95 的 100 次方，大约是 0.5%。

Ninety-nine percent per step gives you thirty-seven percent over a hundred steps.每步 99% 的成功率，在 100 步后只剩 37%。 To get a ninety percent chance of a clean hundred-step run you need about ninety-nine point nine percent per step.要让一次 100 步的运行有 90% 的概率干净完成，你每步需要大约 99.9% 的成功率。

That is the tyranny of compounding, and it tells you something important: for long tasks, per-step accuracy is almost never the answer.这就是复利的暴政，它告诉你一件重要的事：对于长任务，单步准确率几乎从来都不是答案。 You cannot make the steps reliable enough.你没办法把每一步做得足够可靠。 What you need is for errors to be detected and recovered from, so that a failure costs you a step rather than the run.你需要的是错误能够被检测并从中恢复，这样一次失败只会让你损失一步，而不是整次运行。

Which reframes what a good harness is doing. It is not primarily making the model perform better. It is catching the model when it is wrong.这重新定义了一个好的 harness 究竟在做什么。它主要不是在让模型表现得更好，而是在模型出错时接住它。 Run the tests after the edit. Check that the file was actually written.编辑之后运行测试。确认文件确实被写入了。 Notice that the same command has now failed three times and stop trying it. Verify before proceeding.注意到同一条命令现在已经失败了三次，就别再试了。在继续之前先验证。

There is a measurement of how far this has got, from METR, and it is the most useful single number I know for tracking agent progress.对于这件事进展到什么程度，有一个来自 METR 的量度，它是我所知道的追踪 agent 进展最有用的单一数字。 Rather than scoring tasks, they ask: how long a task, measured by how long it takes a human professional, can an agent complete with fifty percent reliability?他们不是给任务打分，而是问：以人类专业人士完成所需的时长来衡量，一个 agent 能以 50% 的可靠度完成多长的任务？

That horizon has been doubling roughly every seven months for six years.六年来，这个时间跨度大约每七个月翻一番。 Frontier models around the time of the study sat near a couple of hours.研究进行前后的前沿模型大约处在两个小时的水平。

I like that metric because it measures the thing that actually matters — sustained competence over a long chain — rather than performance on a single question. 我喜欢这个指标，因为它衡量的是真正重要的东西——在一条长链上持续保持的胜任力——而不是在单个问题上的表现。 And note that the fifty percent reliability figure is a coin flip.还要注意，50% 可靠度这个数字就是抛硬币。 A two-hour horizon at fifty percent is not two hours of work you can trust; it is two hours of work you must check.50% 可靠度下的两小时跨度，并不是你可以信任的两小时工作量；而是你必须去核查的两小时工作量。

Now, the honest limits, because this area is drowning in demonstrations.现在说说那些诚实的局限，因为这个领域充斥着各种演示。

There is an enormous gap between a demo and a system you can leave running. Demos are selected.一次演示和一套你能放着让它自己跑的系统之间，存在着巨大的鸿沟。演示是精选出来的。 They are the successful trace out of many attempts, and the failures are not shown, and the failures are where the engineering is.它们是众多尝试中那条成功的轨迹，失败的部分不会展示出来，而工程的难点恰恰在那些失败里。

Benchmarks are increasingly gamed, sometimes inadvertently.基准测试越来越多地被钻空子，有时是无意的。 Harnesses get tuned against the benchmark until they perform well on it specifically, which measures the tuning rather than the capability.harness 会针对基准不断调优，直到专门在它上面表现良好，而这衡量的是调优本身，而不是能力。 Contamination — benchmark tasks appearing in training data — is a persistent worry that is hard to rule out.污染——基准任务出现在训练数据里——是一个挥之不去、又难以排除的隐忧。

And the deepest gap is that we do not have good ways to evaluate long, open-ended, judgement-laden work, precisely because it is the work where verification is not cheap.而最深的鸿沟在于，我们没有好的办法去评估那些长的、开放式的、充满判断的工作，恰恰是因为在这类工作里验证并不廉价。 Everything we measure well is everything the harness can check, which is a biased sample of everything we care about.我们能测得好的一切，都是 harness 能核查的一切，而这是我们真正在意的一切当中一个有偏的样本。

Let me finish with an example I can vouch for, because it is this podcast. This episode was written by an agent running in a harness.让我用一个我能亲自担保的例子来收尾，因为它就是这档播客。这一集是由一个在 harness 中运行的 agent 写就的。

Not narrated by one — the voice is a local text-to-speech model — but researched, drafted and filed by an agent working in a loop of exactly the kind described.不是由它来配音——声音出自一个本地的文本转语音模型——而是由一个正是以上文所述那种循环方式工作的 agent 完成研究、起草和归档的。

And the interesting part is not the writing.而有意思的部分并不是写作本身。 It is the machinery around it, which exists entirely because of the failure modes in this episode.而是围绕它的那套机制，这套机制之所以存在，完全是因为本集所讲的那些失败模式。

Every citation in the show notes is checked by a script that fetches the URL and confirms it resolves, because a language model asked for references will sometimes produce plausible ones that do not exist.show notes 里的每一条引用都由一个脚本核查，它会抓取 URL 并确认它能够解析，因为一个语言模型在被要求提供参考文献时，有时会生成看似可信、实则并不存在的条目。 That is a verification gate: the harness checking the model's work with something that is not a model.这就是一道验证关卡：harness 用某种并非模型的东西来核查模型的工作。

The pipeline runs in two phases, preparing an episode the night before and publishing the next morning, which exists so that there is a window in which a human can look at it.整个流水线分两个阶段运行，前一天晚上准备一期节目，第二天早上发布，之所以这样设计，是为了留出一个可以让人来看一眼的窗口期。

There is a gate that executes the page's own JavaScript and confirms every promised article actually renders, added after a syntax error silently blanked a section.有一道关卡会执行页面自身的 JavaScript，确认每一篇承诺要有的文章都真的渲染出来了，这是在一次语法错误悄无声息地让某个板块变成空白之后加上的。

And there is a failure alert, added after the most instructive incident in the project's history.还有一个失败告警，是在这个项目历史上最有教育意义的一次事故之后加上的。 A gate correctly refused to publish a broken issue. It wrote the reason to a log. Nobody read the log.一道关卡正确地拒绝发布一期有问题的节目。它把原因写进了日志。没人去读那份日志。 Three consecutive weeks of output silently did not happen, and nobody noticed until it was asked about directly.连续三周的产出就这样悄无声息地没有发生，直到有人直接问起来，才有人注意到。

That is the harness lesson stated as plainly as it can be. The gate worked.这就是那条关于 harness 的教训，用最直白的方式讲出来。关卡起作用了。 The system still failed, for three weeks, because a failure that is not surfaced is indistinguishable from a system that has quietly stopped existing.系统却依然失败了，失败了三周，因为一个没有被呈现出来的失败，和一个悄悄停止存在的系统，是无从区分的。

So: the model is the part that reasons.所以说：模型是负责推理的那部分。 The harness is the part that acts, remembers, verifies, recovers, and — most easily forgotten — tells someone when it has stopped.harness 是负责行动、记忆、验证、恢复的那部分，还有——最容易被遗忘的——在它停下来的时候告诉某个人。

Almost all of the reliability lives in the second one.几乎所有的可靠性都住在第二部分里。 And at the moment, so does almost all of the difference between a system that impresses you for ten minutes and one you can actually leave running.而此刻，一个能让你惊艳十分钟的系统，和一个你真的可以放着让它一直跑的系统，两者之间几乎全部的差距，也住在这里。

This August, a woman died in Los Alamos, New Mexico, at the age of ninety-seven. Her name was Mary Tsingou.今年八月，一位女性在新墨西哥州的洛斯阿拉莫斯去世，享年九十七岁。她的名字叫玛丽·钦古（Mary Tsingou）。 If you have heard of the thing she is now finally named for, you have almost certainly heard it called something else.如果你听说过如今终于以她命名的那个东西，你几乎肯定听到的是另一个名字。 For fifty-three years it was the Fermi-Pasta-Ulam problem, after three men. She was in a footnote.五十三年来，它一直被称为费米-帕斯塔-乌拉姆问题，取自三位男性的名字。而她只出现在一个脚注里。 Today I want to tell you what she actually did, because it is one of the strangest and most important results in twentieth-century science, and because the story of how her name came back is a small echo of the discovery itself.今天我想告诉你她究竟做了什么，因为这是二十世纪科学中最奇特、也最重要的成果之一，也因为她的名字如何重新归位的故事，本身就是这一发现的一个小小回响。

Start in the summer of 1953, at Los Alamos, in front of a machine called the MANIAC.故事要从1953年夏天说起，地点是洛斯阿拉莫斯，一台名叫 MANIAC 的机器前。 It was one of the first electronic computers, and it was slow and temperamental and had less memory than a modern doorbell.它是最早的电子计算机之一，运行缓慢、脾气古怪，内存比现代的门铃还少。 Enrico Fermi, one of the great physicists of the century, had an idea for it that was, at the time, almost heretical.恩里科·费米，这个世纪最伟大的物理学家之一，对它有一个在当时几乎堪称异端的想法。 He did not want to use the machine to compute a number he already knew how to get.他不想用这台机器去计算一个他已经知道怎么算出来的数。 He wanted to use it to watch something he could not solve by hand. He wanted to do an experiment inside the computer.他想用它来观察一个自己无法手算求解的东西。他想在计算机内部做一场实验。

Here is the experiment.实验是这样的。 Imagine a string, but think of it as a row of sixty-four little weights, joined in a line by springs, with the two ends pinned down.想象一根弦，但把它设想成一排六十四个小重物，用弹簧连成一条线，两端固定住。 Pluck it in the smoothest possible way. Not a jagged pluck.以尽可能平滑的方式拨动它。不是那种参差不齐的拨法。 One single, gentle hump, the whole string bowed like the lowest note on a guitar.而是单独一个、轻柔的隆起，整根弦弯成一道弓形，就像吉他上最低的那个音。 If the springs were perfectly ordinary, that hump would just swing back and forth forever, and nothing interesting would ever happen.如果弹簧完全普通，那道隆起就只会永远来回摆动，什么有趣的事都不会发生。 The energy would stay in that one shape. So Fermi and his colleagues changed the springs, just a little.能量会一直停留在那一种形状里。于是费米和他的同事们改动了弹簧，只改了一点点。

They made them very slightly nonlinear. That is a technical word, so let me give you the picture.他们让弹簧带上极其轻微的非线性。这是个技术词汇，我来给你描绘一下画面。 An ordinary spring pushes back in exact proportion to how far you stretch it. Stretch it twice as far, it pushes twice as hard.普通弹簧的回弹力与你把它拉伸的距离严格成正比。拉伸两倍远，它就回推两倍的力。 A nonlinear spring is one where that proportion is a little off. Stretch it twice as far, it pushes back a bit more than twice as hard.非线性弹簧则是这个比例稍稍有点偏差。拉伸两倍远，它回推的力会比两倍稍多一点。 A tiny imperfection, deliberately added. And everyone knew what a tiny imperfection like that should do. It should scramble the string.一个微小的、刻意添加的瑕疵。而所有人都知道这样一个微小瑕疵应该会造成什么后果。它应该会把弦搅乱。

