跟讀練習: The easy way to learn nothing - 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。
229 句
如果句子過短或過長,請點選 Edit 進行調整。
1
Every few years, 15-year-olds in over 90 countries all take the same test, and in the most recent version of this, for the very first time,
2
it asks students how often they use chatbots like ChatGPT for their schoolwork.
3
Here in Canada, the teens who are using AI for their work every week were about two years behind the ones who never use it.
4
Two years behind.
5
I was blown away by this a little bit.
6
I was sort of expecting there to be a discrepancy between the two, but two years behind felt like a lot.
7
So I actually started digging into more research
8
and I found studies where students using AI learned twice as much as students who didn't use it.
9
So which is it?
10
Is it something that helps or hurts?
11
Now, before we go any further into this, the Canadian number comes with a bit of an asterisk because it is a correlation.
12
It doesn't prove that AI caused those lower scores.
13
And the researcher who dug into the results of the data says that themselves.
14
But the uses with the biggest gaps are the ones you'd sort of expect.
15
The ones where AI does the work for you, right?
16
Students who are using it every day to summarize their readings or to draft their writing,
17
they scored about 80 points lower than the students who never used AI.
18
Like with everything else, AI keeps getting pitched as a revolution, not just in how we work, but in how we run our lives.
19
And of course, now in how we learn as well.
20
Learn faster with AI.
21
Learn this in 40 minutes instead of 40 hours.
22
Less effort and better results.
23
And that's the part I mostly want to push back on because it frames learning as something that you can shortcut.
24
The thing is, when it comes to actually learning something, we learn through friction.
25
The struggle isn't the thing getting in the way.
26
The struggle is the learning.
27
And that's because understanding something isn't the same as actually being able to do it.
28
And we've known this for a long time, right?
29
Before the AI revolution and all of that, the whole idea of tutorial hell comes from the same gap.
30
You understand something well enough to be able to sort of get it, understand how it works, but you're not actually able to use it.
31
I always like comparing this to learning how to skateboard.
32
You can watch a two minute video
33
or maybe a bunch of videos on YouTube on how to do a kickflip
34
and you can understand every single step that you'd have to do to be able to do a kickflip.
35
But that first time you step on it, you're probably going to fall on your no matter how many tutorials you've watched.
36
That same gap exists for learning anything, and there's a study that actually shows this really well. In 2019,
37
a group of Harvard researchers published a paper in the Proceedings
38
of the National Academy of Sciences of the United States of America, which is a mouthful, so often that is referred to PNAS for short.
39
Yes, most people call it PNAS.
40
And that study was called Measuring Actual Learning vs. Feeling of Learning in Response to Being Actively Engaged in the Classroom.
41
They took students and gave them identical course material, but one group got a regular lecture and the other group had to actively work through it.
42
And they found that students in the active classroom learn more, but they feel like they learn less.
43
Not only that though, it goes beyond that.
44
The students in the lecture felt like their teacher was better and they enjoyed the class more.
45
And the explanation behind this was basically that when something is explained smoothly, it feels like learning.
46
Beginners aren't great at judging how much they actually know.
47
And the struggle that comes from active learning feels like it's not working
48
when really the struggle is what's causing the learning to happen.
49
The study I just talked about was done before AI chatbots were a thing.
50
What happens when you actually hand students one?
51
Well, the team from the University of Pennsylvania ran an experiment with almost 1,000 high schools in Turkey, which was also published in PNAS.
52
They split the students into three groups for their math practice sessions.
53
One group got the regular chat GPT, however you want to use it.
54
Another group got a version set up as a tutor, which would give hints, but wouldn't hand over the answer.
55
And one group got nothing, just their textbook and notes.
56
And as you, I guess, expect during practice, the regular chat GPT group crushed it.
57
They got 48% more problems right than the kids with no AI.
58
And the ones who had the tutor version did even better, getting 127%.
59
Then they all took the exam with no AI allowed.
60
And you probably guessed the regular chat GPT group scored 17% worse than the kids who had never had it.
