Artificial intelligence · 人工智能
| English | 中文 | Pinyin · 拼音 |
|---|---|---|
| artificial intelligence/ˌɑːtɪˈfɪʃl ɪnˈtelɪdʒəns/ | 人工智能 | rén gōng zhì néng |
| machine learning/məˈʃiːn ˈlɜːnɪŋ/ | 机器学习 | jī qì xué xí |
| training data/ˈtreɪnɪŋ ˈdeɪtə/ | 训练数据 | xùn liàn shù jù |
| deep learning/diːp ˈlɜːnɪŋ/ | 深度学习 | shēn dù xué xí |
| neural networks/ˈnjuːrəl ˈnetwɜːks/ | 神经网络 | shén jīng wǎng luò |
| speech recognition/spiːtʃ ˌrekəɡˈnɪʃn/ | 语音识别 | yǔ yīn shí bié |
| image recognition/ˈɪmɪdʒ ˌrekəɡˈnɪʃn/ | 图像识别 | tú xiàng shí bié |
| machine translation/məˈʃiːn trænˈsleɪʃn/ | 机器翻译 | jī qì fān yì |
| recommendation systems/ˌrekəmenˈdeɪʃn ˈsɪstəmz/ | 推荐系统 | tuī jiàn xì tǒng |
| autonomous vehicles/ɔːˈtɒnəməs ˈvɪəklz/ | 自动驾驶汽车 | zì dòng jià shǐ qì chē |
| optical character recognition/ˈɒptɪkl ˈkærɪktə ˌrekəɡˈnɪʃn/ | 光学字符识别 | guāng xué zì fú shí bié |
| text-to-speech/tekst tə spiːtʃ/ | 文本转语音 | wén běn zhuǎn yǔ yīn |
| bias/ˈbaɪəs/ | 偏见 | piān jiàn |
| accountability/əˌkaʊntəˈbɪlɪti/ | 问责 | wèn zé |
The move no human would have played
- In March 2016 a program called AlphaGo beat Lee Sedol, one of the strongest Go players alive, four games to one. In the second game it played a move that commentators called a mistake; it turned out to be the winning idea.
- Nobody had programmed that move. The system had learned Go by studying millions of positions and playing itself, finding patterns its makers could not have written down.
- Two years later Amazon scrapped an AI recruiting tool after finding it had learned from ten years of CVs to mark down any that mentioned "women's".
- Both are artificial intelligence 人工智能. This lesson is how it learns, what it is used for, and its social, economic and environmental impact.
没有人类会走的那一步
- 2016 年 3 月,一个叫 AlphaGo 的程序以四比一击败了当世最强的围棋手之一李世石。第二局它走了一步解说员称为失误的棋;结果那正是获胜的构想。
- 没有人给它编过那一步。这个系统通过研究数百万个棋局并与自己对弈学会了围棋,发现了制造者写不出来的规律。
- 两年后,亚马逊废弃了一个 AI 招聘工具,因为发现它从十年的简历中学到:凡提到"女子"的一律降分。
- 两者都是人工智能(artificial intelligence)。这一课讲它怎样学习、用来做什么,以及它的社会、经济和环境影响。
What AI is and how modern AI learns
- Artificial intelligence (AI) means computer systems that perform tasks which normally need human intelligence: recognising speech and images, translating, playing games, driving.
- Most modern AI uses machine learning 机器学习: instead of being programmed step by step, the system improves at a task by learning patterns from large amounts of training data 训练数据.
- Deep learning 深度学习 uses neural networks 神经网络 with many layers of connected nodes, and is the leading approach today. Better and larger data gives a better model; bad data gives bad results.
A model is trained on examples, not written as rules
什么是 AI,现代 AI 怎样学习
- 人工智能(AI)指执行通常需要人类智能的任务的计算机系统:识别语音和图像、翻译、下棋、驾驶。
- 大多数现代 AI 使用机器学习(machine learning):系统不是一步一步被编程,而是通过从大量训练数据(training data)中学习规律来改进任务表现。
- 深度学习(deep learning)使用多层相连节点的神经网络(neural networks),是今天的主流方法。数据更好更多,模型就更好;数据差,结果就差。
模型是用样例训练出来的,不是写成规则的
Machine learning differs from traditional programming because it: · 机器学习与传统编程的区别在于它：
ML algorithms improve at a task by learning from large amounts of data, rather than following hand-written rules. · ML算法通过学习大量数据来提升任务表现，而不是遵循手写规则。
Match each AI term to what it means. · 将每个AI术语与其含义匹配。
ML learns from data; deep learning uses many layers; bias and opacity (black box) are key ethical concerns. · ML从数据中学习；深度学习使用多层；偏见和不可解释性（黑盒）是关键伦理问题。
Applications
- Understanding input: speech recognition 语音识别 turns spoken words into text; image recognition 图像识别 finds objects, faces or text in pictures.
- Producing output or decisions: machine translation 机器翻译 between languages; recommendation systems 推荐系统 suggesting videos, products or music; autonomous vehicles 自动驾驶汽车 and robots; medical diagnosis and fraud detection.
