シャドーイング練習: What Is Jev? The AI Model That Doesn't Generate Text - 動画で英語スピーキングを学ぶ

練習モードを選ぶ

各モードは全画面の練習ルームで開きます。

169 文

文が短すぎたり長すぎる場合は、Editをタップして調整してください。

all right i want to talk about jev jev's a new ai model from a startup called type safe

and it's a bit unusual because it's an ai model

that doesn't actually write any text it actually answers questions by picking from a list

and then giving a probability for each one of his answers

now type safe calls this a new type of ai model it's a system one model

which is a name borrowed from Daniel Kahneman in his book one of my favorites thinking fast

and slow Kahneman describes two modes of thinking human thinking

that is and the first of those modes is called system one thinking

and system one thinking is fast and automatic

so what's two times two you instantly know the answer is

four without having to really work it out in your head that's system one thinking the other type

of thinking is classified as system two thinking and that's slow

and deliberate so

if i ask you what's 17 times 24 well you're probably going to need to work through

that one step by step to figure it out

or just look at the answer on your teleprompter

which is uh 408 there we go got it now most

of the ai models we use are built to write stuff

and a chatbot produces its answers one token at a time

in a ford pass now a reasoning model goes a bit further than

that and it actually writes out its chain of thought as we're going through this

so you can actually kind of see what it was thinking

which is basically as close to a system two model as ai gets

but but a lot of decisions inside software are not really system two they're just kind of quick judgment calls

so let me give you an example of that let's say

that we have got a email for customer support and

that email says that I was charged twice this month please fix it all right

so how do we respond to this email well the software

that's handling this email probably has to look at three types

of question one of the questions is is this a refund request yes

or no another question might be

which team needs to process this email maybe the billing team

or maybe it's the technical support team or maybe it's the sales team

and then also how urgent is this request how quickly do we need to respond

now a person just skimming an email like that they would instantly know

that refund yeah it's yes this is a refund request that's kind of a system one bit of work

and it's that kind of work that jev is built for

so a common way to automate a support email problem like

this today using an llm would be to send this email to the llm

and ask for the answers back the answers to these things in a structured format

so maybe we would go through this large language model output some tokens

and these tokens would represent a format maybe in json now

that mostly works

but it doesn't come back with a reliable sense of how

sure the model is how confident is a large language model

that it got the answer to these questions right well we could ask the model

but whatever it responds with wouldn't necessarily actually match the real

probabilities the model used to come up with this answer

but jev well jev does come up with probabilities so

if we send these questions to JEV we're going to get a very different looking response

so it takes in two things it takes in what's called state

which is the data the decision is about so

that might include our support email

that might include some recent customer charges as well all of

that is input into the JEV model

and it also takes in each of these questions as well into the model as input

and our three questions are actually three different types of questions so refund that is a yes

or no question

which type safe calls a null a team this team question it really comes down to a choice as to

which team is this so picking one option from a list

and then urgent

that really comes down to a score on a scale from

low to critical now all three of these questions go in in a single request

and whereas an llm would write its json answer out one token at a time jeff doesn't generate any text

so all three answers they all come back at once

and that's a big part of why this is so much faster

and cheaper than an llm on these types of questions now

we provide questions as input what is coming back i said

jeff doesn't give you text what it gives you coming back our probabilities as the output

so for refund it's going to be a single number let's

say it's going to come back with 0 .9 meaning there is a 90 chance that yes this is a

refund request for the team it's a probability for each one of the options that were part of this choice.

So for example, the billing choice, which is one of the three, may come back with a probability score of 0 .85.

And by the way, the answer to choice can only be the options that we give it.

So for team, it's always going to be billing or technical or sales.

And then urgency, well, that's going to come back if we kind of put this on a scale, pretty high to critical here so this is quite important.

Now, the model can still pick the wrong options or score any of these questions incorrectly,

of course, but the probability tells us how likely that is.

So how do you train a model to give accurate probabilities like this?

Well, let me go back to training traditional LLMs for a moment.

Most language models, they start with a stage known as pre -training.

So you pre -train a model first.

That means the model reads a huge amount of text and it learns to predict the next token.

Now that creates a model that knows a lot of stuff but it isn't much use to talk to yet.