This was not a wild guess. It was the settled expectation of physics, and it has a name: thermalization, or equipartition of energy.这不是什么大胆的猜测。这是物理学中已成定论的预期，而且它有个名字：热化，或者说能量均分。 The reasoning goes like this. Your single smooth hump is one pure mode of vibration.推理是这样的。你那单独一个平滑的隆起，是一种纯粹的振动模式。 But the string can also vibrate in finer, more wrinkled shapes, faster ripples with more wiggles in them.但弦还可以以更细密、更褶皱的形状振动，那是带有更多摆动的、更快的涟漪。 In a perfectly linear string, those shapes never talk to each other. The nonlinearity is what lets them talk.在一根完全线性的弦里，这些形状之间从不交流。而非线性正是让它们能够交流的东西。 So the energy you poured into the one big hump should begin to leak.所以你灌注进那一个大隆起里的能量，应该会开始泄漏。 First into a slightly more wrinkled shape, then a finer one, then finer still, spreading out among all the possible modes until the string is a jittering, random mess with the energy shared roughly equally everywhere.先是漏进一个稍微更褶皱的形状，然后更细的，然后更细，在所有可能的模式之间扩散开来，直到整根弦变成一团抖动的、随机的混乱，能量大致均匀地分摊到各处。 Heat, in other words.换句话说，就是热。 This is the same intuition that tells you a drop of ink in a glass of water will spread until the water is evenly grey, and will never, ever gather itself back into a drop.这和那个告诉你一滴墨水滴进一杯水里会扩散、直到水变成均匀灰色、且永远永远不会自己重新聚回一滴的直觉，是同一个直觉。 Fermi expected to watch the ordered hump dissolve into disorder. He thought the interesting question was only how fast.费米预期看到那个有序的隆起消融为无序。他以为唯一有趣的问题只是：有多快。

Someone had to actually build this experiment and run it, and that someone was Mary Tsingou.总得有人真正把这个实验搭建起来并运行它，而这个人就是玛丽·钦古。 She was a mathematician, one of the first people ever to program the MANIAC.她是一位数学家，是有史以来最早为 MANIAC 编写程序的人之一。 Fermi and Stanislaw Ulam and John Pasta framed the physics question. But framing the question is not the same as answering it.Fermi、Stanislaw Ulam 和 John Pasta 提出了这个物理问题。但提出问题和解答问题不是一回事。 Tsingou devised the algorithm, worked out how to represent this vibrating string as instructions a primitive machine could execute, drew the flowcharts, wrote the code, debugged it, and ran it.Tsingou 设计了算法，琢磨出如何把这根振动的弦表示成一台原始机器能执行的指令，画出流程图，写下代码，调试它，然后运行它。 And I want you to sit with that word for a second: ran it.我想请你在这个词上停留一秒：运行它。 In an ordinary experiment, with glassware and wires, the person who builds the apparatus and takes the readings is an experimentalist, a full author.在一场普通的实验里，用玻璃器皿和电线，搭建仪器并读取数据的人是实验者，是完整的作者。 This was a numerical experiment, Ulam's own phrase, and the apparatus was the code.这是一场数值实验——这是 Ulam 本人的说法——而仪器就是那段代码。 The person who built and ran the apparatus was Mary Tsingou. Leaving her off is not a footnote-sized omission.搭建并运行这台仪器的人是 Mary Tsingou。把她漏掉，不是一个脚注大小的疏忽。 It is like publishing the results of a telescope and thanking, in small print, the person who built the telescope and pointed it.这就好比发表了一架望远镜的观测结果，却只在小字里感谢那个建造望远镜并把它对准目标的人。

Now the result, which is the reason any of us are talking about it. The energy did start to leak, just as predicted.现在说结果，这才是我们所有人今天讨论它的原因。能量确实开始泄漏，正如预测的那样。 It flowed out of the big hump into the more wrinkled modes. For a while the string looked like it was on its way to becoming a random mess.它从那个大的隆起流入了更加褶皱的模态。有那么一阵子，这根弦看起来正走在变成一团随机混乱的路上。 And then it stopped. And then it turned around.然后它停了下来。然后它掉头往回走。 The energy flowed back out of the finer modes and reassembled, almost perfectly, into the original single hump.能量从那些更精细的模态里回流，几乎完美地重新聚拢成最初那个单一的隆起。 After about a hundred and fifty-seven swings of that lowest note, the string was very nearly back where it started, with something like ninety-seven percent of the energy home in the mode it began in.在那个最低音符大约摆动了 157 次之后，这根弦几乎回到了它出发的地方，大约 97% 的能量回到了它最初所在的模态里。 The ink drop had un-spread. This is called Fermi-Pasta-Ulam-Tsingou recurrence, and nobody had a theory for it.那滴墨水又收拢回去了。这被称为 Fermi-Pasta-Ulam-Tsingou 回归，没有人有理论能解释它。

Let me be careful here, because this is where the popular version goes wrong, and where this listener, I think, will want the honest ledger.我在这里要谨慎一点，因为这正是通俗版本出错的地方，也是我想，这位听众会想要一份诚实账目的地方。

The first thing you will read is that this was the birth of chaos theory. That is almost backwards. What they saw was not chaos.你首先会读到的说法是，这是混沌理论的诞生。这几乎是反过来了。他们看到的并不是混沌。 Chaos is the disorder they expected and did not get.混沌是他们预期会出现、却没有得到的那种无序。 What they saw was the opposite: stubborn, delicate order where the textbooks demanded randomness.他们看到的恰恰相反：在教科书要求出现随机性的地方，却出现了顽固而精妙的秩序。 It is more accurate to say this one result seeded two fields at once.更准确的说法是，这一个结果同时孕育了两个领域。 It led toward chaos theory indirectly, by posing a puzzle deep enough that solving it required new mathematics about when order survives.它间接地导向了混沌理论，是因为它抛出了一个足够深刻的谜题，深到解开它需要一套关于秩序何时得以存续的新数学。 But directly, it led to soliton theory.但它直接导向的，是孤子理论。 In 1965, Norman Zabusky and Martin Kruskal took the continuum limit of exactly this system and found solitary waves, humps that hold their shape, pass right through each other, and come out unchanged.1965 年，Norman Zabusky 和 Martin Kruskal 对正是这个系统取了连续极限，发现了孤立波——那些能保持自身形状、能彼此穿透、穿过后毫发无损的隆起。 They coined the word soliton for them. The recurrence is those waves, periodically lining back up.他们为这些波造了 soliton（孤子）这个词。那个回归现象，就是这些波周期性地重新排列成行。 So if you want a one-sentence correction: it was not the birth of chaos.所以如果你想要一句话的更正：这不是混沌的诞生。 It was the birth of the numerical experiment, and it happened to give birth to solitons on the way.这是数值实验的诞生，而它在途中恰好孕育了孤子。

The second correction is more subtle, and it is my favorite part.第二处更正更微妙，也是我最喜欢的部分。 It is tempting to say the experiment proved that energy does not thermalize, that equipartition simply fails. That is also wrong.人们很容易说，这个实验证明了能量不会热化，能量均分原理干脆就失效了。这同样是错的。 We now know the string does eventually scramble. It does reach the random, evenly-shared state everyone expected.我们现在知道，这根弦最终还是会被打乱。它确实会达到所有人预期的那种随机、均匀共享的状态。 It just takes vastly, almost unimaginably longer than those first runs could see. The recurrence is not the end of the story;只不过它花的时间要漫长得多，几乎长到无法想象，长到那最初几次运行根本看不到。回归并不是故事的结尾； it is a long, structured detour on the way to equilibrium, a kind of metastable pause.它只是通往平衡途中一段漫长而有结构的绕行，是一种亚稳态的停顿。 Run the machine long enough, and higher, rarer interactions between the modes finally break the spell and the energy spreads for good.让这台机器运行得足够久，模态之间那些更高阶、更罕见的相互作用最终会打破这个魔咒，能量便一去不复返地散开了。 So the true discovery was not that thermalization fails.所以真正的发现并不是热化失败了。 It was that the road to equilibrium can be enormously long and beautifully organized, full of near-perfect returns, so that a short experiment sees only the returns and misses the destination.而是通往平衡的道路可以极其漫长、又极其井然有序，充满了近乎完美的回归，以至于一次短暂的实验只看到了这些回归，却错过了终点。 Fermi did not find a system that refuses to reach equilibrium. He found that we badly misjudged how it gets there.费米并没有找到一个拒绝达到平衡的系统。他发现的是，我们严重误判了它抵达平衡的方式。

And the third caution: you will occasionally see a headline saying the problem has been solved. Treat that gently.第三点提醒是：你偶尔会看到某个标题宣称这个问题已经被解决了。对此要温和看待。 There are by now several explanations that each work in their own regime, the soliton picture and the theory of persistent orderly motion known as KAM among them, and researchers are still, seventy years later, publishing on the exact route from one hump to true randomness.到目前为止已经有好几种解释，每一种都在各自的范畴内成立，其中包括孤子（soliton）图景，以及关于持续有序运动的理论——即所谓的 KAM 理论——而研究者们在七十年之后，仍在发表关于从一个驼峰到真正随机性之间确切路径的论文。 It is understood in pieces. It is not closed. So what did Fermi really do in 1953? He did not prove a theorem.它是被一块一块地理解的。它并未被盖棺定论。那么费米在 1953 年究竟做了什么？他没有证明一条定理。

He proposed that a computer could be an instrument for discovering things, not just for grinding through arithmetic you already understood.他提出，计算机可以成为一种用来发现事物的仪器，而不仅仅是用来机械地演算你早已理解的算术。 That is the founding act of computational science, and nearly everything from climate models to the simulations that design aircraft descends from it.这是计算科学的奠基之举，从气候模型到设计飞机的仿真，几乎一切都源自于此。 And the person who turned that proposal into a working instrument, and was the first human being to watch the energy come home, spent fifty-three years as an initial that got dropped.而那个把这个提议变成一台可用仪器、并且第一个亲眼看到能量回归的人，却有五十三年一直被简化成一个被丢掉的首字母。 In 2008 a physicist named Thierry Dauxois wrote a short piece titled, roughly, Fermi, Pasta, Ulam, and a mysterious lady, and argued the problem should carry her name.2008 年，一位名叫 Thierry Dauxois 的物理学家写了一篇短文，标题大致是《费米、帕斯塔、乌拉姆，和一位神秘的女士》，并主张这个问题应当以她的名字命名。 It slowly did. She lived to see it. Her name, like the energy in her string, took a very long detour, and then came back.它慢慢地做到了。她活着看到了这一天。她的名字，就像她那根弦里的能量一样，绕了一个极长的弯路，然后又回来了。