- The exam usually gives a scenario and asks which kinds of AI it uses, or how it works.
应用
- 理解输入:语音识别(speech recognition)把说出的话变成文字;图像识别(image recognition)在图片中找到物体、人脸或文字。
- 产生输出或决策:语言之间的机器翻译(machine translation);推荐视频、商品或音乐的推荐系统(recommendation systems);自动驾驶汽车(autonomous vehicles)和机器人;医学诊断和欺诈检测。
- 考试通常给一个情境,问它用了哪些 AI,或者它怎样工作。
Computing concept lab · 计算概念实验
Classify concrete examples by the computing idea they demonstrate. · 根据它们所演示的计算概念对具体示例进行分类。
Which is an everyday application of AI? · 哪项是AI的日常应用？
Recommendation systems (and speech/image recognition, translation, self-driving) are common AI applications. · 推荐系统（以及语音/图像识别、翻译、自动驾驶）是常见的AI应用。
Worked example: reading a foreign label aloud
- A phone app photographs a product label, translates it and reads it aloud. Explain how AI is used. [4]
- Optical character recognition 光学字符识别 analyses the pixels of the photograph to locate the characters, and the patterns of pixels are converted into individual characters and words.
- Machine translation converts the words into the user's language.
- Text-to-speech 文本转语音 produces the spoken output. Each stage is a mark, in pipeline order.
Recognise, translate, speak
例题:朗读外文标签
- 一个手机应用拍下商品标签,翻译并朗读它。解释其中怎样使用 AI。[4]
- 光学字符识别(optical character recognition)分析照片的像素以定位字符,像素图案被转换成单个字符和单词。
- 机器翻译把单词转换成用户的语言。
- 文本转语音(text-to-speech)生成语音输出。每个阶段一分,按流水线顺序。
识别、翻译、朗读
Put the stages of the label-reading app in order. · 将标签读取应用的阶段排序。
Recognise, read, translate, speak. Each stage is a mark in a four-mark answer. · 识别、读取、翻译、朗读。每个阶段在四分题中各占一分。
Benefits
- Accessibility: speech and image AI helps users with impairments; translation helps people who do not speak the language.
- Productivity: repetitive tasks are automated, freeing people for creative work.
- Decision support: AI finds patterns in datasets too large for a person, in medical scans or in transactions.
- It is available at any hour and can be personalised to each user.
好处
- 无障碍:语音和图像 AI 帮助有障碍的用户;翻译帮助不懂这种语言的人。
- 生产力:重复任务被自动化,把人解放出来做创造性工作。
- 决策支持:AI 在人无法处理的大数据集中——医学扫描、交易记录——找出规律。
- 它随时可用,并且可以为每个用户个性化。
A benefit of AI is: · AI的一个好处是：
AI excels at finding patterns in large data (decision support). Fairness and explainability are concerns, not guarantees. · AI擅长在大数据中发现模式（决策支持）。公平性和可解释性是担忧，而非保证。
Social impact
- Benefits: a label reader helps people with a visual impairment, people who cannot read the language, and people with reading difficulties; facial recognition at an airport speeds identity checks and can stop wanted people entering.
- Harms: facial recognition misidentifies some people and tracks everyone without consent, so privacy is lost; people who rely on AI may lose skills; recommendation systems can trap users in narrow views.
- A social-impact answer names the group affected and what changes for them.
社会影响
- 好处:标签朗读器帮助视力障碍者、不识这种语言的人和有阅读困难的人;机场的人脸识别加快身份核查,并能阻止被通缉者入境。
- 危害:人脸识别会误认一些人,并在未经同意的情况下追踪所有人,隐私因此丧失;依赖 AI 的人可能失去技能;推荐系统可能把用户困在狭窄的视野里。
- 社会影响的答案要说出受影响的群体和对他们的改变。
Economic impact
- An AI fault-diagnosis module in a repair garage diagnoses faults faster and more accurately, so more vehicles are repaired a day and costs fall.
- But fewer skilled mechanics may be needed, so jobs are lost, and the module must be bought, trained and maintained.
- More generally, AI raises productivity and creates new jobs in some fields while removing routine jobs in others.
经济影响
- 修理厂里的 AI 故障诊断模块更快更准地诊断故障,所以每天修好的车更多,成本下降。
- 但需要的熟练技工可能减少,所以有人失业;而且模块必须购买、训练和维护。
- 更普遍地说,AI 提高生产力并在某些领域创造新工作,同时在其他领域取消常规工作。
Environmental impact
- Training and running large models uses a great deal of electricity, and data centres use water for cooling; the hardware they run on becomes electronic waste.
- On the other side, AI cuts energy use in buildings, optimises transport routes and monitors the environment.
- Both directions score, and the exam expects you to know they exist.