So if we want to use this as a chatbot then we need to go to a second stage, which is called post -training.

Now, a lot of post -training uses a technique called reinforcement learning.

The model produces an answer, something scores that answer, and the model kind of gets nudged towards the answers that score well.

so the question is what's doing the scoring well one way to do that is with a technique called rl

hf reinforcement learning from human feedback

so people are doing the scoring there human raters look at two responses from the model

and they pick the one they like best

and those choices train a reward model and then

that chatbot is going to be trained to give answers

that the reward model scores highly people tend to like answers

that sound confident so an unfortunate side effect of rlhf is

that a model can learn to sound like really sure of itself even

when it's wrong

but that's just one form of reinforcement learning another form is r l v r that's reinforcement learning with verifiable rewards.

And the scoring here is automatic.

100

Did the model get the right answer to the math problem?

101

Does the code it wrote pass its unit tests?

102

Those sort of things.

103

And that's a big reason that reasoning models today have got so good at math and coding.

104

But the reward here only checks whether the final answer was right.

105

It doesn't reward the model for knowing how sure it should be.

106

And those models spend a

107

while working through their chain of thought before they come up with an answer

108

and that can make them a little bit slow and expensive to run.

109

Now JEV is trained with a different form of reinforcement learning.

110

It's called RLCD, reinforcement learning for calibrated decisions.

111

Now TypeSafe haven't published much about JEV's architecture

112

so i'll stick to what they have said well we know that jeff gives a probability

113

for each option and it's rewarded when those probabilities turn out to be right

114

so that's what the C is all about it's saying the model is calibrate it.

115

And a model that's calibrated kind of basically looks like this.

116

So along the bottom axis here if we map probability

117

and then along the side axis here we can map whether the model is correct.

118

So a well calibrated model is going to have kind of a diagonal line much like this.

119

So when we say that there is a 80 probability that 80 probability is going to be correct about 80

120

of the time that's calibration so back to the the refund question

121

that was the first of the three questions we posed to jeff about

122

that support email so

123

if we have probabilities we can use them to inform code using a threshold like this

124

so we've got a probability threshold here one is certainty zero is definitely not kind of thing

125

so if a probability score is above let's say 0 .9

126

so maybe our score comes back here somewhere

127

that means we can treat this support email as a legit refund request

128

and we can send that straight to the refund queue for processing

129

if however the score comes back anywhere kind of around between here and let's say 0 .1

130

well there's some uncertainty here so the code

131

that received a number like this would know to actually get a human involved

132

so now we need to bring in a person to actually check

133

that because we're not sure and

134

if a score comes in below 0 .1 well i think we can say that's not a refund request

135

so keep calm and carry on we don't need to do anything with that one because the numbers are calibrated,

136

the threshold also tells us roughly how often the automatic path will be wrong.

137

So the more a mistake is going to cost, the higher we would set the threshold.

138

Now this works for more than support emails of course.

139

Jeff can act as a guardrail

140

that can be used around all sorts of things like around a chat bot that's checking messages going in

141

and out for things like well jailbreak attempts

142

but there are limits jev only takes text in as of today it's not very good at math

143

or even counting so you're going to want to leave

144

that stuff to something else

145

and like any ai model it can be tricked by instructions hidden in the data it's reading

146

so is it time to throw away large language models

147

and replace them completely with system one models no hardly

148

but the two can work together so

149

if you think about a workflow maybe we have jev in the workflow

150

that can make quick calls like sorting

151

that customer support email to see whether it's a refund or not

152

and then in the workflow we could go to a large

153

language model to do the slower work work like replying to the customer

154

and then we could go back to use Jev to classify any response

155

that came back from the customer

156

which is pretty much how Danny Kahneman said we work too

157

most of what we do runs on those fast automatic system one processes

158

so that's these two stages here and then the slow and careful system two, which is the large language model in this case,

159

that only kicks in when something needs real thought.

160

Oh, and I wanted to share a thing about the name of Jev as well.

161

Jev is named after William Stanley Jevons, who was an economist who pointed out in 1865

162

that as steam engines got more efficient britain actually used more coal that's known as jevons paradox

163

so perhaps the same thing will happen with system one models

164

when a judgment call is super fast

165

and super cheap like with jev well it can go into places where an llm would be too slow

166

or too expensive today like in every row of a database or on every line of a log file. Who knows?