61
And I guess important thing here maybe is the ones that had the tutor set up.
62
They only did just as well as those that were in the control group.
63
But here's the part that got me the most from the study is the chat GPT group didn't realize it.
64
They didn't think that they'd done any worse or learned any less.
65
And now you might be listening to me saying all this and go, OK, but Kevin, this is all about high school math.
66
I'm a developer and I get it.
67
So let's turn our attention to Anthropic, the company that makes Claude, who ran a study on exactly this earlier this year.
68
They took 52 developers who all knew
69
Python already, but had never used a particular library and they had them build a couple of features with it.
70
Half of them could use an AI assistant and half of them had to write everything themselves.
71
And then they took a quiz on what they just built.
72
The group that had access to AI averaged 50% and the hand coding group averaged 67%.
73
But the biggest gap of all, which is probably the most important one, was on questions about debugging,
74
where they said the largest gap in scores between the two groups was on debugging questions, suggesting that the ability to understand when code is incorrect
75
and why it fails may be a particular area of concern if AI impedes coding development.
76
And I think that makes sense because debugging is where you find out
77
if you actually understand the thing or if you just have code that happens to work.
78
Now, one of the things that usually we're told is, well, with AI, we go faster.
79
But interestingly enough, on this study, the AI group wasn't actually faster.
80
It says they saved on average about two minutes, but it wasn't statistically significant.
81
So in other words, they traded learning and they didn't get speed in return.
82
And this is from a study published by one of the AI giants telling us that it's the greatest productivity tool ever,
83
yet we are still seeing a bit of a lack of evidence around that.
84
Now, one thing with all of this, I want to be really clear here
85
because I don't think anyone who falls into this is being lazy or is doing something wrong on purpose.
86
It's just really, really easy to slide into this because the answer is right there.
87
Even back in the day, we had Stack Overflow and you copy and pasted stuff from there all the time, right?
88
But even that had a bit of friction built into it.
89
You had to figure out what to search for.
90
You had to find the answer that was close to your problem.
91
You usually had to adopt the code to your own code
92
and half the time it wouldn't work the first time and you had to figure out why.
93
So you're doing all of that debugging and all of that other stuff right there.
94
Whereas now it's just a short prompt and it's done.
95
Then we can read that code, but that's the trap.
96
We look at it and we understand every single line.
97
So we think we know what's going on, but it's just like the kickflip again.
98
Understanding it when you watch it isn't the same as being able to do it.
99
Reading code and understanding it is not the same as being able to write or debug it when it breaks.
100
They're two completely different skill sets.
101
And as I said a little bit earlier, we've known this forever, right?
102
We had tutorial hell.
103
We've known this is a thing, but with AI, it's only reinforcing it even more.
104
And I'm not just talking about code here, but obviously this applies to everything.
105
AI can make study guides and flashcards and podcasts for you.
106
And I think a lot of that is just a waste of time, not to mention resources.
107
and you might be going, well, flashcards are useful.
108
And yeah, retrieval is a really important part of learning.
109
And flashcards are great for that.
110
But making those flashcards yourself is a fantastic extra layer of learning because to make them, you'd have to figure out what's important
111
and fill in all of those gaps in your knowledge
112
and actually know all of the stuff that's going to go on those flashcards.
113
So by the time that you're actually using them,
114
you already understand the topic a lot deeper than you would have if you're just using other flashcards.
115
And the knowledge is there and then you're focused on retrieval rather than using the flashcards to be like, oh, I don't actually know that.
116
And then having to learn it more, of course, making them all is more work.
117
But that's, as I've been saying since the beginning, is kind of the whole point.
118
But I still have an answer to question from the very beginning of this video when I said, if AI is bad for learning,
119
what about those studies that I talked about where some students were learning twice as much?
120
And so the first one we'll talk about is Harvard, who ran a study last year.
121
And this time it was on a physics course where students
122
learned from an AI tutor instead of like an active learning class.
123
And the students with the AI tutor learned twice as much and in less time.
