Artificial Intelligence人工智慧
AI 不是被寫出規則的，是被餵了大量例子之後自己找出規律的。這也是為什麼它會出錯得那麼有自信。
CEFR B2閱讀約 6 分鐘 · 12 個單字 · 測驗卷 100 分
3D Tour3D 導覽影片：先看一遍，再自己點點看
3D 導覽影片 · 約 2 分鐘
The Robot Training Lab機器人訓練實驗室
12 個可以點的單字 · 可以隨時暫停、轉動場景
逐字稿（14 段）
- 1Welcome to the robot training lab. Nobody wrote rules for this AI. It learned from examples.歡迎來到機器人訓練實驗室。沒有人替這個 AI 寫規則，它是從例子裡學的。
- 2This is artificial intelligence. Artificial means it is made by people, not natural.這就是人工智慧。「人工的」表示它是人做出來的，不是天生的。
- 3To teach it, people label thousands of pictures: cat, or not cat.要教它，人們先替幾千張圖片做標記：是貓，或不是貓。
- 4This is training. The system is shown the examples one after another.這就是訓練：把這些例子一張接一張拿給系統看。
- 5Then it works out the pattern by itself. Nobody writes the rule, and nobody can read it afterwards.接著它自己找出規律。沒有人寫下規則，事後也沒有人讀得懂。
- 6A model only learns from what it was shown. Unbalanced examples give unbalanced answers. That is bias.模型只能從看過的東西學。例子偏了，答案也會偏，這就叫偏誤。
- 7A language model predicts the words that most likely come next.語言模型預測的是接下來最可能出現的字。
- 8Its answer is plausible. It looks like it fits. But plausible is not the same as true.它的答案看似合理，好像塞得進去。但看似合理不等於正確。
- 9That is why an AI can sound confident and still be completely wrong.這就是為什麼 AI 可以講得很有自信，卻完全講錯。
- 10Sometimes it invents a fact or a source that does not exist. People call this a hallucination.有時它會捏造不存在的事實或出處，這叫做「幻覺」。
- 11So always verify anything that matters. Check it yourself.所以任何重要的東西都要查證，自己去核對。
- 12Ask it to show its reasoning. Mistakes are much easier to spot when you can see the steps.要它把推理過程講出來。看得到步驟，錯誤就容易找得多。
- 13Use its output to help you understand. Handing it in as your own is not learning.用它的輸出來幫助你理解。把它當成自己的交出去，就不是在學習。
- 14Now it's your turn. Click around the lab and explore.現在換你了。在實驗室裡到處點點看吧。
Learning From Examples從例子裡學
For most of computing history, humans wrote every rule. To recognise a cat, you would have to describe a cat in code — and nobody ever managed it.在電腦發展史的大部分時間裡，規則都是人寫的。要認出一隻貓，你得用程式碼描述貓長什麼樣——而從來沒有人做到過。
Machine learning takes the opposite approach. You show the system thousands of pictures labelled cat and not cat, and it works out the pattern itself.機器學習反過來做。你給系統看幾千張標好「貓」和「不是貓」的圖片，讓它自己找出規律。
Nobody writes the rule. Nobody can even read it afterwards — the knowledge is spread across millions of numbers.沒有人寫下那條規則。事後也沒有人讀得懂它——那份知識分散在幾百萬個數字裡。
This is powerful and also the source of most of its problems. A system that learned from examples will repeat whatever was in those examples, including the mistakes.這既是它的威力所在，也是大部分問題的來源。從例子裡學來的系統，會複製例子裡的一切——包括其中的錯誤。
What to Keep in Mind使用時要記得的事
- Training data decides everything訓練資料決定一切
- A model can only learn from what it was shown. If the examples were unbalanced, its answers will be unbalanced in the same way. This is called bias.模型只能從被給看過的東西裡學。如果例子本身就偏了，它的答案也會照樣偏。這叫做偏誤（bias）。
- It predicts, it does not know它在預測，不是在知道
