保持提问，
也记录修正。

1. 1940s图灵与恩尼格玛Turing and Enigma二战中，图灵在布莱切利园设计机器，破译德军的恩尼格玛密电。1950 年，他写下一个问题：机器会思考吗？During the Second World War, Alan Turing designed machines at Bletchley Park to break the German Enigma cipher. In 1950 he wrote down a question: can machines think?Turing 1950 · Computing Machinery and IntelligenceBletchley Park
2. 1956它有了名字The field gets a name达特茅斯的一个夏天，这个领域被命名为“人工智能”。随后，感知机和会下跳棋的程序追问：机器能从经验中学习吗？One summer at Dartmouth, the field was named “artificial intelligence”. Soon the perceptron and a checkers-playing program asked the next question: can a machine learn from experience?The Dartmouth proposal (1955)Rosenblatt 1958 · The PerceptronSamuel 1959 · Checkers
3. 1956–1974第一波浪潮The first wave会聊天的 ELIZA，会看路的机器人 Shakey……乐观的人相信，二十年内机器就能做人能做的一切。ELIZA could chat; Shakey the robot could find its way. Optimists believed that within twenty years machines would do any work a person can do.Weizenbaum 1966 · ELIZAShakey the robot
4. 1974–1990s寒冬里的灯A lamp in the winter承诺落空，经费断流，AI 两度入冬。仍有人守着那盏灯：反向传播让多层网络学会了学习，LeNet 读懂了手写数字。Promises went unkept and funding dried up: AI went through two winters. A few kept the lamp lit — backpropagation taught multi-layer networks to learn, and LeNet learned to read handwritten digits.AI winterRumelhart, Hinton & Williams 1986 · BackpropLeCun et al. 1989
5. 1997深蓝Deep BlueIBM 的深蓝以 3½ 比 2½ 击败国际象棋世界冠军卡斯帕罗夫，AI 重新回到聚光灯下。IBM’s Deep Blue beat world chess champion Garry Kasparov 3½ – 2½, and AI was back in the spotlight.Deep Blue vs. Kasparov
6. 2006–2015爆发前夜The eve of the boom为游戏而生的显卡成了神经网络的引擎；ImageNet 带来一千四百万张标注图片。2012 年，AlexNet 用两块显卡把识别错误率一口气降了近十个百分点；2015 年，152 层的 ResNet 把它降到了 3.6%。Graphics cards built for games became the engine of neural networks, and ImageNet supplied fourteen million labelled images. In 2012 AlexNet, trained on two GPUs, cut the error rate by almost ten points; in 2015 the 152-layer ResNet brought it down to 3.6%.ImageNetKrizhevsky et al. 2012 · AlexNetHe et al. 2015 · ResNet
7. 2015一家非营利机构A non-profit2015 年底，OpenAI 在旧金山成立。它最初是一家非营利研究机构，创始人们承诺捐出 10 亿美元，目标是让通用人工智能造福全人类。（2019 年起，它转向“有限盈利”的结构。）At the end of 2015, OpenAI was founded in San Francisco as a non-profit research company. Its founders pledged $1 billion, with the goal of making artificial general intelligence benefit all of humanity. (In 2019 it moved to a “capped-profit” structure.)Introducing OpenAI (2015)
8. 2016第 37 手Move 37AlphaGo 以 4 比 1 战胜李世石。第二局第 37 手，一步人类棋手万分之一才会下的棋，让人们重新想象围棋。AlphaGo beat Lee Sedol 4 – 1. Move 37 of the second game — a move a human player would choose about one time in ten thousand — made people imagine Go anew.Silver et al. 2016 · NatureAlphaGo vs. Lee Sedol
9. 机器学习 → 深度学习一层一层地学Learning, layer by layer机器学习让机器从数据中找规律；深度学习把神经网络叠得更深——前馈网络逐层传递，卷积网络看懂图像，Transformer 让每个词都能看见整句话。Machine learning lets machines find patterns in data; deep learning stacks neural networks ever deeper. Feed-forward networks pass signals layer by layer, convolutional networks see images, and the Transformer lets every word see the whole sentence.LeCun, Bengio & Hinton 2015 · Deep learning
10. 2017注意力就是你所需要的Attention is all you need谷歌的论文《Attention Is All You Need》提出 Transformer：不再逐字阅读，而是让每个词同时“注意”所有词。“它”指的是谁？注意力会告诉你。Google’s paper “Attention Is All You Need” introduced the Transformer: instead of reading word by word, every word attends to every other word at once. What does “it” refer to? Attention will tell you.Vaswani et al. 2017
11. 2020规模的魔法The magic of scale“规模定律”发现：模型越大、数据越多、算力越足，表现就越好，而且可以预测。GPT-3 有 1750 亿个参数，只看几个例子就能学会新任务。两年后 ChatGPT 上线，许多人第一次和 AI 聊起天来。Scaling laws showed that bigger models, more data and more compute make models better — predictably. GPT-3 had 175 billion parameters and could pick up a new task from just a few examples. Two years later ChatGPT launched, and millions chatted with an AI for the first time.Kaplan et al. 2020 · Scaling lawsBrown et al. 2020 · GPT-3Introducing ChatGPT