124
The World Bank ran a six week program in Nigeria using an AI tutor after school
125
and the students made gains equal to roughly two years of typical learning, which is pretty big claim right there.
126
But if we take a look at both of these and look at how they're set up, the Harvard Tutor was custom-built for that course.
127
It was designed to walk students through problems step by step and make them do the thinking, not give them the answers.
128
And the Nigeria program had teachers guiding students on how to use it.
129
Even the Turkish math study that we talked about earlier, the Tutor version of it, the one that only gave hints,
130
it didn't hurt the students at all.
131
Those students did perform just as well as the ones that didn't have any access to AI at all as well.
132
And in the Anthropics study, not everyone who used AI did badly.
133
The developers who did well were the ones who asked more conceptual questions, things like why does this work this way?
134
And then they fixed their own errors themselves.
135
Those people scored about as well as the people who didn't use AI at all, and it was actually one of the faster approaches.
136
So it means that AI with all of this isn't necessarily the problem.
137
It's trying to use it as a shortcut that is the problem.
138
Every time it helped, it was set up to create friction and to make people think and every time it hurt, it was removing that friction.
139
And I know there'll be some pushback.
140
The people who are probably leaving comments like this aren't even watching all the way through to here,
141
but that's completely fine because I do want to talk about this anyway of the people we're going to say,
142
who cares because AI is writing the code anyway, so why do I even need to be able to write it myself?
143
I sort of feel like this is similar to the question my kids asked me of, well, we have a calculator,
144
so why do I need to learn long division?
145
And I get it in a way, but I also don't think it's actually the same thing and the research that we were just talking about,
146
it all points to why.
147
And that's because a calculator is always right.
148
You don't need to check its work.
149
AI isn't like that.
150
It gets things wrong and on top of that, of course, it gets things wrong confidently.
151
And what this means is we still need people with strong technical skills.
152
If we just take a look at what happened not too long ago with PHP,
153
I made a video on that talking about the disaster that that was, and there was a ton of comments of people saying,
154
well, just use AI to fix it.
155
The thing is, though, the issue is way beyond that though, right?
156
Upgrading from PHP 5 to a stable version isn't a one-shot prompt.
157
It's going to break things, and if you're running WordPress, you're probably going to need a new theme and a new set of plugins,
158
and it's a huge job.
159
It can be a job that AI helps you with, but it's not one of those just get it to fix it type of things.
160
And in the case of coding, I don't think the job goes away.
161
It changes.
162
Instead of writing code, maybe you're reviewing it.
163
I hear that from a lot of people, and not everyone loves that, but it is, or is becoming part of it.
164
But if you're reviewing it, it also means you're the one deciding if it's right and
165
if it's accessible and if it's going to fall apart at different screen sizes
166
and if it's doing something weird or maybe you see something in the code
167
that is a little bit of a red flag because it works fine now, but you know because you have the technical skills around this
168
that it's something that will come back to bite you six months from now.
169
And if you remember, if we go back again to that Anthropic study, the biggest gap was on debugging.
170
The exact skill you need when the AI code doesn't work.
171
The people who leaned on AI the most were the worst at figuring out what was wrong with the code,
172
and that's the part that Anthropic itself was worried about, because if people are going to be supervising AI,
173
they need the skills to actually catch the mistakes.
174
There was another study from Microsoft
175
and Carnegie Mellon where they surveyed a few hundred people who use AI at work every week,
176
and what they found was that the more people trusted AI, the less they thought critically about what it gave them.
177
And the less confident people were in the task itself, the more likely they were to just accept whatever the AI spat out.
178
And I'm sure you've seen this before in your own things, if you've used it at all,
179
where you find the mistakes and the things that are on the topics you know the best, or it might not always be a mistake,
180
but it's like, oh, it's not quite right.
181
So those are the ones you push back on.
182
Whereas if you're not an expert in the subject, everything it's doing, it just sounds right.
183
So you just tend to, okay, that sounds good.
184
And you go with it.
185
And that makes complete sense.
186
If you don't know CSS, how would you know the CSS it gave you is bad?