167

All right, that's Jev.

168

Where would a System 1 model fit into the software that you work on?

169

Let me know in the comments.

アプリをダウンロード

流暢に話すために必要なすべて

AI発音採点文ごとに採点

IPA練習すべての音をマスター

単語学習自分の単語帳を作る

単語ゲーム遊びながら覚える

QRをスキャン

Download on theApp Store Get it onGoogle Play

📺 同じチャンネル

✨ おすすめ動画

26 文

How does AI actually work?

Daily ConversationUniversity StudentsWorking Adults

183 文

The Great A.I. Debate: Should We Fear the Digital Revolution? [2015] | Intelligence Squared

Daily ConversationUniversity StudentsWorking Adults

33 文

Enhancing AV Simulation with Generative AI - DRIVE Labs Ep. 36

Daily ConversationUniversity StudentsWorking Adults

215 文

Creator of C# and TypeScript: "AI will NEVER Replace Coders, Here's Why" | Anders Hejlsberg

Daily ConversationUniversity StudentsWorking Adults

407 文

Why AI Tokens are so Expensive - Computerphile

Daily ConversationUniversity StudentsWorking Adults

182 文

We Let $200 GPT-Pro Think for 7 Days on Unsolved Math

Daily ConversationUniversity StudentsWorking Adults

553 文

How AI Is Changing Strategy—and 4 Skills Leaders Need Next

Daily ConversationUniversity StudentsWorking Adults

75 文

What AI can (and can't) currently do in a real lab | Jennifer Doudna

Daily ConversationUniversity StudentsWorking Adults

159 文

AI Is Stealing From You

Daily ConversationUniversity StudentsWorking Adults

63 文

A $6.3 billion open-weight model just got embarrassed by the French...

Daily ConversationUniversity StudentsWorking Adults

188 文

NVIDIA GTC 2025 - Keynote Highlights by NVIDIA CEO Jensen Huang

Daily ConversationUniversity StudentsWorking Adults

このレッスンの語彙とスピーキングのポイント

この動画には、シャドーイング用の文が169文、単語が2101語あります。音声の長さは14:51です。話す速さは1分あたり約141語で安定しており、シャドーイングしやすいペースです。単語の85%は英語の頻出3,000語に含まれます。残りは練習の前に確認しておきましょう。

この動画の重要語彙

動画に出てくる覚えておきたい単語15語を、発音と意味つきで紹介します。

単語	発音	意味
probability 名詞	/ˌpɹɑ.bəˈbɪl.ə.ti/	確率
refund 名詞	/ɹɪˈfʌnd/	払い戻し, 返金
reward 名詞	/ɹɪˈwɔɹd/	見返り, 賞
token 名詞	/ˈtoʊkən/	トークン
automatic 形容詞	/ˌɔː.təˈmæt.ɪk/	自動の, 自動的な
threshold 名詞	/ˈθɹɛʃ(h)oʊld/	敷居
mathematics 名詞	/mæθ(.ə)ˈmæt.ɪks/	数学
format 名詞	/ˈfɔːɹ.mæt/	書式, フォーマット
technique 名詞	/tɛkˈniːk/	技術
output 名詞	/ˈaʊtpʊt/	出力
judgment 名詞	/ˈd͡ʒʌd͡ʒ.mənt/	判決, 裁き
urgent 形容詞	/ˈɜː.d͡ʒənt/	緊急の, 急を要する
axis 名詞	/ˈæksɪs/	軸
trick 名詞	/ˈtɹɪk/	欺き, 企み
unusual 形容詞	/ʌnˈjuːʒ(u)əl/	珍しい

動画に出てくる句動詞

単語	意味
come back 動詞	戻る, 帰る
throw away 動詞	棄てる, 捨てる

この動画の文法

話し手がよく使っている文型を、動画の実際の表現とともに紹介します。

文型	動画での表現
受動態 be + 過去分詞 — 誰がするかより、何が起きるかに焦点を当てる	is called · is classified · are built
関係詞節 who / which + 節 — 人や物について情報を加える	choice, which is · stage, which is · two, which is
現在完了形 have/has + 過去分詞 — 過去の出来事が今も関係している	haven't published · have said

