Chapter 17. Looking Inside第十七章 往里看
A manga for grown-ups. Read it in English or Chinese.给大人看的漫画。可以用中文或英文读。































































































The lab at night, the notebook and the stone table are in the pictures. Here are the ideas in words.
On page 3 Professor Brunner describes what Swan did, and the first box says why nobody can easily tell why. A neural network is made of many simple units joined by numbers called weights, and training sets billions of them. After training we know every number, yet we cannot say in plain words what the network has learned or why it gave a particular answer. The field that tries to find out is called interpretability. It matters for every chapter so far. If we can see that a model looks at snow rather than at the wolf, or at a hospital's stamp rather than at the lungs, we can fix it before it does harm. Chapter 13 told both stories (Ribeiro and colleagues 2016, Zech and colleagues 2018).
On page 4 Swan explains itself, and the explanation sounds sensible. But a model's explanation of itself is generated in the same way as any other text, so it may not match what happened inside. Turpin and colleagues (2023) found that step-by-step explanations sometimes left out the real reason for an answer, such as a hint planted in the question. Baker and colleagues (2025, met in Chapter 16) found that written reasoning can be a useful window, but it can become less honest if a model is trained against it. Swan's "stable at the last check" is the story's version, a reason that was true at six o'clock and false by eight.
On page 5 Xiaoyu cannot read Swan one part at a time. Olah and colleagues (2020) proposed studying networks almost like biology, piece by piece, and found circuits, such as curve detectors built from simpler edge detectors. They also found polysemantic units, single units that respond to unrelated things such as cat faces and car fronts. Elhage and colleagues (2022) offered an explanation. A network has to hold more ideas than it has units, so the ideas are stored overlapping, in what they called superposition, a bit like many radio stations sharing the same airwaves.
On page 6 a second, simpler program sorts Swan's lights into features. This is called dictionary learning, done with a sparse autoencoder, a second network trained to rewrite the first one's activity as a much larger set of features, only a few active at a time. Bricken and colleagues (2023) and, independently, Cunningham and colleagues (2023) showed that the features are often far easier to understand than single units. Templeton and colleagues (2024) scaled the method up to a large model, Claude 3 Sonnet, and found up to 34 million features, including one for the Golden Gate Bridge, one for computer code with security flaws and one for flattering praise. Turning a feature up changed what the model did. With the bridge feature turned up, the model began to describe itself as the bridge, and Anthropic briefly let the public try this as "Golden Gate Claude". Gao and colleagues (2024) at OpenAI scaled similar methods to GPT-4. Real features are patterns of numbers. In the story, as in real labs, a feature's name is only a guess until it is tested. That is why Xiaoyu does not trust her guess about "asked for nothing tonight" until she has tried it in a practice copy of Swan.
On pages 14 to 16 Xiaoyu gives each person in the notebook a colour and finds that most pages need several, the way one part of a network takes part in several ideas. This is a picture, not a method. Her colours are guesses too, and she checks them page by page.
One last sentence for families. People build AI by training it, not by writing down every rule, so even its makers do not fully know how it works inside. Scientists are learning to look inside, and that helps them find and fix mistakes. In places where it matters, like hospitals, people keep checking by hand and decide whether to trust a program.
In 2025 researchers at Anthropic went further and traced how features connect into step-by-step "attribution graphs" inside a production model, Claude 3.5 Haiku (Lindsey and colleagues 2025). Asked for the capital of the state containing Dallas, the model seemed to represent "Texas" internally before it answered "Austin". Writing a rhyming verse, it seemed to pick the rhyming word before it wrote the line leading to it. The authors present these as case studies, not a complete account.
