Chapter 20. When You Forget, I'll Remember第二十章 你忘了,我记得
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 cards, the tea and the old pictures are on the pages. Here are the ideas in words.
On page 2 a box says that the name "artificial intelligence" is 95 years old. On 31 August 1955 John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon wrote a proposal for a two-month summer study at Dartmouth College. It holds the first use of the phrase "artificial intelligence", and it already listed "neuron nets" as a topic, so learning networks were there at the birth. Grandpa's birthday a year later is the story's own.
On page 7 the assistant that wrote the toast could not count its own characters. Chat models do not read letters the way we do. They receive token IDs, each one standing for a chunk of text, and that is why they have been oddly bad at tasks like counting the letters in a word. GPT-2, for example, used a vocabulary of 50,257 tokens. This is one reason, not the whole reason. How a model counts Chinese characters is not something this book tries to say.
On page 8 An asks whether Bean has a brain. Artificial neurons were first described by McCulloch and Pitts in 1943 and were only loosely inspired by brain cells. A brain grows, feels, sleeps and belongs to a living body. For a child, "lots of tiny number-switches that learned patterns" is a truer picture.
On pages 8 and 9 Xiaoyu gives An four plain things to hold on to. A machine like Bean learns from examples that people give it. This is one of the big ideas in the AI4K12 list, and children as young as four can grasp it, especially when they do the teaching themselves, as in PopBots. Bean can be wrong and still sound sure. The researchers who wrote "Why language models hallucinate" point out that most tests give a point for a right answer and nothing for "I don't know", so a model tuned for those tests learns to guess. Bean cannot love An back, while people and pets can. And people check Bean and fix it when it is wrong. Two more belong to families, not to the pages. For anything important or personal, ask a grown-up, since children's trust in a device and in a person changes with age and with the kind of question, and younger children need help deciding whom to ask for what. And keep secrets safe, since children in one study understood little about data privacy.
On pages 2, 8 and 9 Xiaoyu answers only what An asks, in a few words, and lets An laugh. "Does Bean have a brain?" gets a short answer, and so does "What can't Bean do?". This is careful practice drawn from how people talk with young children about hard things. It is not a method proven by studies of AI. Follow the child's lead, say true things plainly, and stop when she has had enough.
On page 9 the box asks whether a machine can be a friend. AI companions can be designed to feel like friends. For young children this is risky. The relationship goes one way, and a company designed the system. The US Federal Trade Commission opened an inquiry into AI chatbots acting as companions on 11 September 2025, and UNICEF's 2025 guidance singles out AI companions used by children as a concern. The research says nothing about machines that speak as someone who has died, so the book makes no claim about them.
One honest sentence for families. The helper is a clever tool that learned from lots of examples people showed it.
A good picture book about AI shows a helper that learns from examples a child can see, makes mistakes that people check, is made by people and is never alive.
On pages 5 and 6 the laptop beats Uncle Wu at chess, and a robot vacuum is stopped by a sock. This is Moravec's paradox from Chapter 1. What grown-ups find hard, like chess, is easy for machines, and what a toddler does without thinking, like pouring tea or crossing a room with socks on it, is very hard (Moravec 1988).
Bridge City and everyone in it are made up, and so is the whole world of 2050 with its garden towers and solar boats. Bean and everything Bean says are made up, and so is Bean rolling round a sock. Grandpa, his notebook, the letter, his village card game, the long-life noodles and the birthday are made up too. No real product is described. The assistant that could not count, the vacuum and the sock, and the arguments about what counts as clever belong to the world as it is now. Every fact in the boxes is real.
Everything in the 2026 chapters is our world as it is now. Everything after it is imagined, a hopeful guess. Thank you for coming along.
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.