This week the headlines said the human heart can heal itself.这周的新闻标题说，人的心脏能够自我修复。 A team in Sydney reported that after a heart attack, the surviving muscle starts making new muscle cells.悉尼的一个团队报告说，心脏病发作之后，存活下来的肌肉开始生成新的肌肉细胞。 The word regenerate showed up everywhere, and underneath it, the old hope that one day a damaged heart could be coaxed back to full strength."再生"这个词到处都是，而在它背后，是那个古老的希望——有朝一日，受损的心脏能被重新引导，恢复到完全的力量。

It is a good story, and most of it is true. But the interesting part is not the headline.这是个好故事，其中大部分是真的。但有意思的部分不在标题里。 The interesting part is that for most of the last century the textbook said the opposite, that the heart you are born with is the heart you die with, no new muscle ever, and that the textbook was also, very nearly, right.有意思的部分在于，上个世纪的大部分时间里，教科书讲的恰恰相反：你出生时的那颗心脏，就是你死时的那颗心脏，永远不会有新的肌肉；而这本教科书，也几乎是对的。 Both the cheerful new story and the grim old one are wrong in the same small way, and the thing that settled it is one of the most elegant measurements in biology.那个欢快的新故事和那个冷峻的旧故事，在同一个小地方都错了，而最终把这件事定下来的，是生物学中最优雅的测量之一。 So today, three things. How anyone could possibly know whether a living human heart makes new cells.所以今天讲三件事。人究竟怎么可能知道一颗活着的人类心脏是否在生成新细胞。 What the honest number turned out to be. And the expensive fraud that grew in the gap between that number and the one everyone wanted.那个诚实的数字最后究竟是多少。以及在那个数字和所有人想要的数字之间的缝隙里，滋长出的一场代价高昂的造假。

Start with why the old dogma was believable. A heart muscle cell is built to do one thing, forever.先说为什么旧的教条曾经可信。一个心肌细胞天生就是为了永远做一件事而造的。 It fills itself almost wall to wall with the protein machinery of contraction, the tiny ratchets that shorten when they are told to, and it wires itself to its neighbors so the whole sheet squeezes in time.它几乎从一壁到另一壁把自己填满了收缩用的蛋白质机械——那些接到指令就缩短的微小棘轮——并且把自己和邻居连成电路，好让整片肌肉齐步挤压。 To divide, a cell has to take all of that apart. It has to dissolve its internal scaffolding, copy its chromosomes, and pull itself in two.要分裂，一个细胞必须把这一切拆开。它必须溶解自身内部的支架，复制自己的染色体，再把自己拉成两半。 A heart cell cannot do that, because the heart cannot stop. There is no maintenance window.心脏细胞做不到这一点，因为心脏不能停。没有停机检修的窗口。 So it seemed obvious that heart cells, once built in the womb, simply never divide again.所以看起来很明显：心脏细胞一旦在子宫里造好，就再也不会分裂了。 When people have a heart attack, a branch of the coronary plumbing clogs, the muscle downstream starves, and a chunk of it dies and turns to scar.人心脏病发作时，冠状动脉管路的一条分支堵塞，下游的肌肉挨饿，其中一块坏死，变成疤痕。 The scar never becomes muscle again. The permanence seemed to follow. Here is the problem. You cannot watch.疤痕永远不会再变回肌肉。这种永久性看起来是顺理成章的。问题就在这里。你没法看。

You cannot sit and stare at a living human heart and count which cells are new.你没法坐在那儿盯着一颗活着的人类心脏，去数哪些细胞是新的。 By the time you have the tissue, in an autopsy, the person is dead and nothing is dividing.等你拿到组织时——在尸检中——人已经死了，什么都不在分裂。 For decades there was no way to ask the question directly, only indirect hints that pointed both ways.几十年里，都没有办法直接问这个问题，只有指向两个方向的间接暗示。

The trick that broke it open is almost absurd, and it comes from the worst thing humans did to the atmosphere.把它撬开的那个办法近乎荒诞，而它来自人类对大气做过的最糟糕的一件事。 Between the mid nineteen fifties and nineteen sixty three, the United States and the Soviet Union tested hundreds of nuclear weapons above ground.在二十世纪五十年代中期到一九六三年之间，美国和苏联在地面以上测试了数百件核武器。 Those blasts flung neutrons into the air, and the neutrons turned ordinary nitrogen into carbon fourteen, the radioactive form of carbon.那些爆炸把中子抛进空气，而中子把普通的氮变成了碳14，也就是碳的放射性形态。 In about eight years the amount of carbon fourteen in the air roughly doubled across the northern hemisphere.大约八年间，整个北半球空气中碳14的量大致翻了一番。 Then, in nineteen sixty three, the test ban treaty stopped the atmospheric blasts, and the level began to fall, not because the carbon decayed, but because the oceans and the growing plants drank it down year by year.然后，在一九六三年，禁试条约叫停了大气层核爆，那个水平开始下降——不是因为碳衰变了，而是因为海洋和不断生长的植物年复一年地把它吸了下去。 So the air carries a dated signature.所以空气带着一个可以定年的签名。 Every year since the peak has a slightly lower carbon fourteen level than the year before, a smooth downhill curve that we have measured precisely.从峰值起的每一年，碳14水平都比前一年略低，形成一条平滑的下坡曲线，我们已经把它精确地测量过了。

Now follow the carbon. Plants pull it out of the air. We eat the plants, or we eat the animals that ate them.现在跟着碳走。植物把它从空气中吸出来。我们吃植物，或者吃那些吃了植物的动物。 The carbon ends up in everything we build our bodies from, including, crucially, the DNA inside a cell. And here is the key.碳最终进入我们用来构筑身体的一切之中，其中关键性地，包括细胞内部的 DNA。而这就是要害所在。 When a cell divides, it copies its DNA fresh, using carbon from whatever you have been eating lately.当一个细胞分裂时，它会用你近来所吃东西里的碳，重新复制自己的 DNA。 From that moment, the DNA in that cell is locked. It does not get rewritten unless the cell divides again.从那一刻起，那个细胞里的 DNA 就被锁定了。除非细胞再次分裂，否则它不会被改写。 So the carbon fourteen level frozen into a cell's DNA is a timestamp. It tells you the year that cell was last born.所以冻结在一个细胞 DNA 中的碳14水平，就是一枚时间戳。它告诉你那个细胞最后一次诞生是在哪一年。 A Swedish group, Jonas Frisen's lab at the Karolinska Institute, realized you could read that timestamp.瑞典的一个研究组，卡罗林斯卡学院 Jonas Frisen 的实验室，意识到你可以读取这个时间戳。 Take heart muscle from people who lived through the bomb era, measure the carbon fourteen in the DNA of their heart cells, and compare it to the atmospheric curve.取来那些经历过核弹年代的人的心肌，测量他们心肌细胞 DNA 中的碳十四含量，再和大气曲线做对比。 If the cells were all made before birth, the DNA would read old, from before the spike.如果这些细胞全都是出生前生成的，那么 DNA 读出来会显得很老，来自峰值之前。 If the heart makes new cells, some of the DNA would read younger than the person's birthday. They published the answer in two thousand nine.如果心脏会生成新细胞，那么其中一部分 DNA 读出来会比这个人的生日更年轻。他们在 2009 年发表了答案。

The heart does make new muscle cells. And the rate is tiny.心脏确实会生成新的肌肉细胞。而且这个速率非常小。 At age twenty five, about one percent of your heart muscle cells are replaced in a year.在 25 岁时，你的心肌细胞每年大约有 1% 被更新。 By seventy five it has drifted down to under half a percent.到 75 岁时，这个比例已经下滑到不足 0.5%。 Add it all up and fewer than half the muscle cells you die with are ones you grew after birth.全部累加起来，你去世时的心肌细胞中，出生后新长出来的还不到一半。 So the old dogma was wrong, but it was wrong the way a rounding error is wrong. The heart renews, just barely, and more slowly as you age.所以旧的教条错了，但它错的方式就像一个舍入误差那样。心脏会更新，只是勉强更新，而且随着年龄增长越来越慢。 Both sides of the argument had missed this, because a one percent signal is invisible unless you have a clock, and the bomb gave them a clock.争论的双方都错过了这一点，因为 1% 的信号是不可见的，除非你有一个时钟——而核弹给了他们一个时钟。

Now the part that costs money and trust. While the careful measurement was being made, a much louder claim was already out in the world.现在说到耗费金钱与信任的部分。就在这个审慎的测量正在进行的同时，一个响亮得多的说法早已流传于世。 Starting around two thousand one, a cardiologist named Piero Anversa, eventually at Harvard and the Brigham hospital in Boston, reported that the heart contains its own stem cells, a reserve population he labeled by a marker called c kit, cells that could rebuild lost muscle if you could just wake them up.大约从 2001 年开始，一位名叫 Piero Anversa 的心脏病学家——他后来在哈佛以及波士顿的布莱根医院任职——报告说心脏含有自己的干细胞，一个他用一种名为 c-kit 的标记物来标定的储备细胞群，这些细胞只要你能唤醒它们，就能重建失去的肌肉。 This was the dream. Not a heart that renews at a crawl, but a heart with a dormant repair crew waiting for the call.这就是那个梦想。不是一颗以爬行速度更新的心脏，而是一颗拥有休眠修复队伍、等待召唤的心脏。 It launched a whole field, and clinical trials in real patients. It was not real.它催生了整整一个领域，以及在真实患者身上进行的临床试验。而它并不是真的。

In two thousand eighteen, after years of doubt, Harvard and the Brigham concluded that thirty one papers from Anversa's lab contained fabricated or falsified data and asked the journals to retract them.2018 年，在多年的质疑之后，哈佛和布莱根医院认定 Anversa 实验室的 31 篇论文含有伪造或篡改的数据，并要求期刊撤稿。 The hospital paid ten million dollars to the United States government to settle claims that the work had been used to win federal grants.该医院向美国政府支付了 1000 万美元，以了结关于这些工作曾被用来赢得联邦经费的指控。 A major trial built on the idea was halted. The lab was already closed.一项建立在这一想法之上的大型试验被叫停。实验室此时已经关闭。 And notice what the quiet carbon measurement had been saying the whole time.而请注意那个安静的碳测量自始至终在说什么。 The turnover is small, and the new cells come from ordinary muscle cells dividing, not from some special stem cell reservoir.更新量很小，而且新细胞来自普通肌肉细胞的分裂，而非来自某种特殊的干细胞储库。 The honest number was there, contradicting the miracle, years before the fraud collapsed.那个诚实的数字一直都在那里，与奇迹相矛盾，比这场欺诈崩塌早了许多年。 The fraud grew precisely in the gap between the small true number and the large wished for one.这场欺诈恰恰生长在那个小小的真实数字与那个庞大的、被人期盼的数字之间的缝隙里。

