跟讀練習: How AI gets its character - 通過影片學習英語口語
選擇練習模式
每種模式都會開啟全屏練習室。
51 句
如果句子過短或過長,請點選 Edit 進行調整。
1
Hi there.
2
My name is Maggie and I lead the education team at Anthropic.
3
Today I'm here to talk to you about how AI assistants end up with a disposition.
4
We'll look at the two training stages that turn raw prediction into something useful, the fingerprints those stages leave behind and how knowing those fingerprints help you get better results.
5
Why does an AI try to be helpful in the first place?
6
Why is it polite?
7
Why does it refuse certain things?
8
Knowing that an AI predicts the next word doesn't really answer any of that.
9
Helpfulness is built deliberately in layers and each layer influences your experiences with AI each day.
10
Modern AI assistants are built in two stages.
11
Stage one is pre-training.
12
The model sees enormous amounts of data and learns one thing.
13
Given everything so far, guess what comes next?
14
That's it, repeated billions of times.
15
Stage two is fine-tuning.
16
The document completer from stage one gets trained again this time on curated examples of helpful behavior and
17
reward signals shaped by human preferences this is the layer that turns the ai model into an assistant imagine
18
you could talk to a model that had only been through stage one no fine-tuning at all you type what is the capital of France.
19
A raw pre-trained model doesn't answer your question, it continues your document.
20
Maybe it outputs Paris, what's the capital of Germany, Berlin, what's the capital of Spain, and so on, because it's seen that pattern in quizzes.
21
Maybe it writes a paragraph from a geography textbook, maybe it generates more questions.
22
It has no concept of you, no concept of helping.
23
It's purely continuing a document in whatever direction seems statistically likely.
24
The assistant behavior you actually experience with AI tools today is a trained overlay on top of that.
25
Fine-tuning is what makes generative AI systems usable and useful.
26
But because it relies on human judgments about what good looks like, the texture of those judgments shows up in these models' personality.
27
Often these personality traits are what make generative AI so effective.
28
But there can be a shadow side to AI's helpfulness.
29
Four shadow areas are, one, sycophancy.
30
When people prefer agreeable responses, the model learns to validate readily and back down under light pushback, even when it was right the first time.
31
Two, verbosity.
32
When thoroughness scores better during training, the model defaults to longer answers, even when brevity could serve you better for a specific situation.
33
Three, over-caution.
34
When safety training leans conservative, the model can hedge heavily or refuse requests that are actually safe.
35
And four, loose confidence calibration.
36
The model's stated confidence is only loosely tied to its actual liability.
37
Confidence is genuinely hard to train, so it's particularly important to be vigilant here.
38
These aren't bugs in one particular model.
39
They're things that show up in all AI models.
40
However, the quality and type of fine-tuning done on a model directly shapes how these things manifest, and it will likely be different from model to model.
41
At Anthropic, we train Claude to be broadly safe, ethical, and helpful.
42
You can even read Claude's entire constitution to see how we train Claude and how we intentionally shape Claude's personality.
43
Why does this matter to you?
44
Understanding how AI is made and why it behaves the way it does puts you in control when it comes to AI.
45
If your AI assistant caves the moment you push back, that's sycophancy, and you should factor that in when assessing responses.
46
If you're getting essays when you want bullets, that's the verbosity default kicking in.
47
If you're getting heavy caveats on a harmless question, that's over-caution.
48
We'll address what to do about this in the upcoming lessons.
49
The assistant you talked to wasn't born helpful.
50
That behavior was built layer by layer, and sometimes the seams show.
51
Learning to spot these seams is part of using AI well.
同頻道
FOR YOU
推薦影片
為什麼通過這個影片練習口語?
練習口語的過程中,使用生動的例子是非常有幫助的。在這段影片中,講者深入講解了人工智慧是如何被訓練得以表現出有用的行為。學習者可以通過模仿講者的表達方式,提高自己的口語能力。隨著聽力理解的提升,學習者能夠更自然地參與對話,提升與他人交流的自信心。這是一個非常適合 看YouTube學英語 的機會,尤其是在討論現代科技和人工智慧話題時,能夠幫助學員擴充套件詞彙和句型。
語法與表達在語境中的應用
在這段影片中,講者使用了許多重要的語法結構和表達方式,對學習者來說很有啟發性。以下是一些關鍵結構:
- “Why does an AI try to be helpful?” - 這個問題結構引導了討論,適用於在對話中詢問對方的觀點。
- “The model sees enormous amounts of data and learns one thing.” - 使用了“sees... and learns”的形式,可以幫助學習者理解並練習複合句的構造。
- “It has no concept of you, no concept of helping.” - 重複的句型使得資訊更為強調,適合練習描述和解釋。
- “The assistant behavior you actually experience...” - 這個結構使得表述更加具體且形象,為表達提供了生動的例子。
- “These aren't bugs in one particular model.” - 使用“aren't”作為否定形式,幫助學習者理解如何在不同語境下進行反駁或澄清。
常見發音陷阱
在影片中,某些詞彙和發音可能會對學習者造成挑戰。以下是一些需要注意的發音陷阱:
- “AI” - 確保清晰發音,避免模糊的母音音響。
- “helpful” - 注意這個詞的重音和發音,應該突出“help”的發音.
- “behavior” - 發音時注意連讀以增強說話的流暢度。
- “training”與“training”之間的區別 - 確保音調的變化,以表達不同的意思。
通過反覆模仿和提高英語發音,學習者能夠更好地掌握這些發音,增強他們的口語流利度。運用英語影子跟讀的方法,可以在日常練習中事半功倍。
什麼是跟讀法?
跟讀法 (Shadowing) 是一種有科學依據的語言學習技巧,最初開發用於專業口譯員的培訓,並由多語言者Alexander Arguelles博士普及。這個方法簡單而強大:您在聽英語母語原聲的同時立即大聲重複——就像是一個延遲1-2秒緊跟說話者的影子。與被動聽力或語法練習不同,跟讀法強迫您的大腦和口腔肌肉同時處理並模模擬實的講話模式。研究表明它能顯著提高發音準確性,語調,節奏,連讀,聽力理解和口語流利度——使其成為雅思口語備考和真實英語交流最有效的方法之一。










![Understand AI in 14 minutes – with Anthropic's Chloe Lubinski [ARC 2026]](https://img.youtube.com/vi/aBUniZHgCnE/mqdefault.jpg)