注意したい発音

話し手はdoesn't, we're, you'reなど、短縮形や弱形を17回使っています。聞こえたとおりの短い形で発音しましょう。

「th」の音: threshold /ˈθɹɛʃ(h)oʊld/, mathematics /mæθ(.ə)ˈmæt.ɪks/
「sh」と「zh」の音: threshold /ˈθɹɛʃ(h)oʊld/, unusual /ʌnˈjuːʒ(u)əl/, efficient /ɪˈfɪʃənt/, instruction /ɪnˈstɹʌkʃən/, calibration /ˌkæl.ɪˈbɹeɪ.ʃən/
長い単語（アクセントの位置に注意）: probability /ˌpɹɑ.bəˈbɪl.ə.ti/, automatic /ˌɔː.təˈmæt.ɪk/, mathematics /mæθ(.ə)ˈmæt.ɪks/, necessarily /ˌnɛs.əˈsɛɹ.ə.li/, unusual /ʌnˈjuːʒ(u)əl/

この動画での練習方法

まず声を出さずに動画を最後まで聞き、知らない単語をメモします。
通常の速度で一文ずつシャドーイングし、話し手のリズムに合うまで繰り返します。
自分の声を録音して元の音声と比べます。probability, refund, rewardなどの単語に特に注意しましょう。

シャドーイングとは？英語上達に効果的な理由

シャドーイング（Shadowing）は、もともとプロの通訳者養成プログラムで開発された言語学習法で、多言語習得者として知られるDr. Alexander Arguelles によって広く普及されました。方法はシンプルですが非常に効果的：ネイティブスピーカーの英語を聞きながら、1〜2秒の遅延で声に出してすぐに繰り返す——まるで「影（shadow）」のように話者を追いかけます。文法ドリルや受動的なリスニングと異なり、シャドーイングは脳と口の筋肉が同時にリアルタイムで英語を処理・再現することを強制します。研究により、発音精度、抑揚、リズム、連音、リスニング力、そして会話の流暢さが大幅に向上することが確認されています。IELTSスピーキング対策や自然な英語コミュニケーションを目指す方に特におすすめです。

シャドーイングのやり方: ステップ別の完全ガイドを読む →

シャドーイング練習: What Is Jev? The AI Model That Doesn't Generate Text - 動画で英語スピーキングを学ぶ

練習モードを選ぶ

アプリをダウンロード

Every Generation Thinks This Time Is Different: The AI Job Myth

What is Event Driven Architecture (EDA)?

Agentic AI Frameworks Explained: Workflows, Multi-Agent, & Production

AI, Machine Learning, Deep Learning and Generative AI Explained

Spec-Driven Development: AI Assisted Coding Explained

AI Agents vs Business Rules: Which Should Make Decisions?

AI & Education: Generative AI & the Future of Critical Thinking

EDR vs. EPP vs. NGAV

MCP vs ADK: How Modern AI Agents Connect and Work Together

Cybersecurity Architecture: Data Security

The 7 Skills You Need to Build AI Agents

How Smart Organizations Will Use AI: Jevons Paradox & Workforce Impact

What are AI Agents?

Cybersecurity Architecture: Five Principles to Follow (and One to Avoid)

Zero Trust Explained in 4 mins

How does AI actually work?

The Great A.I. Debate: Should We Fear the Digital Revolution? [2015] | Intelligence Squared

Enhancing AV Simulation with Generative AI - DRIVE Labs Ep. 36

Creator of C# and TypeScript: "AI will NEVER Replace Coders, Here's Why" | Anders Hejlsberg

Why AI Tokens are so Expensive - Computerphile

We Let $200 GPT-Pro Think for 7 Days on Unsolved Math

How AI Is Changing Strategy—and 4 Skills Leaders Need Next

What AI can (and can't) currently do in a real lab | Jennifer Doudna

AI Is Stealing From You

A $6.3 billion open-weight model just got embarrassed by the French...

NVIDIA GTC 2025 - Keynote Highlights by NVIDIA CEO Jensen Huang

このレッスンの語彙とスピーキングのポイント

この動画の重要語彙

動画に出てくる句動詞

この動画の文法

注意したい発音

この動画での練習方法

シャドーイングとは？英語上達に効果的な理由