So where do the few new cells actually come from, and why are there so few? The answer came from animals that kept the trick we lost.那么这为数不多的新细胞究竟来自哪里，又为什么如此之少？答案来自那些保留了我们已失去的这一本领的动物。 A zebrafish can have a chunk of its heart cut out and grow it back within two months.斑马鱼可以被切掉一块心脏，并在两个月内把它长回来。 A newborn mouse, in its first few days of life, can do the same.一只新生小鼠，在它生命最初的几天里，也能做到同样的事。 But by about one week old the mouse loses it, and its heart muscle switches over to the permanent, non dividing state we are stuck with.但大约到一周大时，小鼠就失去了这种能力，它的心肌切换到我们所困于其中的那种永久性、不再分裂的状态。 When researchers traced where the regrown muscle comes from, it was not a stem cell.当研究人员追踪重新长出的肌肉来自何处时，答案并不是干细胞。 It was the existing muscle cells briefly dismantling themselves, dividing, and rebuilding.而是现有的肌肉细胞短暂地把自己拆解开、分裂、再重建。 Mammals can do this as newborns and then trade it away.哺乳动物在新生阶段能做到这一点，然后把它拿去交换掉了。 We give up the ability to regrow heart muscle in exchange for muscle cells that are bigger, stronger, packed with contraction machinery, and able to pump against adult blood pressure without ever pausing.我们放弃了重新长出心肌的能力，换来的是更大、更强、装满收缩机器、并且能够对抗成年血压持续泵血而从不停歇的肌肉细胞。 It is a trade, not a flaw. Relentless performance, paid for with the loss of repair.这是一场交易，而非一个缺陷。不知疲倦的性能，代价是失去修复的能力。

Which brings us back to this week's study, and lets us read it honestly. The Sydney group, led by Robert Hume, did something genuinely new.这就把我们带回本周这项研究，也让我们能诚实地读它。由 Robert Hume 领衔的悉尼团队做了一件真正新颖的事。 They collected living heart tissue from patients during bypass surgery, rather than waiting for an autopsy, and they watched for cells actually in the act of dividing.他们没有等到尸检，而是在患者接受搭桥手术期间采集了活体心肌组织，并观察那些正处于分裂过程中的细胞。 They found that after injury the rate of division ticks up.他们发现，损伤之后分裂的速率会有所上升。 That is the first direct sign in living human hearts of the same repair response seen in mice. But listen to what Hume himself said.这是活体人类心脏中，首次直接观察到与小鼠身上相同的修复反应的迹象。但请听听 Hume 本人是怎么说的。 It is not enough. A heart attack can kill roughly a third of the muscle in the affected region.这还不够。一次心脏病发作可以杀死受累区域大约三分之一的肌肉。 A process that renews half a percent a year, even nudged upward, cannot refill that. The scar stays. So here is the ledger, kept straight.一个每年只更新半个百分点的过程，哪怕被略微推高，也无法填补那么多。疤痕留了下来。所以,把这笔账算清楚。

What is proven. The adult human heart renews its muscle, slowly, from existing muscle cells, and the rate rises a little after injury.已被证实的是：成年人类心脏确实会从现有的肌肉细胞出发，缓慢地更新其肌肉，而且损伤之后这一速率会略有上升。 What was oversold. That the heart can heal itself, that heart failure could be reversed by waking a hidden stem cell.被过度吹嘘的是：心脏能够自我修复，心力衰竭可以通过唤醒一种隐藏的干细胞而逆转。 That stem cell was never found, and the most confident version of it was fabricated. What the evidence actually supports.那种干细胞从未被找到，而其中最信誓旦旦的版本是伪造出来的。证据真正支持的是—— A real but feeble repair program, dialed almost to off during our first week of life, which we might one day learn to dial back up.一个真实但微弱的修复程序，在我们生命的第一周几乎被调到关闭状态，而有朝一日我们或许能学会把它重新调高。 That is a genuine goal. It is not a cure that exists. And the lesson this listener will recognize.那是一个真切的目标。它并不是一种已经存在的疗法。而这位听众会认出的教训是——

The elegant measurement did two jobs at once. It proved the heart renews, and it capped how much.那个精巧的测量同时完成了两件事。它证明了心脏会更新，也给出了更新幅度的上限。 The same number that overturned one dogma quietly refuted the opposite dream.推翻一个教条的那个数字，也悄悄驳倒了与之相反的那个梦想。 The trouble is that a true small number is a weaker headline than a false large one, and for nearly twenty years the false large one won. 麻烦在于，一个真实的小数字，作为标题不如一个虚假的大数字有分量，而在将近二十年里，虚假的大数字占了上风。 The clock was in the bomb all along. Somebody just had to agree to read it.时钟一直就在那颗炸弹里。只是需要有人愿意去读它。

This month, two kinds of headline ran within days of each other.这个月，有两类头条新闻在几天之内接连出现。 One said dark energy was dead, that the accelerating universe had turned out to be an illusion.一类说暗能量已死，说加速膨胀的宇宙其实是一场幻觉。 The other said dark energy had survived a major challenge, that the universe is still accelerating after all. Same month. Same field.另一类说暗能量挺过了一次重大挑战，说宇宙终究仍在加速。同一个月。同一个领域。 In one case the same journal, back to back. And much of it the same data — a catalogue of just over seventeen hundred exploding stars.其中一例还发在同一份期刊上，前后脚刊出。而且用的很大程度上是同一批数据——一份仅一千七百多颗爆炸恒星的星表。

That should stop you.这该让你停下来想想。 When competent people read the same measurements and reach opposite conclusions, the question worth asking is not who is right.当有能力的人读同一批测量数据却得出相反的结论，值得问的问题不是谁对。 It is what kind of question this is, that lets it happen at all. So: what was actually measured, and what was inferred.而是这究竟是一类什么样的问题，才会让这种情况发生。所以：到底测量了什么，又推断出了什么。

The gap between those two words is the whole story. Start with the tool. A Type Ia supernova is a particular kind of explosion.这两个词之间的落差，就是全部故事所在。先从工具说起。Ia 型超新星是一种特定类型的爆炸。

A dead star, a white dwarf, pulls matter off a companion until it reaches a fixed weight, and at that weight it detonates.一颗死亡的恒星，一颗白矮星，从伴星身上吸走物质，直到达到一个固定的质量，而在这个质量上它会引爆。 Because the trigger is always the same weight, the explosions come out at nearly the same brightness.由于触发的始终是同一个质量，这些爆炸出来的亮度几乎相同。 So they work like streetlights of known wattage: how bright one looks tells you how far away it is. Faint means far.所以它们的作用就像已知瓦数的路灯：一盏看上去有多亮，就能告诉你它离你有多远。暗意味着远。 Through the nineteen-nineties astronomers collected these explosions across billions of light years.整个 1990 年代，天文学家在跨越数十亿光年的范围内收集这些爆炸。

In nineteen ninety-eight two teams, one led by Saul Perlmutter, the other by Brian Schmidt and Adam Riess, reported the same surprise.1998 年，两个团队——一个由 Saul Perlmutter 领衔，另一个由 Brian Schmidt 和 Adam Riess 领衔——报告了同一个意外。 The distant supernovae were too faint.遥远的超新星太暗了。 By around a fifth, fainter than they should have been if the expansion of the universe were slowing down under its own gravity, which is what everyone expected.如果宇宙的膨胀正在自身引力作用下减速——这也是所有人所预期的——那么它们比本该有的亮度暗了大约五分之一。 Too faint means too far.太暗意味着太远。 Too far means that in the billions of years since that light set out, the universe had carried those explosions further away than a coasting or slowing cosmos would.太远意味着，在那束光出发以来的数十亿年里，宇宙已经把那些爆炸带到了比一个匀速滑行或减速的宇宙更远的地方。 The expansion was speeding up. And to speed up, it needs something pushing. They called that something dark energy.膨胀正在加速。而要加速，就需要某种东西在推动。他们把那种东西称为暗能量。 It won the Nobel Prize in twenty eleven. Now here is the move that most tellings skip. Nobody saw the universe accelerate. You cannot.它在 2011 年赢得了诺贝尔奖。而现在要说的，是大多数讲述都跳过的一步。没有人看见宇宙加速。你看不见。

What you see is a dot: a brightness and a redshift, for each supernova.你看见的是一个点：对每颗超新星而言，是一个亮度和一个红移。 To turn a scatter of dots into the sentence the expansion is speeding up, you have to fit a model of the whole history of the cosmos to those dots and read the answer off the model.要把一堆散点变成「膨胀正在加速」这句话，你必须用一个描述整个宇宙历史的模型去拟合这些点，再从模型上把答案读出来。 And that model rests on two assumptions.而这个模型建立在两个假设之上。 First, that the universe is smooth and the same in every direction, so that a single expansion rate can describe the whole thing.第一，宇宙是平滑的，在各个方向上都相同，这样单一的膨胀率就能描述整个宇宙。 Second, that a supernova ten billion years ago was the same as one today, so that faint means only far and nothing else.第二，一百亿年前的一颗超新星与今天的一颗相同，这样暗就只意味着远，而不意味着别的什么。 Neither of those is something you observe. They are things you assume, so that the dots will speak.这两点都不是你能观测到的东西。它们是你假设出来的，好让这些点开口说话。

This is what anyone who works on inverse problems will recognise at once. You never measure the thing you actually care about.任何研究反问题的人都会一眼认出这一点。你从来测不到你真正关心的那个东西。 You measure something downstream, and read the cause back through a model.你测的是下游的某个东西，再通过一个模型把原因反推回去。 And the standing danger is that more than one cause can produce the same signal. The data then cannot choose between them. You have to.而始终存在的危险是：不止一个原因能产生同一个信号。这时数据本身无法在它们之间做出选择。你得来选。 Using assumptions the data itself is silent about. And that is exactly what has happened here.而选择所依据的假设，恰恰是数据本身保持沉默的。而这正是这里所发生的事。

Take the same seventeen hundred supernovae, and you can fit them several ways. The standard way keeps both assumptions and adds dark energy.拿同样这一千七百颗超新星，你可以用好几种方式去拟合。标准的方式保留两个假设，再加上暗能量。

That is the textbook. But drop the first assumption, that the universe is smooth. It is not, really.这是教科书里的说法。但把第一条假设拿掉——宇宙是均匀的。其实并非如此。

It is voids, enormous empty bubbles, separated by thin walls where the galaxies live.它是一个个空洞，巨大的空泡，被星系所在的薄壁隔开。 A New Zealand physicist named David Wiltshire has argued for years that this lumpiness matters more than we admit, because clocks do not tick at the same rate everywhere.一位名叫 David Wiltshire 的新西兰物理学家多年来一直主张，这种不均匀性比我们承认的更重要，因为各处的钟并不以相同的速率走时。 A clock in the middle of an empty void runs ahead of a clock inside a dense wall. There is no single cosmic time to hand out to everyone.空洞中央的钟比致密壁内的钟走得快。没有一个可以分发给所有人的统一宇宙时间。 When he fits the supernovae with that in mind, and with no dark energy at all, the fit comes out competitive with the standard one.当他带着这一点、并且完全不引入暗能量去拟合超新星时，得到的拟合与标准模型不相上下。 In his picture part of the acceleration is a bookkeeping error.在他的图景里，一部分加速是一种记账误差。 The error you make by pretending two clocks that run at different rates are the same clock.也就是你把两个以不同速率走时的钟当成同一个钟时所犯的误差。