The hardware behind AI has a lifetime, and an afterlife
环境影响
- 训练和运行大模型消耗大量电力,数据中心用水冷却;运行它们的硬件最终成为电子垃圾。
- 另一方面,AI 减少建筑的能耗、优化交通路线并监测环境。
- 两个方向都得分,考试期望你知道两者都存在。
AI 背后的硬件有寿命,也有"身后事"
Match each example to the heading it belongs under. · 将每个示例与其所属标题匹配。
Jobs and costs are economic, people and privacy are social, energy and waste are environmental. · 就业和成本是经济方面的，人和隐私是社会方面的，能源和废物是环境方面的。
Training and running large AI models has no significant environmental cost. · 训练和运行大型AI模型没有显著的环境成本。
Large models consume a great deal of electricity and cooling water, and their hardware becomes e-waste. AI also helps cut energy use elsewhere, so both directions belong in an answer. · 大型模型消耗大量电力和冷却水，其硬件也会变成电子垃圾。AI也有助于在其他地方减少能源使用，因此两个方向都应包含在答案中。
Worked example: answering an "impact" question
- Explain the economic impact of a garage installing an AI fault-diagnosis system. [4]
- Give the impact and its consequence, and give both directions.
- Faults are diagnosed faster and more accurately, so more vehicles are repaired per day and revenue rises. Fewer mechanics' hours are needed per car, so costs fall.
- Skilled mechanics may lose their jobs, so there is unemployment and a loss of skills; and the system must be bought and maintained, so there is an initial and ongoing cost.
例题:回答一道"影响"题
- 解释修理厂安装 AI 故障诊断系统的经济影响。[4]
- 给出影响和它的后果,并给出两个方向。
- 故障诊断更快更准,所以每天修好的车更多,收入上升。每辆车需要的技工工时更少,所以成本下降。
- 熟练技工可能失业,所以有失业和技能流失;系统必须购买和维护,所以有初始和持续的成本。
A full-mark impact point states the impact and its , usually joined by "so". · 满分的影响点陈述了影响及其，通常用“所以”连接。
"Fewer mechanics are needed, so skilled workers lose their jobs" is a mark; "fewer mechanics" alone is not. · "需要更少技工，所以熟练工人失去工作"是一个得分点；仅"更少技工"不算。
Concerns
- Bias 偏见: unfair patterns in the training data become unfair decisions, as in Amazon's recruiting tool. Job displacement: routine roles disappear.
- Privacy: training uses large amounts of personal data. Transparency: large models are black boxes whose decisions are hard to explain.
- Accountability 问责: when the AI is wrong, is the developer, the operator or the user responsible? Misuse: deepfakes, misinformation, surveillance.
- Professionals must understand the limits of what they build, inform users, and reduce harm.
Bias goes in with the data and comes out in the decisions
担忧
- 偏见(bias):训练数据中不公平的模式变成不公平的决定,就像亚马逊的招聘工具。岗位替代:常规岗位消失。
- 隐私:训练使用大量个人数据。透明度:大模型是黑箱,决定难以解释。
- 问责(accountability):AI 出错时,是开发者、运营者还是用户负责?滥用:深度伪造、虚假信息、监控。
- 专业人员必须了解自己所建系统的局限,告知用户,并减少伤害。
偏见随数据进入,随决定出来
AI "bias" usually arises because: · AI“偏见”通常是因为：
If the training data reflects unfair patterns, the model learns and repeats them — e.g. in hiring or lending. · 如果训练数据反映了不公平的模式，模型就会学习并重复这些模式——例如在招聘或贷款中。
Select all · 所有 genuine concerns about AI. · 选择关于 AI 的所有真实担忧。
Bias, opacity and misuse are real concerns. AI certainly can be wrong — accountability for that is itself a concern. · 偏见、不透明性和滥用是真实的担忧。AI 当然可能会出错——对此的责任归属本身就是一个担忧。
Marks that slip away
- AI is not "a robot" or "a computer that thinks". Define it as systems performing tasks that normally need human intelligence.
- An impact without a consequence is half a mark. "Jobs are lost" needs "so skilled workers are unemployed".
- Social, economic and environmental are three headings. Answer under the one the question names.
- Bias comes from the data the model learned from, not from the computer having opinions.
容易丢掉的分
- AI 不是"一个机器人"或"会思考的计算机"。要定义为执行通常需要人类智能的任务的系统。
- 没有后果的影响只算半分。"失业"需要"所以熟练工人失去工作"。
- 社会、经济和环境是三个标题。在题目点名的那个标题下回答。
- 偏见来自模型学习的数据,不是计算机有了自己的观点。
You've got it
- AI performs tasks that normally need human intelligence; machine learning learns patterns from training data rather than following written rules; deep learning uses layered neural networks
- applications: speech and image recognition understand input; translation, recommendation and autonomous vehicles produce output; OCR → translation → text-to-speech reads a label aloud
- impact under three headings, each with a consequence and both directions: social, economic, environmental
- concerns: bias from data, job displacement, privacy, transparency, accountability, misuse
你掌握了
- AI 执行通常需要人类智能的任务;机器学习从训练数据中学习规律而不是遵循写好的规则;深度学习使用分层的神经网络
- 应用:语音和图像识别理解输入;翻译、推荐和自动驾驶产生输出;OCR → 翻译 → 文本转语音朗读标签
- 影响分三个标题,各带后果和两个方向:社会、经济、环境
- 担忧:来自数据的偏见、岗位替代、隐私、透明度、问责、滥用