12. 2020 → 2024蛋白质与诺贝尔奖Proteins and the Nobel Prize2020 年，AlphaFold 2 几乎解决了困扰生物学五十年的蛋白质折叠问题，随后预测了两亿多种蛋白质的结构。2024 年，哈萨比斯和江珀获得诺贝尔化学奖；同一年，霍普菲尔德和辛顿获得诺贝尔物理学奖——当年守着寒冬那盏灯的人。In 2020 AlphaFold 2 all but solved protein folding, a fifty-year-old problem in biology, and went on to predict the structures of over 200 million proteins. In 2024 Hassabis and Jumper shared the Nobel Prize in Chemistry; the same year Hopfield and Hinton won the Nobel Prize in Physics — Hinton, who kept the lamp lit through the winter.Jumper et al. 2021 · AlphaFold 2AlphaFold Protein Structure DatabaseNobel Prize in Chemistry 2024Nobel Prize in Physics 2024
13. 2014 → 2022AI 学会画画AI learns to draw2014 年的 GAN 让两个网络互相较劲：一个造假，一个识破。后来的扩散模型换了思路——从一团噪声出发，一步步去噪，直到画面浮现。2022 年，DALL·E 2、Midjourney 和 Stable Diffusion 让任何人都能“用文字作画”。In 2014 GANs pitted two networks against each other: one forges, the other spots the fake. Diffusion models took another path — start from pure noise and remove it step by step until a picture appears. In 2022 DALL·E 2, Midjourney and Stable Diffusion let anyone paint with words.Goodfellow et al. 2014 · GANHo et al. 2020 · Diffusion (DDPM)Rombach et al. 2022 · Stable Diffusion
14. 2022涌现Emergence一只椋鸟只看身边几只邻居，成千上万只却能画出流动的形状。模型也是这样：规模跨过某个门槛，一些从没专门教过的能力会突然出现——这被称为“涌现”。（它是真实的跃变，还是评测方式造成的错觉？学界仍在争论。）Each starling watches only a few neighbours, yet thousands together draw a flowing shape. Models seem to do the same: past some threshold of scale, abilities nobody taught them suddenly appear — “emergence”. (Real leap, or an artifact of how we measure? Researchers still argue.)Wei et al. 2022 · Emergent abilitiesSchaeffer et al. 2023 · A mirage?Anderson 1972 · More is differentBallerini et al. 2008 · Starlings
15. 2023Claude让模型好好说话，靠的是 RLHF——人类反馈强化学习：人来比较哪个回答更好，模型照着改。Anthropic 发布 Claude，又加上了一部“宪法”，力求有益、诚实、无害——不只要更聪明，也要更值得信任。Models learned to be good conversational partners through RLHF — reinforcement learning from human feedback: people compare answers, and the model learns from their choices. Anthropic released Claude and added a written “constitution”, aiming for helpful, honest and harmless — not just smarter, but more trustworthy.Ouyang et al. 2022 · InstructGPT (RLHF)Bai et al. 2022 · Constitutional AIIntroducing Claude
16. 2025先想，再答Think first, then answerDeepSeek-R1 开源了推理模型：靠强化学习学会写下长长的思维链——会验算，会回头，也会说“等等”。画面里是网友为它创作的“鲸鱼娘”：爱吃白饭，也爱烧 token。DeepSeek-R1 released an open-weight reasoning model that learned, through reinforcement learning, to write long chains of thought — checking its work, backtracking, even saying “wait”. In the picture: the “whale girl” fans drew for it, who loves plain rice and burning tokens.DeepSeek-AI 2025 · DeepSeek-R1
17. 多模态看、听、说、画See, hear, speak, drawAI 不再只读文字：它看图，听声音，开口说话，提笔作画，把世界的多种语言连在一起。AI no longer only reads text: it looks at pictures, listens, speaks and paints, joining up the many languages of the world.Radford et al. 2021 · CLIPHello GPT-4o
18. 2025–2026百模齐发A hundred models bloom2025 年起，新模型几乎每周都在发布：开源的、闭源的，大的、小的，全世界加起来已有上百款。把它们放进 Artificial Analysis 排行榜的“智能 × 成本”坐标里，最靠左上的那条线，就是帕累托前沿。From 2025 on, new models arrived almost every week — open and closed, large and small — hundreds of them worldwide. Place them on the Artificial Analysis leaderboard’s intelligence-versus-cost chart, and the line along the upper left is the Pareto frontier.Artificial Analysis
19. 2025–2026从聊天到动手From chatting to doingAI 开始自己动手：打开终端、读代码、改代码、跑测试，一连工作几个小时。它们能独立完成的任务长度，大约每七个月翻一番。AI began to act on its own: opening a terminal, reading and editing code, running the tests, working for hours at a stretch. The length of tasks agents can finish on their own has been doubling roughly every seven months.METR 2025 · Measuring long tasksSWE-benchClaude Code
20. 2026今天TodayClaude Fable 5 与 GPT-6 相继登场，Grok、DeepSeek、Qwen、Gemini、Kimi、GLM 同场竞速；AI 智能体为千禧年难题之一——纳维–斯托克斯问题——给出了候选反例，正等待数学界检验。Claude Fable 5 and GPT-6 arrived, with Grok, DeepSeek, Qwen, Gemini, Kimi and GLM racing alongside. A team of AI agents proposed a counterexample for one of the Millennium Prize Problems — Navier–Stokes — now awaiting the verdict of mathematicians.Quanta · AI and a Millennium Prize ProblemClay Institute · Navier–StokesGPT-6 Astra
21. 1950 → ∞未完待续To be continued从 1950 年的那个问题出发，我们走了七十多年。机器会思考吗？——我们仍在回答。More than seventy years on from that question of 1950. Can machines think? We are still answering.

配乐：约翰·帕赫贝尔《D 大调卡农》（c. 1680），在浏览器中实时合成与改编。 · Music: Pachelbel’s Canon in D (c. 1680), arranged and synthesized live in the browser.