Now drop the other half of that assumption, that the universe looks the same in every direction.现在再把那条假设的另一半拿掉——宇宙在每个方向上看起来都一样。 Subir Sarkar, at Oxford, with colleagues at the Tata Institute in Mumbai, has found something pointed.牛津的 Subir Sarkar，与孟买塔塔研究所的同事一起，发现了一个很有指向性的现象。 The apparent acceleration is not uniform across the sky.表观的加速在全天空并不均匀。 It is strongest along the direction we happen to be moving, toward a hotspot in the afterglow of the Big Bang, and it fades as you look further out.它在我们恰好正在运动的那个方向上最强——朝向大爆炸余辉中的一个热点——而当你往更远处看时，它就减弱。 That is a strange thing for dark energy to do.对暗能量来说，这是件很奇怪的事。 Dark energy is supposed to be the energy of empty space itself, and empty space has no favourite direction.暗能量本应是空间本身的能量，而空间没有偏爱的方向。 A signal that points somewhere looks less like the vacuum and more like our own local motion, never quite subtracted off.一个指向某处的信号，看起来不太像真空，倒更像是我们自身的局部运动，从未被彻底扣除干净。

And there is a third thread, the one behind this month's louder headline. Drop the second assumption, that a supernova is a supernova.还有第三条线索，也就是本月更大标题背后的那条。把第二条假设拿掉——超新星就是超新星。 Explosions from young populations of stars may be very slightly dimmer, on average, than ones from old populations.来自年轻恒星族的爆发，平均而言可能比来自年老恒星族的爆发略微暗一点点。 And the mix of young and old shifts as you look back in time.而年轻与年老的混合比例会随着你回溯时间而变化。 If the distant supernovae come from systematically younger stars, and are intrinsically a touch fainter, they would look farther away than they really are.如果遥远的超新星系统性地来自更年轻的恒星，本征上稍微更暗一些，它们看起来就会比实际更远。 Which would mimic acceleration exactly.而这恰好会模拟出加速的假象。 Sarkar's group applied a correction for this stellar age to the catalogue, and reported that the sign flipped. Not acceleration.Sarkar 的团队对目录中的这一恒星年龄效应做了修正，并报告称符号发生了翻转。不是加速。 Deceleration. So is dark energy dead? No. And here the refereeing matters.而是减速。那么暗能量是不是已经死了？没有。而在这里，同行评议的把关很关键。

In the same journal, another group led by Maria Vincenzi found the data still accelerating.在同一份期刊上，由 Maria Vincenzi 领衔的另一个团队发现数据仍然在加速。 And a team at Southampton took the age-correction study apart and found two concrete problems.而南安普顿的一个团队把那项年龄修正的研究拆开来看，发现了两个具体问题。 It had left out a standard, well-known step — that supernovae in big galaxies come out slightly brighter than ones in small galaxies.它漏掉了一个标准的、众所周知的步骤——大星系中的超新星比小星系中的略微更亮。 And it had assumed a galaxy's average age was the age of the particular stars that exploded, which overstates the gap by a factor of three to five, because even ancient galaxies keep pockets of young stars.而且它假定一个星系的平均年龄就是那批发生爆炸的特定恒星的年龄，这把差距高估了三到五倍，因为即便是古老的星系也保有一小片一小片的年轻恒星。 Put both of those back, and the deceleration goes away. As things stand, the flip looks fragile.把这两点都补回去，减速就消失了。就目前而言，这次翻转看起来很脆弱。

But do not swing to the opposite overselling, that acceleration is simply settled and done.但也别摆到另一个过度兜售的极端，说加速已经板上钉钉、盖棺定论。 Two things keep it standing, both worth saying plainly. First, the supernovae are not the only evidence.有两点让它站得住脚，都值得直白地说出来。第一，超新星并不是唯一的证据。 The afterglow of the Big Bang tells us the universe is geometrically flat, which fixes its total budget of energy.大爆炸的余辉告诉我们，宇宙在几何上是平直的，这就固定了它的总能量预算。 Add up all the matter there is, ordinary and dark, and you fall about thirty percent short. Something else has to make up the difference.把所有物质加起来——普通物质和暗物质——你还差大约百分之三十。必须有别的东西来补上这个差额。 And the faint ripples of sound frozen into the arrangement of galaxies give a completely separate ruler for the expansion, and it too calls for dark energy.而那些微弱的声波涟漪，被冻结在星系的排布之中，为宇宙膨胀提供了一把完全独立的标尺，它同样需要暗能量。 Three legs, then, not one. To kill dark energy you would have to explain away all three. And yet.所以是三条腿，而不是一条。要想干掉暗能量，你得把这三条腿全都解释掉。然而。

Every one of those three legs is read inside the very same assumption, a smooth and evenly spread universe. 这三条腿中的每一条，都是在同一个假设之下被解读的，那就是一个平滑而均匀铺展的宇宙。 The premise that Wiltshire and Sarkar are prodding is not tucked away inside the supernovae alone.Wiltshire 和 Sarkar 所戳的那个前提，并不只是藏在超新星这一条里。 It is the frame all three measurements are read in.它是所有三项测量被解读时所共用的框架。 So the famous concordance, the way three independent methods seem to agree, is a concordance of models that quietly share one premise.所以那著名的一致性，也就是三种独立方法看起来彼此吻合的方式，其实是一堆悄悄共享同一前提的模型之间的一致性。 That does not make it wrong. It makes it less of an independent triple-check than it is usually sold as.这并不意味着它错了。它只是意味着，它并不像人们通常兜售的那样，是一次独立的三重检验。

And there is a third position, quieter, and gaining ground, that people keep confusing with the first.此外还有第三种立场，更安静，正在赢得地盘，而人们总是把它和第一种混为一谈。 This year the large galaxy survey called DESI, and a new compilation out of the University of Queensland of nearly three thousand supernovae, both hint that dark energy is real but not constant.今年，名为 DESI 的大型星系巡天，以及昆士兰大学新汇编的近 3000 颗超新星数据，都暗示暗能量是真实存在的，但并非恒定不变。 That it has been weakening over cosmic time. That is not no dark energy. It is dark energy, but not the simple unchanging kind.它一直在随宇宙时间而减弱。这不是没有暗能量。这是暗能量，只不过不是那种简单而不变的类型。 Keep the three apart. No acceleration at all: radical, contested, and right now looking weak.把这三者分开来看。完全没有加速：激进、有争议，而且眼下看起来站不住脚。 Acceleration driven by a constant: the standard picture.由一个常数驱动的加速：标准图景。 Acceleration driven by something that changes: the interesting middle, and where the smart money seems to be drifting.由某种会变化的东西驱动的加速：有意思的那个中间地带，而聪明的赌注似乎正朝这边漂移。

So what is the lesson to carry out of this. Not that scientists cannot tell whether the universe is accelerating. That is the lazy read.那么该从这里带走的教训是什么。不是科学家没法判断宇宙是否在加速。那是偷懒的读法。 The lesson is sharper. The most important number in cosmology was never measured.教训更为锋利。宇宙学中最重要的那个数字，从来就没有被测量过。 It was inferred, read back through a model from the faintness of a scatter of dots.它是被推断出来的，是透过一个模型、从一撮暗点的暗弱程度里反推回来的。 The inference rests on an assumption, that the universe is smooth and the same everywhere, that has never been directly confirmed.这个推断建立在一个假设之上，即宇宙是平滑的、处处相同的，而这个假设从未被直接证实过。 Only assumed, because it makes the mathematics work. And when several universes fit the same dots, shouting the data louder does not help.只是被假定了，因为它能让数学算得通。而当好几个宇宙都能拟合同一撮暗点时，把数据喊得更大声也没有用。 What helps is finding a test the models disagree on. Sarkar's is exactly that.有用的是找到一个模型之间会各执一词的检验。Sarkar 的检验正是如此。 A real vacuum energy cannot point in a direction, so if the signal points, that tells you something the brightness alone never could.真正的真空能不可能指向某个方向，所以如果信号有指向性，那它告诉你的东西，是单凭亮度永远给不出来的。

Here is the honest sentence. The universe is very probably accelerating.下面是那句诚实的话。宇宙很可能确实在加速。 But acceleration is a conclusion we compute, not a thing we see, and the premise underneath it, that the cosmos is smooth, is the load-bearing wall that nobody has actually inspected.但加速是我们算出来的一个结论，而不是我们看到的一样东西，而它底下那个前提，也就是宇宙是平滑的，是那面承重墙，却没有人真正检查过。 The fight this month was never about the universe dying. It was about whether we have ever really looked behind that wall.这个月的这场争论，从来就不是关于宇宙的死亡。它关乎的是，我们究竟有没有真正往那面墙背后看过。 And slowly, tediously, exactly as they should, a few people are finally trying to.而缓慢地、乏味地，正如他们本该做的那样，终于有少数几个人开始尝试去看了。

Ada Yonath died on the thirty-first of August, at eighty-seven.Ada Yonath 于 8 月 31 日去世，享年 87 岁。 She was the first Israeli woman to win a Nobel Prize, and the first woman to win the chemistry prize in forty-five years.她是第一位获得诺贝尔奖的以色列女性，也是 45 年来第一位获得化学奖的女性。 The obituaries all reach for the same word, which is perseverance, because for about twenty years a large part of her own field told her she was wasting her life.所有的讣告都用到了同一个词——坚持，因为大约有二十年时间，她自己领域里的很大一部分人都告诉她，她是在浪费自己的人生。 She liked to repeat what they called her. A dreamer. The village fool. The so-called scientist.她喜欢重复别人给她起的称呼。一个梦想家。村里的傻瓜。所谓的科学家。 I want to take that seriously tonight, because the easy version of her story is that sheer stubbornness beat an impossible problem, and the easy version is not quite true.今晚我想认真对待这一点，因为她故事的简化版本是：纯粹的固执击败了一个不可能的难题——而这个简化版本并不完全属实。 Stubbornness bought her the time. Something more specific won. Start with what she was trying to see.固执为她争取到了时间。真正取胜的是某种更具体的东西。先从她想要看清的东西说起。