- A language model produces the words that most plausibly come next. Plausible is not the same as true, which is why it can be confidently wrong.語言模型產生的是「接下來最合理的字」。合理不等於正確，這就是為什麼它會非常有自信地講錯。
- Made-up answers have a name編造的答案有個名字
- When a model invents a fact, a source or a quotation that does not exist, people call it a hallucination. Always check anything that matters.當模型編出不存在的事實、出處或引文，這叫做「幻覺」（hallucination）。任何重要的東西都要自己查證。
- It is a tool, not an author它是工具，不是作者
- Using AI to help you understand something is learning. Handing in its output as your own is not, and you will not have learned anything either.用 AI 幫助你理解某件事是在學習。把它的輸出當成自己的交出去不是——而且你也什麼都沒學到。
- Ask it to show its working要它把過程講出來
- Asking for the steps or the reasoning makes mistakes far easier to spot, in the same way a maths answer with working is easier to check.要求它列出步驟或推理過程，錯誤會容易看出來得多——就像數學題有寫算式比較好檢查。
Two Ways to Build a System兩種建構方式
| Item 項目 | Traditional program 傳統程式 | Machine learning 機器學習 |
|---|---|---|
| Who writes the rules 誰寫規則 | A person 人 | The system finds them 系統自己找 |
| What you supply 你提供什麼 | Instructions 指令 | Examples 例子 |
| Can you read the rule 規則讀得懂嗎 | Yes 可以 | Usually not 通常不行 |
| Wrong answers 出錯時 | Traceable to a line 可追到某一行 | Hard to locate 難以定位 |
| Good for 適合 | Exact rules 規則明確的事 | Fuzzy patterns 難以言明的規律 |
資料查證至 2025 年底
算薪水用傳統程式，因為規則明確而且必須完全正確。認出照片裡的貓用機器學習，因為沒有人寫得出「貓的規則」。選錯工具，兩邊都會做得很糟。
Did You Know?你知道嗎？
- The term artificial intelligence was coined in 1956, so the field is older than most people think.artificial intelligence 這個詞 1956 年就出現了，這個領域比多數人以為的老。
- Progress came in waves, with long quiet periods in between that researchers call AI winters.進展是一波一波來的，中間有長期的沉寂，研究者稱之為「AI 寒冬」。
- A model does not remember your conversation unless it is designed to. Each request usually starts fresh.模型不會記得你的對話，除非它被設計成那樣。每次請求通常都是重新開始。
- Being able to describe a problem clearly in words is now a genuinely useful skill, because that is how you get a useful answer.「把問題用文字講清楚」現在是真正有用的能力，因為那決定你能不能得到有用的答案。
Key Words單字表
artificial/ˌɑːr.t̬əˈfɪʃ.əl/adj.
人工的
Artificial intelligence is not natural intelligence.人工智慧不是天然的智慧。
pattern/ˈpæt̬.ɚn/n.
規律、模式
It works out the pattern itself.它自己找出規律。
label/ˈleɪ.bəl/v. / n.
標記
The pictures are labelled cat or not cat.圖片被標記為貓或不是貓。
training/ˈtreɪ.nɪŋ/n.
訓練
Training data decides everything.訓練資料決定一切。
bias/ˈbaɪ.əs/n.
偏誤
Unbalanced data creates bias.不均衡的資料造成偏誤。
predict/prɪˈdɪkt/v.
預測
It predicts the next words.它預測接下來的字。
plausible/ˈplɑː.zə.bəl/adj.
看似合理的
Plausible is not the same as true.看似合理不等於正確。
confident/ˈkɑːn.fə.dənt/adj.
有自信的
It can be confidently wrong.它可能非常有自信地講錯。
invent/ɪnˈvent/v.
捏造
It may invent a source.它可能捏造出處。
verify/ˈver.ə.faɪ/v.
查證
Always verify anything important.重要的東西一定要查證。
reasoning/ˈriː.zən.ɪŋ/n.
推理
Ask it to show its reasoning.要它把推理過程講出來。
output/ˈaʊt.pʊt/n.
輸出
Do not hand in its output as your own.不要把它的輸出當成自己的交出去。
讀完了，來考一下
測驗卷共 100 分，有單字配對、選擇、填空與簡答。 按下去會開啟列印版，可以直接印出來，或在列印視窗選「另存為 PDF」。