187
It looks fine.
188
It works on your screen.
189
Ship it. So really, I feel like we're actually in the state now where it's the opposite of what people are saying with the whole,
190
like, you don't need to learn it anymore thing.
191
The less you know, the less useful AI actually is because you can't tell good output from bad.
192
The more you know, the more you can get out of it
193
because you know what to ask for and when it's wrong.
194
I also know some people are going to comment, and again, these people probably aren't watching at this point,
195
but they're just going to say that I'm old and out of touch with the new way of doing things,
196
and that I just need to try the newest models because they're better than before.
197
I've been hearing that since the very beginning, and I know the models now are really good.
198
I know they can be impressive, and I see their outputs all the time,
199
but I still believe that expertise still matters and will continue to matter, and that means that learning is still important,
200
and taking shortcuts to that learning doesn't work.
201
The people getting great results out of these tools aren't the ones who skipped learning the fundamentals.
202
They're the ones who already have them.
203
So if you're trying to learn something new, whether or not you want to use AI, it doesn't really matter.
204
But just try and learn it yourself.
205
Read the docs.
206
Build something small.
207
Create an ugly demo.
208
Hit the wall and struggle.
209
And when you hit the wall, ask why.
210
And if you do decide to turn to AI, ask it why something's not working, not what isn't working.
211
Ask it to explain the concept and not write the code for you.
212
And if it does write something for you, build it yourself again afterwards.
213
Don't just read it and nod along
214
because that's where it feels like you've learned something when you haven't learned anything at all.
215
And I know there's going to be comments again about different ways that people have used AI to help them learn,
216
like making it act as a tutor and other things like that or taking a mess of notes and organizing it.
217
And I do think all of those things are great use cases for it.
218
And that's what the research supports as well.
219
But that's not what's being sold to us.
220
What's being sold to us is learn this in 30 minutes instead of 30 days.
221
And from everything I've read and seen, the 30 days isn't the problem.
222
It's where the learning actually happens and there are no shortcuts.
223
And if you just so happen to want to be learning CSS, I'm going to say check out my course,
224
CSS Demystified, because the whole point of that course isn't to learn about the syntax or anything like that.
225
It's to learn why CSS works the way it does.
226
You can start working with the browser instead of fighting against it.
227
So the link for that is just down below.
228
And with that, I'm going to say thank you very much for listening.
229
And of course, until next time, don't forget to make your corner of the internet just a little bit more awesome.
同頻道
FOR YOU
推薦影片
本課的詞彙與口語要點
這節口語課以影片“The easy way to learn nothing”為素材。 影片中反覆出現的詞有:learning, students, learn, code, study。 這段影片共有 229 個句子、3249 個單詞可供跟讀。 講話部分時長為 14:31。 說話人語速較快,每分鐘約 224 個詞,會出現較多連讀和弱讀。 88% 的單詞屬於英語最常用的 3,000 詞,詞彙容易理解。
影片中的重點詞彙
影片中 15 個值得學習的單詞,附發音和釋義:
| 單詞 | 發音 | 釋義 |
|---|---|---|
| tutor 名詞 | /ˈtjuːtəː/ | 私人教師 /私人教師, 家庭教師 /家庭教師 |
| skill 名詞 | /skɪl/ | 技巧, 技能 |
| friction 名詞 | /ˈfɹɪkʃən/ | 摩擦 |
| tutorial 名詞 | /ˌtjuːˈtɔːɹɪəl/ | 輔導 /輔導, 教程 |
| topic 名詞 | /ˈtɑpɪk/ | 主題 /主題, 話題 /話題 |
| gap 名詞 | /ɡæp/ | 間隙 /間隙 |
| output 名詞 | /ˈaʊtpʊt/ | 輸出 /輸出 |
| whereas | /(h)wɛɹˈæz/ | 然而 |
| classroom 名詞 | /ˈklɑːsɹʊm/ | 教室, 課堂 /課堂 |
| lecture 名詞 | /ˈlɛk.t͡ʃə/ | 演講 /演講 |
| hint 名詞 | /hɪnt/ | 暗示 |
| researcher 名詞 | /ˈɹiˌsɝ.t͡ʃɚ/ | 研究員 /研究員, 研究者 |
| calculator 名詞 | /kæl.kjə.leɪ.tɚ/ | 計算器 /計算器, 電子計算器 /電子計算器 |
| necessarily 副詞 | /ˌnɛs.əˈsɛɹ.ə.li/ | 必然地, 必定地 |
| stable 名詞 | /ˈsteɪ.bəl/ | 馬廄, 馬棚 /馬棚 |
影片中出現的短語動詞
| 單詞 | 釋義 |
|---|---|
| figure out 動詞 | 弄清楚 |
| check out 動詞 | 退房 |
| find out 動詞 | 找出, 查明 |
| go away 動詞 | 離開 /離開, 走 |
| go back 動詞 | 回去, 返回 |
| hand over 動詞 | 交出, 移交 |
值得反覆跟讀的句子
影片中簡短而完整的句子,可以直接用在日常對話裡:
- What happens when you actually hand students one?