Every living cell is full of machines that build proteins, and the machine is called the ribosome. Think of it as a two-part reading head.每一个活细胞里都充满了制造蛋白质的机器，这台机器叫做核糖体。可以把它想象成一个两部分组成的读取头。 A messenger tape comes in carrying the instructions copied from your genes, and the ribosome runs along it three letters at a time.一条信使带子进来，携带着从你的基因中拷贝出来的指令，核糖体沿着它一次读取三个字母地运行。 For each three-letter word it grabs the matching building block and welds it onto a growing chain. That chain, folded up, is a protein.对于每一个三字母的词，它抓取匹配的构件，焊接到一条不断延长的链上。这条链折叠起来，就是一个蛋白质。 This is happening inside you right now, millions of times a second, in every cell you have.此刻这件事正在你体内发生，每秒钟数百万次，在你的每一个细胞里。 The ribosome is the thing that turns the genetic code from writing into substance.核糖体就是把遗传密码从文字变成实体的那个东西。 If you want to understand life at the level of moving parts, this is close to the central part.如果你想在运动部件的层面上理解生命，这几乎就是最核心的部分。 And in the nineteen-seventies, nobody knew what it looked like. To see something that small you use X-ray crystallography.而在 1970 年代，没有人知道它长什么样。要看清如此小的东西，你得用 X 射线晶体学。

You cannot simply photograph a single molecule, because it is far smaller than the wavelength of visible light.你没法简单地给单个分子拍照，因为它远小于可见光的波长。 So you do something indirect.所以你要用一种间接的办法。 You get billions of identical copies to line up in a perfect repeating grid, a crystal, and you shine X-rays through it.你让数十亿个完全相同的拷贝排列成一个完美的重复网格——一块晶体——然后让 X 射线穿过它。 The copies all scatter the beam the same way, the scattered waves reinforce each other into a pattern of spots you can record, and with a great deal of mathematics you run that pattern backwards into a three-dimensional map.这些拷贝都以相同的方式散射光束，散射的波彼此加强，形成一个你可以记录下来的斑点图案，再借助大量的数学运算，把这个图案反演成一张三维图。 Chemists had been doing this to proteins for decades. But the whole thing depends on the first step, getting the molecules to line up.化学家用这种方法研究蛋白质已经有几十年了。但整件事都取决于第一步——让分子排列起来。 Small, rigid molecules crystallize willingly. The ribosome is neither small nor rigid. It is enormous, by molecular standards.小而刚性的分子会心甘情愿地结晶。核糖体既不小也不刚性。以分子的标准来说，它极其庞大。 It is a loose assembly of dozens of parts. It is half protein and half RNA. And it flexes constantly, because flexing is its job.它是由数十个部件松散组装而成的。它一半是蛋白质，一半是 RNA。而且它不停地弯曲，因为弯曲正是它的工作。 Asking it to sit still in a perfect grid struck most people as like asking a bowl of spaghetti to freeze into a diamond.要它在一个完美的网格里乖乖不动，在大多数人看来，就像要一碗意大利面冻结成一颗钻石。 That is why they laughed. Here is the first thing that actually worked, and it came from a bicycle accident.这就是他们发笑的原因。下面是第一件真正奏效的事，而它源于一场自行车事故。

Yonath was laid up recovering, reading, and she came across an account of hibernating bears.Yonath 卧床养伤，读着书，偶然读到一段关于冬眠熊的记述。 When a bear goes dormant, its cells stop making protein for months, and rather than let the idle ribosomes fall apart, the cell packs them away in neat, ordered arrays, ready to switch back on in spring.当一头熊进入休眠，它的细胞会连续数月停止制造蛋白质，而细胞并不让这些闲置的核糖体散架，而是把它们整齐有序地打包收好，等到春天再重新启动。 She read that and thought, roughly, if a bear can order its ribosomes, why can't I.她读到这段，心里大致想：如果一头熊能把它的核糖体排列整齐，我为什么不能。 Nature was already crystallizing the thing, for its own reasons, to survive the winter. That quietly reframed the problem.自然界为了它自己的理由，为了熬过冬天，早就在给这个东西结晶了。这悄然地重新定义了这个难题。 It was not impossible. It was a problem of persuading the molecule that it was going dormant.它并非不可能。它只是一个说服分子相信自己正在进入休眠的问题。

The second thing that worked was choosing tougher organisms.第二件奏效的事，是选择更强韧的生物。 Instead of ribosomes from ordinary bacteria, she went hunting in extreme places. Bacteria from hot springs. Microbes from the Dead Sea.她没有用普通细菌的核糖体，而是到极端的地方去搜寻。来自温泉的细菌。来自死海的微生物。 Organisms that shrug off radiation.那些对辐射不以为然的生物。 Their ribosomes have evolved to hold together under heat and salt and stress, which makes them stiffer, more uniform, more willing to line up.它们的核糖体已经进化到能在高温、盐分和各种胁迫下保持稳定，这让它们更坚硬、更均一、更愿意排列成阵。 The sample was not a minor detail of the experiment. In a real sense the sample was the experiment.样品并不是实验中的一个次要细节。在某种真实的意义上，样品就是实验本身。 Find the right ribosome and half the problem dissolves before you start.找到合适的核糖体，一半的难题在你动手之前就已经化解了。

The third thing became the most important beyond her own work, and it solved a trouble I have not mentioned yet.第三件事的重要性超出了她自己的研究，它还解决了一个我尚未提及的麻烦。 The very X-rays that read the crystal also destroy it.读取晶体的那些 X 射线，同时也在摧毁它。 A ribosome crystal is delicate, and a strong beam burns through it before you can finish measuring.核糖体晶体很脆弱，强束流会在你测完之前就把它烧穿。 Yonath's answer was to freeze the crystals, cool them to nearly two hundred degrees below zero, where the damage slows down enough to collect a full picture before the crystal dies.Yonath 的答案是把晶体冷冻起来，将它们冷却到接近零下两百度，在那样的温度下损伤放缓到足以在晶体死去之前采集到完整的图像。 This is called cryo-crystallography, and here is the part worth holding onto. It is now how almost all of structural biology is done.这被称为冷冻晶体学，而下面这一点值得记住：如今几乎所有的结构生物学都是这么做的。 Nearly every protein structure you have ever seen a picture of was read from a flash-frozen crystal.你见过图片的几乎每一个蛋白质结构，都是从急速冷冻的晶体中读取出来的。 She developed the technique for a problem everyone thought was hers alone, and it turned into a tool for the whole field.她为一个人人都以为只属于她自己的问题开发了这项技术，而它最终变成了整个领域的工具。 That is often how method work pays off. It looks like a private workaround and becomes public infrastructure. 方法性的工作往往就是这样得到回报的。它看起来像是一个私人的权宜之计，最后却成了公共的基础设施。

Even with all three, it took roughly twenty-five thousand attempts to get the first crude crystals, in nineteen-eighty, and another twenty years to sharpen the picture down to individual atoms, around the year two thousand.即便三样都齐备，在一九八零年，也花了大约两万五千次尝试才得到最初的粗糙晶体，又用了二十年才把图像锐化到单个原子的程度，大约在两千年前后。 Along the way, in nineteen-ninety-three, she found something lovely, a tunnel running through the large half of the machine, the chute the new protein slides out through as it is assembled.一路走来，在一九九三年，她发现了一样美妙的东西：一条贯穿这台机器较大那一半的隧道，也就是新合成的蛋白质在被组装时滑出去的那条通道。 Nobody had known it was there. Now the part where I have to be careful, because this is where the popular telling blurs the credit.此前没有人知道它的存在。现在到了我必须谨慎的部分，因为正是在这里，通俗的讲述模糊了功劳的归属。

The Nobel in two thousand and nine went to three people. Yonath, Venkatraman Ramakrishnan, and Thomas Steitz.二〇〇九年的诺贝尔奖颁给了三个人：Yonath、Venkatraman Ramakrishnan 和 Thomas Steitz。 They did not do the same thing.他们做的并不是同一件事。 The single most startling discovery to come out of these structures is that the ribosome is what is called a ribozyme.从这些结构中得出的最令人震惊的单一发现，是核糖体属于所谓的核酶（ribozyme）。 When you finally look at the exact spot where the chemistry happens, where the new bond is actually welded, there is no protein there.当你最终看向化学反应发生的确切位置，也就是新化学键真正被焊接起来的地方，那里没有蛋白质。 The catalysis is done by RNA.催化是由 RNA 完成的。 That mattered enormously, and it was seen most sharply in Steitz's structure of the large subunit, not Yonath's.这一点意义极其重大，而它在 Steitz 关于大亚基的结构中被看得最为清晰，而非 Yonath 的结构。 Her own crown was the small subunit, the reading head, and the tunnel, and the antibiotics.她自己的桂冠是小亚基——那个阅读头——以及那条隧道和那些抗生素。 It takes nothing away from her to say this plainly.把这一点直白地说出来，丝毫无损于她的成就。 The prize was shared precisely because three groups converged on one machine from different sides, and the honest account keeps track of who saw what.这个奖之所以被共享，恰恰是因为三个团队从不同的侧面汇聚到同一台机器上，而诚实的叙述会记清楚谁看到了什么。

But sit with the ribozyme point, because it is the deepest thing here.但请在核酶这一点上多停留片刻，因为它是这里最深刻的东西。 Today the cell keeps its master copy in DNA and does its work with proteins, and RNA looks like a messenger in between, a middleman.今天，细胞把它的母版拷贝保存在 DNA 中，用蛋白质来完成工作，而 RNA 看起来像是夹在中间的信使，一个中间人。 Finding that the oldest, most essential machine in every living thing does its central chemistry with RNA is like finding a fossil.发现每一个生命体中最古老、最不可或缺的机器竟用 RNA 来完成其核心化学反应，就像发现了一块化石。 It says that before DNA, before protein enzymes, there was a world running on RNA alone, and the ribosome is a survivor from it, still sitting at the center of every cell, still doing the old job the old way.它表明，在 DNA 之前，在蛋白质酶之前，曾有一个仅靠 RNA 运转的世界，而核糖体是从那个世界幸存下来的遗迹，它仍然坐在每一个细胞的中心，仍然用古老的方式做着古老的工作。 The machine that reads the genetic code is older than the division of labor the code now takes for granted.这台读取遗传密码的机器，比这套密码如今习以为常的分工还要古老。 That is not a small claim, and it is worth saying it is an inference, not a photograph of the past.这不是一个小小的论断，也值得说明它是一个推论，而非一张过去的照片。 But it is an inference the structure supports rather than one it merely allows.但它是一个由结构所支持的推论，而不仅仅是一个结构允许成立的推论。

There is also a payoff you can carry to a pharmacy, and it needs the same honesty.还有一项成果可以带进药房，它同样需要这份诚实。 More than half of the antibiotics we use work by jamming the bacterial ribosome.我们使用的抗生素中，超过一半是通过卡住细菌的核糖体来起作用的。 They drop into its gears and stop it building proteins, and they spare you because your ribosomes are shaped differently enough to ignore them.它们嵌进核糖体的齿轮里，让它无法继续合成蛋白质，而它们之所以不伤害你，是因为你的核糖体形状差异足够大，足以对它们视而不见。 Yonath worked out exactly where more than twenty of these drugs bind, and exactly how a single change in the bacterial machine can shrug them off, which is what drug resistance is, seen at the level of atoms.Yonath 精确弄清了其中二十多种药物结合的确切位置，也弄清了细菌机器中一个单一的改变如何就能把它们甩开——这正是耐药性，从原子层面看到的样子。 That is a real map. But a map is not a cure.那是一张真正的地图。但地图不是解药。 Knowing precisely where a drug sits has not, so far, produced a wave of new antibiotics, because the barriers to new drugs are as much economic and biological as they are structural.精确知道一种药物落在哪里，迄今为止并没有催生出一波新抗生素，因为新药面临的壁垒既是结构性的，也同样是经济和生物学上的。 She handed the field the coordinates. The coordinates did not build the road, and it would be overselling her to pretend they did.她把坐标交给了整个领域。坐标并没有修好那条路，若假装它们做到了，就是在过度吹捧她。

