ฝึกพูดภาษาอังกฤษด้วยเทคนิค Shadowing จากวิดีโอ: AI Concepts Every Product Manager Should Know in 25 Minutes | Crash Course

เลือกโหมดฝึก

แต่ละโหมดจะเปิดห้องฝึกแบบเต็มจอ

241 ประโยค

หากประโยคสั้นหรือยาวเกินไป กดที่ Edit เพื่อปรับแก้

Whether you're interviewing for a PM role or already in one, it's getting harder to keep up without understanding AI.

Because every company is building with it, and they expect you to know how it actually works

so you don't ship something that looks impressive but fails in the real world.

That's why in this video, I skip all the buzzwords

and break down all the AI concepts you need to know as a product manager

so you can build something that can justify a huge token bill.

Let's get into it.

First, let's start by understanding what an AI model actually is.

At a high level, a model is a system that learns patterns from data like text,

images, audio, or numbers, and then uses those learned patterns to make predictions or generate outputs.

One of the most important modern model types is the Large Language Model, or LLM.

LLMs are models designed to understand and process human language.

Models like ChatGPT are powered by LLMs, which is why users can ask questions conversationally and receive responses back in natural language.

The important thing to understand here is that LLMs are not thinking of a response in the same ways humans do.

Instead, they predict likely outputs based on patterns learned from massive amounts of text and training data.

Over time, these models have become even more advanced and evolved into multimodal models, meaning they can process more than just text,

including images, audio, and other types of input.

And as these models have become more advanced, they've also become capable of creating new content, which is what we refer to as Generative AI.

This includes capabilities like generating images, writing code, and even drafting OKRs.

Now that we understand what these models are and what they're capable of, of, let's talk about how they actually learn from data to produce useful outputs.

At a high level, there are two foundational approaches to how models learn, which are supervised and unsupervised learning.

Supervised learning is when the model is trained on label data, where each example comes with a correct answer.

For example, let's say you are training an AI model that powers a customer support chatbot.

You might give the model a set of customer support tickets and label each one as a billing issue, bug, or feature request.

The model then learns to associate certain inputs with those known outputs.

On the other hand, unsupervised learning is

when the model is given data without correct answers and has to discover patterns and relationships on its own.

Back to the customer support ticket example, instead of being told whether a ticket is a billing issue,

bug, or feature request, the model would have to identify those relationships on its own by finding patterns

and similarities across the data.

From a PM perspective, this matters because supervised learning requires label data, which can be expensive and time-consuming to create.

Unsupervised learning can work without those labels, but you have less control over what patterns the model learns or how the outputs are organized.

And in most modern AI products, you're often working with a mix of both.

So even if you train these models on a lot of high-quality data, the outputs can still vary even when you give the model the exact same input.

That's because AI models can be non-deterministic, meaning the same input does not always produce the same output.

Instead of following fixed rules, the model is choosing from multiple possible responses that could all be reasonable.

This is because during training, the model learns patterns from data instead of memorizing exact answers.

So when you give it the same input twice, it isn't pulling from a fixed answer bank.

Instead, it predicts what response is most likely to make sense based on the patterns it learned during training.

And because there can be multiple reasonable ways to respond, the output can vary slightly each time.

For example, if you ask the model, write a professional follow-up email after an interview, one response might start with, thank you for your time,

while another might start with, it was great speaking with you today.

Both responses are reasonable, which is why the output can vary even for the exact same input.

On the other hand, a deterministic model means the same input will always produce the same output.

However, most modern AI models are non-deterministic, meaning they can produce slightly different outputs for the same input.

And for many use cases, that's actually useful, because there isn't actually only one perfect answer.

If a user asks for a meeting summary, restaurant recommendation, or travel itinerary, there are multiple reasonable ways the model can respond.

Great, so now we understand what models are, how they learn, and how they behave.

These models are already incredibly powerful, but even more value gets unlocked when you build systems and then products on top of them,

which is why we're seeing such a massive influx of AI companies.

And while a lot of these companies are getting dismissed as ChatGPT wrappers, though I'm sure a lot of them are, they're doing much more than simply calling a model's API or LLM,

which brings us to our next section, how AI products are built.

One of the first problems you run into

when building an AI product is designing systems

that give the model the context it needs to answer questions or execute tasks reliably.

For example, imagine you work at a clothing store

and you're building a customer support chatbot to help with returns on your platform.

If a customer asks, what is your refund policy?

The model might give a generic response based on publicly available return policies it saw drain training data, but it wouldn't actually know your store's real refund policy.

So how do you give models the right information when they need it?

That's where RAG comes in.

RAG stands for Retrieval Augmented Generation.

At a high level, it's a way to give the model access to more context

so it can produce more accurate and relevant answers.

Instead of relying only on what the model learned during training, the system first retrieves relevant information and gives it to the model as context before responding.

So going back to the same customer support chatbot example, When a customer asks what is your refund policy,

RAG allows the system to first look up your company's actual refund policy from internal documentation before responding.

So now the model isn't just responding based on what it learned during training.