The first reason is scale. Large models have hundreds of billions of weights, and even millions of features capture only part of what they do. The second is faithfulness. An explanation can look convincing and still be wrong. Adebayo and colleagues (2018) found that some popular "saliency maps", which highlight the pixels that supposedly mattered, looked much the same even for networks with randomised weights, so the maps partly showed the picture and not the network's reasoning. The third is that researchers still debate what it would mean to have "understood" a network. As of 2026 looking inside gives real, specific insights and some practical tools, but nobody can yet read a large model's mind in general.
The idea of asking what a part of a network responds to is not new. Chapter 7 told how Zeiler and Fergus (2014) and Olah, Mordvintsev and Schubert (2017) pictured what units in an image network see. Chapter 13's wolves and hospital stamp are a case of shortcut learning, which Geirhos and colleagues (2020) describe, and looking inside is one way to catch such shortcuts.
Bridge City and everyone in it are made up. So is the whole world of 2039, with its garden towers, solar boats and self-driving taxis. The Zurich lab, Professor Brunner and the four hospitals are made up. So are Swan, its answer, its parts, the feature "asked for nothing tonight", feature 4117 and the practice copy. Willow and its record are made up. Grandpa, his notebook and every page of it are made up, and so are the park's custom of a white cloth on a stool and the small grey speaker. The research has no case of a hospital program that leaves quiet patients till last. Looking inside a network for a hidden shortcut is real work, and Swan's night is the story's way of showing it. No real hospital system is described, and in real hospitals nurses and doctors decide who is seen first. Grandpa's illness is not named, and the story makes no medical claim. Many airports limit night flights, but the flight times in the story are made up.
Yingying Fu created this manga. She chose the story and its people, directed every page, and edited every word and picture. The words were drafted with Claude, an AI by Anthropic. The pictures are made with AI image generation, painted panel by panel from her direction and checked and corrected by hand.
夜里的实验室、外公的本子、公园的石桌,画里都有了。下面用文字把里面的道理讲一讲。