So what is the honest shape of this life.那么，这段人生诚实的模样究竟是什么。 A person told for two decades that she was chasing a mirage, who was not carried by faith alone but by three transferable ideas.一个被人告知了二十年在追逐海市蜃楼的人，她靠的不只是信念，而是三个可迁移的想法。 A bear's trick for ordering the sample. A deep freeze to outrun the damage. A hardy organism that already wanted to crystallize.一只熊为样品排序的窍门。一场深度冷冻，用以跑赢损伤。一种本就想要结晶的顽强生物。 Persistence did not solve the problem.坚持并没有解决问题。 Persistence kept her in the room long enough for those ideas to arrive, and then the ideas did the work.坚持只是让她在房间里待得足够久，久到那些想法能够到来，然后是那些想法完成了工作。 And what she and two others finally saw was not only a machine but an ancestor, an RNA relic that has been quietly translating the code of life since before there was much else to translate it with.而她与另外两人最终看到的，不只是一台机器，还是一位祖先——一件 RNA 遗迹，早在几乎没有别的东西可供翻译生命密码之前，它就已经在悄悄地翻译着这套密码了。 The village fool was reading about hibernating bears. And she was right.那个村里的傻瓜当年在读关于冬眠熊的书。而她是对的。

Last day. And we finish with Shannon trying to stop what he had started. By the mid nineteen fifties, information theory had escaped.最后一天。我们以 Shannon 试图阻止他所开创的东西作为收尾。到二十世纪五十年代中期，信息论已经失控了。

It had begun as a tool for telephone engineers and it was being applied to psychology, linguistics, economics, biology, management, literary criticism.它起初是电话工程师的工具，后来却被应用到心理学、语言学、经济学、生物学、管理学、文学批评。 There were conferences. There was press coverage. People were computing the entropy of poems and of organisations and of nervous systems.有各种会议，有媒体报道。人们在计算诗歌的熵、组织机构的熵、神经系统的熵。

And in March nineteen fifty-six, Shannon published a one-page editorial in the IRE Transactions on Information Theory.1956 年 3 月，Shannon 在《IRE Transactions on Information Theory》上发表了一篇一页的社论。 It is titled The Bandwagon. He wrote that information theory had become something of a scientific bandwagon.标题叫 The Bandwagon（跟风潮）。他写道，信息论已经变成了某种科学上的跟风潮。

That having started as a technical tool for the communication engineer, it was receiving extraordinary publicity.它起初是通信工程师的一件技术工具，如今却受到了非同寻常的追捧。 That applications were being made to biology, psychology and linguistics, and that these had, in his words, been taken too literally by enthusiasts.各种应用被搬到生物学、心理学和语言学中，用他的话说，这些应用被热心者们理解得太字面化了。

His central warning is the sentence I want you to have.他的核心告诫，就是我想让你记住的那句话。 Establishing such applications, he wrote, is not a trivial matter of translating words to a new domain, but rather the slow tedious process of hypothesis and experimental verification.他写道，建立起这类应用，并不是把词汇翻译到一个新领域那么简单的事，而是假设与实验验证这一缓慢而枯燥的过程。

That is the founder of a field, at the height of its fashion, telling people that using his vocabulary is not the same as using his theory.这是一个领域的创立者，在它最时兴的巅峰时刻，告诉人们：使用他的词汇，并不等于使用他的理论。

So today: where does it transfer, where does it not, and why. Start with the sharpest failure, because it is the most instructive. 所以今天要谈的是：它能迁移到哪里、不能迁移到哪里，以及为什么。先从最尖锐的失败讲起，因为它最有教益。

Information theory says nothing about meaning.信息论对意义只字未提。

That is not an oversight, it is a design decision, made on the first page of the nineteen forty-eight paper.这不是疏忽，而是一个设计决定，在 1948 年那篇论文的第一页就作出了。 Shannon set semantics aside because he could not make it tractable and did not need it.Shannon 把语义搁在一边，因为他无法让它变得可处理，而且他也不需要它。

Which means the theory measures how much, never what. It will tell you a message carries four point two bits.这意味着这套理论衡量的是有多少，从不涉及是什么。它会告诉你一条消息携带了 4.2 比特。 It will not tell you whether the message was true, useful, important, or about anything at all.它不会告诉你这条消息是否真实、有用、重要，或者究竟是不是关于任何东西的。

So any application where the quantity of interest is meaning is misapplying it.所以任何以意义为关注量的应用，都是在误用它。 The entropy of a poem is a computable number and it is a fact about letter frequencies, not about the poem.一首诗的熵是一个可计算的数字，它是关于字母频率的事实，而不是关于这首诗的事实。

There was a genre of nineteen fifties work computing entropy of texts and organisations and calling the result a measure of information content in the ordinary sense.五十年代有一类工作，计算文本、组织机构的熵，然后把结果称作日常意义上的信息含量的度量。 The numbers were right. The interpretation was borrowed illegitimately from the word.数字是对的，但那种解读是从这个词身上非法借来的。

Second failure mode, subtler and more common today: estimation. We covered this with mutual information, and it generalises.第二种失败模式，更微妙，也在今天更常见：估计。我们在讲互信息时谈过这一点，它可以推广。

Every quantity in this course is defined over a probability distribution you do not have. You have samples.这门课里的每一个量，都是定义在一个你并不拥有的概率分布之上的。你拥有的是样本。 Estimating entropy from limited samples is biased, and estimating mutual information is worse, and the bias runs upward — toward reporting structure that is not there.从有限样本中估计熵是有偏的，估计互信息更糟，而且这种偏差是向上的——朝着报告出并不存在的结构。

The theory is exact. Your estimates are not, and the gap between them is where a great deal of unreliable work lives.理论是精确的。你的估计不是，而两者之间的差距，正是大量不可靠工作栖身之处。 That is not a criticism of the theory; it is a warning about the distance between a mathematical object and a data set.这不是对理论的批评；这是对一个数学对象与一个数据集之间距离的告诫。

Third: the theory is about a specific setup. A sender, a receiver, a channel, a code agreed in advance.第三：这套理论是关于一种特定架构的。一个发送者、一个接收者、一条信道、一套事先约定好的编码。 When people apply it to systems that do not have those parts — an ecosystem, a market, a brain considered as a whole — the mapping is usually metaphorical, and the mathematics does not transfer just because the words do.当人们把它应用到不具备这些部件的系统上——一个生态系统、一个市场、被当作整体来看的大脑——这种对应通常是隐喻性的，数学并不会仅仅因为词汇迁移了就跟着迁移。

Now the successes, because they are substantial and they have a common feature.现在讲成功，因为这些成功很实在，而且它们有一个共同特征。

In biology, the genetic code is genuinely a code, in the technical sense: a mapping from sixty-four triplets to twenty amino acids plus stop, agreed in advance, shared across life.在生物学中，遗传密码在技术意义上确实是一套 code（编码）：一个从 64 个三联体到 20 种氨基酸外加终止信号的映射，事先约定好，被所有生命共享。 And information-theoretic analysis pays off.而信息论式的分析是有回报的。 The code's structure is measurably error-tolerant — similar codons tend to specify similar amino acids, so many single-base errors produce a chemically similar substitution rather than a catastrophic one.这套编码的结构具有可度量的容错性——相似的密码子往往指定相似的氨基酸，因此许多单碱基错误产生的是化学性质相近的替换，而非灾难性的替换。 That is an error-correcting property, measurable in the terms we built on day five, and it is evidence about how the code evolved.这是一种纠错特性，可以用我们在第五天建立的那套术语来度量，它也是关于这套编码如何演化的证据。

In neuroscience, the analysis works where the setup genuinely applies: a stimulus, a neural response, and a question about how much the response tells you about the stimulus.在神经科学中，只要设定确实适用，这套分析就能奏效：一个刺激、一个神经响应，以及一个关于响应能告诉你多少刺激信息的问题。 That is mutual information used correctly, and the field has been forced into real sophistication about estimator bias precisely because the naive approach overestimates so reliably.这正是被正确使用的互信息（mutual information），而这个领域之所以在估计量偏差问题上被迫走向真正的精深，恰恰是因为朴素方法如此可靠地高估结果。

In machine learning it is not an application at all — it is the native language. Cross-entropy loss is a description length.在机器学习里，它根本不是某种应用——它就是母语。交叉熵损失（cross-entropy loss）就是一种描述长度。 KL divergence measures the cost of a wrong model.KL 散度衡量的是一个错误模型的代价。 The information bottleneck framework describes learning as compressing input while preserving information about the target.信息瓶颈（information bottleneck）框架把学习描述为在保留关于目标的信息的同时压缩输入。 When you train a model you are minimising bits, literally. And in your own area, the honest uses are of that kind.当你训练一个模型时，你就是在最小化比特数，字面意义上如此。而在你自己的领域，诚实的用法也是这一类。

Mutual information as a screening tool for nonlinear or threshold-driven dependence, where a linear measure would report nothing.把互信息作为一种筛查工具，用于探测非线性或阈值驱动的依赖关系——在那里线性度量什么也报告不出来。 Minimum description length as a principled account of why an over-parameterised model that fits beautifully should still be distrusted.把最小描述长度（minimum description length）作为一种有原则的说法，解释为什么一个过度参数化、拟合得漂亮无比的模型仍然值得怀疑。 And the framing that a model is a compression of observations, so that a model whose description is as long as the data it explains has not explained it.以及这样一种框定：模型是对观测的压缩，因此一个其描述与它所解释的数据一样长的模型，其实并没有解释任何东西。

Notice what the successes share.注意这些成功案例的共同之处。 In every case there is an actual code, or an actual channel, or an explicit probability distribution, and the mathematics is applied to that object rather than to a resemblance.在每一种情形里，都存在一套真实的编码，或一条真实的信道，或一个明确的概率分布，而数学是被应用在那个对象上，而非应用在某种相似性上。

The test I would offer is: can you write down the distribution? If yes, you can compute.我会给出的检验标准是：你能把那个分布写下来吗？如果能，你就能计算。 If you are reasoning by analogy from the words entropy and information, you are doing what Shannon warned against in nineteen fifty-six.如果你是从熵和信息这两个词出发做类比推理，那你正在做香农（Shannon）在 1956 年警告过的事。

Now let me close the whole course, and there are two things I want to leave you with. The first is the arc.现在让我为整门课程收尾，有两件事我想留给你。第一件是这段历程的轨迹。