- I'm a developer and I get it.
- You don't need to check its work.
- Instead of writing code, maybe you're reviewing it.
影片中的語法
說話人最常用的結構,並附上影片中的原話:
| 結構 | 影片中的用法 |
|---|---|
| 現在完成進行時 have/has been + -ing — 以前開始、現在仍在繼續的動作 | I've been saying · I've been hearing |
| 定語從句 who / which + 從句 — 補充說明人或事物 | ones who never · students who didn · Students who are |
| 被動語態 be + 過去分詞 — 強調發生了什麼,而不是誰做的 | was blown · is referred · was called |
| 現在完成時 have/has + 過去分詞 — 過去發生但與現在仍有關聯的事 | we've known · you've watched · you've seen |
需要注意的發音
說話人用了 82 次縮略和弱讀形式,例如 you're, don't, isn't。請按聽到的簡短形式來說。
- “th” 音: math /mæθ/, smoothly /ˈsmuːðli/, mouthful /ˈmaʊθfʊl/
- “sh” 和 “zh” 音: friction /ˈfɹɪkʃən/, explanation /ˌɛkspləˈneɪʃən/, crush /kɹʌʃ/, correlation /ˌkɔɹ.əˈleɪ.ʃən/, flashcard /ˈflæʃˌkɑɹd/
- 長單詞——注意重音位置: developer /dɪˈvɛləpɚ/, calculator /kæl.kjə.leɪ.tɚ/, necessarily /ˌnɛs.əˈsɛɹ.ə.li/, explanation /ˌɛkspləˈneɪʃən/, experiment /ɪkˈspɛɹ.ə.mənt/
中文母語者容易讀錯的音:
- /v/ — 不要讀成 /w/,上齒輕觸下唇: developer /dɪˈvɛləpɚ/, actively /ˈæktɪvli/, impressive /ɪmˈpɹɛsɪv/, productivity /ˌpɹɒdʌkˈtɪvəti/, retrieval /ɹɪˈtɹiːvəl/
- 詞尾子音 — 要讀清楚,後面不要加母音: math /mæθ/, topic /ˈtɑpɪk/, gap /ɡæp/, output /ˈaʊtpʊt/, whereas /(h)wɛɹˈæz/
- /r/ 和 /l/ 的區別: tutorial /ˌtjuːˈtɔːɹɪəl/, classroom /ˈklɑːsɹʊm/, necessarily /ˌnɛs.əˈsɛɹ.ə.li/, roughly /ˈɹʌf.li/, critically /ˈkɹɪt.ɪ.kə.li/
如何用這段影片練習
- 先完整聽一遍影片,不要開口,記下不認識的單詞。
- 先用 0.75 倍速逐句跟讀,熟練之後再回到正常速度。
- 錄下自己的聲音並與原聲對比,特別注意 tutor, skill, friction 這類單詞。
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。