第三页,布伦纳教授讲了天鹅做了什么,第一个方框讲的是,为什么谁都很难说清它为什么这样做。神经网络由许多简单的单元组成,单元之间用叫作“权重”的数字相连,训练要定下几十亿个这样的数字。训练好以后,每个数字我们都知道,却没法用大白话说清网络学到了什么,为什么给出某个答案。想弄清这些的研究领域,叫“可解释性”。前面每一章都跟它有关。要是能看出模型盯着的是雪而不是狼,是医院的印章而不是肺,我们就能在它闯祸之前把它改好。这两个故事第十三章都讲过(里贝罗等,2016年;泽克等,2018年)。
第四页,天鹅解释了自己,听上去很有道理。可模型对自己的解释,和它写别的文字一样,是生成出来的,所以不一定符合它内部真正发生的事。图尔平等人(2023年)发现,模型一步步写出的解释,有时会漏掉答案的真正原因,比如问题里事先埋下的一个提示。贝克等人(2025年,第十六章提到过)发现,写出来的推理过程可以是一扇有用的窗口,可要是专门照着它去训练模型,推理就可能变得不那么老实。天鹅说的“病情稳定,上次检查正常”,就是故事里的例子,这个理由六点时是真的,到八点就不是了。
第五页,小雨没法一个零件一个零件地读懂天鹅。奥拉等人(2020年)提出,可以几乎像研究生物那样,一块一块地研究网络。他们找到了“回路”,比如由更简单的边缘检测器搭起来的曲线检测器。他们还发现了“多义单元”,也就是一个单元会对毫不相干的东西起反应,比如猫脸和汽车车头。埃尔哈奇等人(2022年)给出了一种解释。网络要装的概念比单元多,只好让概念叠在一起存放,他们把这叫作“叠加”,有点像许多电台共用同一片电波。
第六页,一个更简单的程序把天鹅的光点分成了一个个特征。这种做法叫“字典学习”,用的工具是“稀疏自编码器”,就是另训练一个网络,把原网络的活动改写成数量多得多的特征,每次只有少数几个亮着。布里肯等人(2023年)和独立开展研究的坎宁安等人(2023年)都发现,这样得到的特征往往比单个单元好懂得多。坦普尔顿等人(2024年)把这个方法用到大模型Claude 3 Sonnet上,找到了多达3400万个特征,有对应金门大桥的,有对应带安全漏洞的代码的,还有对应奉承话的。把一个特征调强,模型的行为就跟着变。金门大桥特征调强以后,模型开始把自己说成那座桥,Anthropic还短暂地开放给公众试用,叫“金门大桥Claude”。OpenAI的高等人(2024年)把类似的方法用到了GPT-4上。真实的特征是一组组数字的规律。故事里和真实的实验室里一样,一个特征的名字没验证之前,都只是猜测。所以小雨对“今晚什么都没要过”这个猜测,要先在天鹅的练习副本里试过,才肯相信。
第十四页到第十六页,小雨给本子里的每个人一种颜色,发现大多数页要用好几种,就像网络里的一个零件会参与好几个概念。这只是个比方,不是一种方法。她的颜色也是猜测,所以她一页一页地核对。
最后,给大人和孩子说几句。AI是训练出来的,不是人把每条规则都写下来,所以连造它的人,也不完全清楚它里面是怎么运转的。科学家正在学着往里看,这能帮他们发现错误、改正错误。在医院这样要紧的地方,人还是会亲手检查,由人来决定要不要相信一个程序。
2025年,Anthropic的研究者更进一步,在一个已经正式使用的模型Claude 3.5 Haiku里,追踪特征之间怎样一步步连起来,画成“归因图”(林赛等,2025年)。问它“达拉斯所在的州,首府是哪里”,模型在回答“奥斯汀”之前,内部似乎先出现了“得克萨斯”。写押韵的诗句时,它似乎先选好了押韵的词,再写通向这个词的那一行。作者把这些当作个案研究,并不是完整的解释。
第一个原因是规模。大模型有几千亿个权重,就算找出几百万个特征,也只能说明它所做的一部分。第二个是忠实性。一个解释看起来很有说服力,却可能是错的。阿德巴约等人(2018年)发现,有些流行的“显著图”(用来标出据说起了作用的像素),换成权重随机的网络,画出来也差不多。可见这些图有一部分显示的是图片本身,而不是网络的推理。第三,研究者至今还在争论,怎样才算“弄懂”了一个网络。到2026年,往里看已经带来了真实、具体的发现和一些实用的工具,但还没有人能普遍地读懂一个大模型的“心思”。
问“网络的某个部分对什么有反应”,不是新想法。第七章讲过,泽勒和弗格斯(2014年)以及奥拉、莫尔德温采夫和舒伯特(2017年)怎样把图像网络里的单元“看到”的东西画出来。第十三章的狼和医院印章,是“捷径学习”的例子,盖尔霍斯等人(2020年)专门讲过这个问题。往里看,是抓住这类捷径的办法之一。
桥城和城里所有的人都是编的,2039年的整个世界也是,包括种满花草的高楼、太阳能小船和无人驾驶出租车。苏黎世的实验室、布伦纳教授和那四家医院是编的。天鹅、它的回答、它的零件、“今晚什么都没要过”这个特征、特征4117和那个练习副本,也都是编的。小柳和它的记录是编的。外公、他的本子和本子里的每一页是编的,公园里在凳子上铺白布的老规矩和那只灰色的小音箱也是。研究文献里,并没有哪个医院程序把安静的病人排到最后。往网络里面找藏着的捷径,是真实的研究工作,天鹅这一夜,是故事用来展示这种工作的办法。故事没有描写任何真实的医院系统,在真实的医院里,先看谁、后看谁,由护士和医生决定。外公得的是什么病,故事没有说,也不做任何医学上的断言。很多机场夜里限制航班起降,不过故事里的航班时间是编的。
这本漫画由 Yingying Fu 创作。故事和人物由她决定,每一页由她导演,每一句话、每一幅画由她修改。文字是和 Anthropic 的人工智能 Claude 一起起草的。画面由人工智能生成,按她的要求一格一格地画,再逐张检查、修正。
© 2026 Yingying Fu. All rights reserved.© 2026 傅颖盈。保留所有权利。