We started with Morse counting type in a printer's shop and ended with a colloidal particle in an optical trap measuring the heat from erasing one bit.我们从一家印刷厂里数着莫尔斯电码的字模开始，最终落到光镊中一颗测量抹除一个比特所产生热量的胶体粒子。

In between: a quantity that had to be a logarithm because information must add while possibilities multiply.其间：一个必须是对数的量，因为信息必须相加，而可能性却是相乘。 A floor on compression nobody can go under.一条谁也无法逾越的压缩下限。 A cliff in the noise, below which perfect communication is available and above which nothing works.噪声中的一道悬崖，在它之下完美通信触手可及，在它之上一切都无从谈起。 Forty-five years to find codes that reach it, and the answer sitting unread in a nineteen sixty-three thesis.花了四十五年才找到能触及这条极限的编码，而答案早已静静躺在一篇 1963 年的博士论文里无人翻阅。 A cipher that is provably unbreakable and operationally impossible.一种可被证明无法破解、却在操作上不可行的密码。 A rival definition of information, based on shortest programs, that is uncomputable but tells you what randomness means.一个基于最短程序的、与之抗衡的信息定义，它不可计算，却告诉你随机性意味着什么。 And an exchange rate between bits and joules.以及比特与焦耳之间的一个兑换率。

All of it from one paper by one person, published in a company journal, in nineteen forty-eight.所有这一切，都来自一个人的一篇论文，发表在一份公司内部期刊上，时值 1948 年。

The second thing is the deeper point, and it is about the refusal. Shannon's first move was to declare meaning irrelevant.第二件是更深层的一点，它关乎那一次拒绝。香农的第一步，是宣告意义无关紧要。

That looks like a retreat. It is the opposite: it is the move that made everything possible.这看起来像是退缩。恰恰相反：这正是让一切成为可能的一步。 Every prior attempt to define information had drowned in meaning, because meaning does not submit to mathematics.此前每一次定义信息的尝试都淹没在意义之中，因为意义不服从于数学。 By cutting it away, Shannon found that what remained — pure selection from a set of possibilities, with probabilities — was enough to build a complete theory, prove hard limits, and hand engineering an exact number to aim at.通过把意义切除，香农发现所剩下的东西——从一组可能性中带着概率进行的纯粹选择——已足以构建一套完整的理论，证明严格的极限，并交给工程一个可以瞄准的精确数字。

The lesson generalises beyond this subject.这个教训的意义超出了本学科本身。 Sometimes the way to make a problem tractable is to admit you cannot handle its most interesting part, cut it off cleanly, and see what is left.有时候，让一个问题变得可解的办法，就是承认你处理不了它最有趣的那部分，把它干净利落地切掉，看看剩下什么。 What is left is often more than you expected. But you have to remember you did it.剩下的往往比你预期的要多。但你得记住自己做过这件事。

Almost every misuse of information theory in the last seventy-five years comes from people picking up the theory and forgetting the amputation.过去七十五年里，几乎所有对信息论的误用，都源于人们拿起这套理论，却忘了那次截肢。 They use a quantity built by explicitly excluding meaning, and then interpret its output as if it were about meaning.他们用的是一个明确排除了意义而构建出来的量，然后又把它的输出解读得好像它讲的就是意义。 The word invites the mistake. Von Neumann's joke about nobody understanding entropy turned out to have a long tail.这个词本身就在引诱人犯错。冯·诺依曼那句谁也不懂熵的玩笑，结果余波绵长。

So: what did you actually get, over ten days? A precise definition of information as reduced uncertainty.那么：这十天里你到底得到了什么？一个关于信息的精确定义——信息即被削减的不确定性。

The knowledge that compression has a hard floor and where it is. The knowledge that noise does not have to mean errors.还有这样的认识：压缩有一个硬性下限，以及它在哪里。还有：噪声未必意味着错误。 A tool, mutual information, that sees relationships correlation cannot, along with the specific ways it lies to you.一个工具，互信息（mutual information），它能看到相关性看不到的关系，同时也看到它误导你的那些具体方式。 A second definition of information that applies to single objects and formalises Occam's razor.信息的第二个定义，它适用于单个对象，并把奥卡姆剃刀（Occam's razor）形式化了。 And the fact that forgetting has a price in joules.以及这样一个事实：遗忘是要以焦耳为代价的。

And one habit, which is the same habit this podcast keeps arriving at from different directions.还有一个习惯，也正是这档播客从不同方向反复抵达的同一个习惯。 Ask what was actually measured, and whether it is the same thing as what is being claimed.追问实际被测量的是什么，以及它是否和被声称的东西是同一回事。

Shannon measured selection from a set of possibilities. Everything else people have claimed on his behalf, they claimed on their own.香农测量的是从一组可能性中做出的选择。人们以他的名义声称的其余一切，都是他们自己声称的。

Tomorrow we go back to the usual: one thing a day, unconnected to the last.明天我们回到常态：一天讲一件事，与前一件互不相关。 But if you want to go further with this, the single best book is David MacKay's Information Theory, Inference, and Learning Algorithms, which is free online, and which is unusual in treating information theory and machine learning as one subject rather than two — which, after ten days, I hope now seems obviously correct rather than eccentric.但如果你想在这方面走得更远，最好的一本书是 David MacKay 的《Information Theory, Inference, and Learning Algorithms》，它在网上免费，而且它不同寻常之处在于把信息论和机器学习当作一门学科而非两门来对待——经过这十天，我希望这如今看起来是显而易见地正确，而不是古怪。

For eight days I have treated information as pure mathematics, and that was faithful to how Shannon built it.八天来，我一直把信息当作纯粹的数学来对待，这忠实于香农当初构建它的方式。 There is no physics in the nineteen forty-eight paper.1948 年那篇论文里没有任何物理学。 Entropy there is a property of a probability distribution, and it would mean the same thing in a universe with different laws.那里的熵是一个概率分布的属性，在一个物理定律不同的宇宙里，它也会意味着同样的东西。

But the word was already taken.但这个词已经被占用了。 Entropy was a thermodynamic quantity for eighty years before Shannon, and von Neumann told him to borrow it partly because the mathematics matched.在香农之前，熵作为一个热力学量已经存在了八十年，冯·诺依曼建议他借用它，部分原因就是数学形式相吻合。

Today: is that a coincidence? The answer is no, it is not a coincidence, and the connection has been measured in a laboratory in joules.今天要谈的是：这是巧合吗？答案是否定的，这不是巧合，而且这种联系已经在实验室里以焦耳为单位被测量出来。

But the road there runs through a puzzle that took a hundred and fifteen years to resolve, and the resolution is the good part.但通往那里的道路要穿过一个花了一百一十五年才解决的谜题，而这个谜题的解答才是精彩之处。

Start with thermodynamic entropy. Ludwig Boltzmann, in the eighteen seventies, gave the statistical account.先从热力学熵说起。路德维希·玻尔兹曼在 1870 年代给出了统计学上的解释。

A gas in a box has a macroscopic state — its temperature, pressure, volume.盒子里的气体有一个宏观状态——它的温度、压强、体积。 It also has a microscopic state: the exact position and velocity of every molecule.它还有一个微观状态：每一个分子的确切位置和速度。 Enormously many microstates correspond to the same macrostate.极其大量的微观态对应着同一个宏观态。

Boltzmann's entropy is the logarithm of the number of microstates consistent with the macrostate, times a constant.玻尔兹曼的熵是与该宏观态相容的微观态数目的对数，再乘以一个常数。

Now look at that and compare with day two. The logarithm of the number of equally likely possibilities. That is Hartley's measure.现在看着这个，和第二天讲的内容比较一下。等可能情形数目的对数。那就是哈特利的度量。 That is Shannon's entropy for a uniform distribution. They are the same formula.那就是均匀分布下的香农熵。它们是同一个公式。

Boltzmann had it in the eighteen seventies, applied to molecules; Shannon had it in nineteen forty-eight, applied to messages.玻尔兹曼在 1870 年代把它用在分子上；香农在 1948 年把它用在消息上。

Which suggests a reading: thermodynamic entropy is a measure of your ignorance about the microstate.这提示了一种解读：热力学熵是对你关于微观态的无知程度的度量。 It is how many bits you would need to specify exactly which arrangement of molecules you have, given only the macroscopic description.它是指，在只给定宏观描述的情况下，你需要多少比特才能确切指明你手上的分子处于哪一种排布。

That reading is due mainly to Edwin Jaynes, who in nineteen fifty-seven argued that statistical mechanics is really a form of inference — you assign the distribution that maximises entropy subject to what you know, because that is the least presumptuous choice.这种解读主要归功于埃德温·杰恩斯，他在 1957 年论证说，统计力学其实是一种推断形式——你在已知条件的约束下选取使熵最大化的分布，因为那是最不武断的选择。 On this view, thermodynamics is applied information theory.在这种观点下，热力学是应用信息论。

It is elegant and it is not universally accepted, and I should be straight about that.它很优雅，但并未被普遍接受，这一点我应当说清楚。 The objection is that entropy governs engines and refrigerators and is measured with thermometers, and it seems odd for a quantity central to the behaviour of steam to depend on what an observer happens to know.反对意见是，熵支配着发动机和冰箱，是用温度计测量的，而对于一个在蒸汽行为中起核心作用的量来说，让它依赖于某个观察者恰好知道什么，似乎有些古怪。 That debate is live. What is not in dispute is the mathematical identity, and what is also not in dispute is what follows. Now the puzzle.这场争论仍在进行。没有争议的是那个数学上的恒等式，同样没有争议的是由此得出的结论。现在来看这个谜题。

James Clerk Maxwell, in eighteen sixty-seven, proposed a thought experiment to probe the second law of thermodynamics — the law that says entropy never decreases, that heat flows from hot to cold and not back.詹姆斯·克拉克·麦克斯韦在 1867 年提出了一个思想实验，用来探究热力学第二定律——这条定律说熵永不减少，热量从热流向冷而不会反过来。

Imagine a box of gas divided in two, with a small door.设想一个装有气体的盒子被分成两半，中间有一扇小门。

At the door sits a being, later called Maxwell's demon, who can see individual molecules.门口坐着一个存在物，后来被称为麦克斯韦妖，它能看见单个分子。 When a fast molecule approaches from the right, the demon opens the door and lets it through to the left.当一个快分子从右边靠近时，妖就打开门，让它通过到左边。 When a slow one approaches from the left, it lets that through to the right. The demon does no work. The door is frictionless and massless.当一个慢分子从左边靠近时，它就让那个分子通过到右边。妖不做功。门是无摩擦、无质量的。

It only observes and decides. After a while, fast molecules are on the left and slow ones on the right. Fast means hot.它只是观察并作出决定。过一段时间后，快分子都在左边，慢分子都在右边。快意味着热。

So one side has become hot and the other cold, spontaneously, with no work done.于是一侧变热、另一侧变冷，自发地发生，而没有做任何功。