牌、茶和那些旧画面,都在书页里了。这里用文字,把背后的道理讲一讲。
第二页的方框说,“人工智能”这个名字今年95岁了。1955年8月31日,约翰·麦卡锡、马文·明斯基、纳撒尼尔·罗切斯特和克劳德·香农写了一份计划书,提议夏天在达特茅斯学院开展为期两个月的研究。“人工智能”这个词,第一次出现就是在这份计划书里。计划书还把“神经网络”列为研究题目,可见会学习的网络从一开始就在。外公晚一年同一天出生,是故事的安排。
第七页,写祝酒词的助手,连自己写了几个字都数不清。聊天模型读文字,跟我们不一样。它收到的是词元的编号,每个编号代表一小段文字,所以数一个单词里有几个字母这类事,它们一向做得出奇地差。比如GPT-2的词表有50,257个词元。这只是一个原因,不是全部。至于模型怎样数汉字,这本书不作推测。
第八页,安问豆豆有没有脑子。人工神经元由麦卡洛克和皮茨在1943年首先提出,只是大致参照了脑细胞。大脑会生长,有感觉,要睡觉,长在活的身体里。对孩子来说,“好多好多小小的数字开关,学会了一些规律”,这个说法更贴近事实。
第八、九页,小雨给安讲了四件实在的事。豆豆这样的机器,是从人给的例子里学的。这是AI4K12提出的核心概念之一,四岁的孩子就能懂,自己动手教的时候尤其容易懂,PopBots课程就是这么做的。豆豆会出错,出错时口气照样很肯定。《语言模型为什么会产生幻觉》的作者指出,多数测试答对得分,答“我不知道”不得分,为这些测试调出来的模型,自然就学会了猜。豆豆没法像安爱它那样爱安,人和小动物会。人还会检查豆豆,出了错就修。另外两条留给家长,书页里没有讲。重要的事、私人的事,要问大人。孩子信不信机器、信不信人,会随年龄和问题的类型而变,年纪小的孩子需要大人帮着判断什么事该问谁。还有,秘密要守好,有一项研究发现,孩子对数据隐私了解很少。
第二、八、九页,安问什么,小雨就答什么,三言两语,还让安笑。“豆豆有脑子吗?”答得很短,“豆豆有什么做不到的?”也一样。这种做法来自大人跟幼儿谈难题的经验,是一种谨慎的做法,并没有人工智能方面的研究证明它有效。跟着孩子的问题走,真话平平实实地说,孩子听够了就停。
第九页的方框问,机器能不能当朋友。人工智能伙伴可以做得像朋友一样。对年幼的孩子来说,这有风险,因为这份关系是单向的,系统又是公司设计的。2025年9月11日,美国联邦贸易委员会对充当伙伴的人工智能聊天机器人展开调查。联合国儿童基金会2025年的指南,也专门把儿童使用人工智能伙伴列为需要关注的问题。用已故者的口吻说话的机器,现有研究没有涉及,所以这本书对它们不下判断。
最后,给家长一句实在话。这个帮手是一件聪明的工具,是从人给它看的大量例子里学出来的。
好的人工智能绘本,会让孩子看见帮手怎样从例子里学习,出了错有人检查,也讲明帮手是人做的,从来不是活的。
第五、六页,笔记本电脑下棋赢了吴伯伯,扫地机器人却被一只袜子卡住。这就是第一章讲过的莫拉维克悖论。大人觉得难的事,比如下棋,机器做起来反倒容易。小娃娃不假思索就会的事,比如倒茶,比如穿过一间扔着袜子的屋子,机器却很难做到(莫拉维克,1988年)。
桥城和城里的人都是编的,2050年的整个世界也是,包括长满花园的高楼和太阳能小船。豆豆和它说的每一句话是编的,豆豆绕着袜子滚过去也是。外公、他的本子、那封信、他村里的牌、长寿面和那个生日,也都是编的。书里没有写任何真实的产品。数不清字数的助手、扫地机器人和袜子,还有关于什么才算聪明的争论,都是今天这个世界里的事。方框里讲的事实,都是真的。
2026年那几章,写的就是我们今天的世界。后面的一切都是想象,是一个满怀希望的猜想。谢谢你一路同行。
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。