It's responding using relevant context specific to your product or use case.

RAG is powerful for two main reasons.

First, it allows the model to use contextual data and knowledge instead of relying only on what it learned during training.

And secondly, the context can be updated at any time without needing to rebuild or retrain the model.

RAG allows the system to retrieve relevant information before responding.

But how does the system actually know what information is relevant in the first place?

That's where embeddings come in.

Embeddings convert information like text,

images, and audio into numerical representations so AI systems can compare and understand how similar different data is to each other.

Back to the customer support chatbot example, let's say a customer asks, can I get my money back?

If the system has a document titled internal refund policy, embeddings will help the system understand that the document is closely related to the user's request,

even though the wording is completely different.

But once you've converted all of your data into embeddings, you need a way to store and search through them efficiently.

That's where vector databases come in.

A vector database is a system designed to store embeddings and quickly find the most similar ones.

So in the same example, if a customer asks, can I get my money back?

The system converts the question into an embedding, compares it against embeddings stored in the vector database and retrieves the most relevant information,

which in this case would be the internal refund policy.

You can think of it like a search engine, but instead of matching keywords, it's matching meaning.

Vector embeddings also allow the system to scale.

So even if you have thousands or millions of data points, vector databases enable you to quickly find the most relevant ones.

And this is what allows Reg to actually find and retrieve relevant information in AI products.

But even with the right information, the model might still not behave the way you want.

What if the customer support chatbot responded like a b?

That's where we have fine-tuning, which adjusts the way the model behaves.

Instead of giving the model more contextual information, Fine-tuning further trains the model to respond in a more specific or consistent way.

For the same customer support chatbot example, you might fine-tune a model to respond in a company's brand voice,

respond in a structured format such as for complaints, start with an apology,

handle a particular type of task more consistently such as make

sure every response ends with the option to talk to a real customer service rep.

You can think of it like this, RAG and embeddings help the system retrieve the right information,

100

while fine-tuning shapes how the model responds using that information.

101

But even with the right knowledge and the right behavior, the model is still mostly just responding.

102

Agents take this one step further.

103

Instead of just generating responses, agents allow the model to take actions.

104

This means the model can interact with tools, call APIs, or perform tasks on behalf of the user.

105

So going back to the same customer support chatbot example, instead of just answering with a refund policy, an agent could look up the user's order,

106

check if it's still eligible for a refund, initiate the refund, and confirm when it's completed.

107

Instead of the user having to take the next step, the system can handle the workflow end-to-end.

108

That means fewer steps for the user, faster outcomes, and a much more autonomous product.

109

For the typical consumer, this is a big shift because instead of just getting answers, they can now get things done way faster.

110

And you're already seeing how powerful this can be with tools like coding agents.

111

GitHub Copilot, for example, claims developers complete tasks up to 55% faster when using it.

112

And what makes these agents even more powerful is that they can connect to other systems and complete work across multiple tools,

113

not just within their own product.

114

One of the most popular ways AI systems interact with external tools is through Model Context Protocol, or MCP.

115

At a high level, MCP is a standardized way for models to connect to external tools,

116

data sources, and APIs so they can complete end-to-end workflows without needing custom integrations for every tool.

117

So going back to the customer support chatbot example, imagine your order is arriving tomorrow and you ask the chatbot to remind you when it's delivered.

118

Instead of just giving you the delivery update, the AI system could connect to Google Calendar through MCP and create the reminder for you automatically.

119

In this case, MCP gives the model a standardized way to interact with external tools like Google Calendar

120

so it can actually take actions on the user's behalf instead of only responding with information.

121

Now that we understand the basics of how AI products work, let's look at some of the most common patterns used to design AI products.

122

We'll start with patterns that require the most human involvement and then end with systems that are the most autonomous.

123

The first pattern is the chat interface.

124

This is where users interact directly with AI by asking questions or giving instructions, like with ChatGPT or Gemini.

125

It's the most familiar way people use AI today.

126

The benefit is that it's flexible as users can ask anything, explore ideas, and guide the interaction however they want.

127

But the downside is that it requires effort from the user to get the outcome they want.

128

Users have to know what to ask, how to ask it, and what to do with that answer.

129

For example, if you ask ChatGPT to help you plan a trip to Japan, you still need to guide the conversation, refine the prompts, decide which recommendations to follow,

130

and actually book the flights, hotels, and activities yourself.

131

So while the chat pattern is powerful, it puts more responsibility on the user to guide the process and get the results they want.

132

Then we have the co-pilot pattern.

133

This is where AI assists the user while they're doing a task, but doesn't take full control.

134

A great example of this is Superhuman, an AI email co-pilot.

135

It helps you draft replies, summarize conversations, and organize your inbox.

136

It can suggest responses and speed up your workflow, but you're still in control of what actually gets sent.

137

So AI is augmenting your workflow, not replacing it.

138

The next pattern is human-in-the-loop workflows, where instead of the AI simply assisting you,

139

it does most of the work and the human mainly steps in to review or approve the final output.

140

A good example is Harvey AI, a platform that helps lawyers save time on legal work.

141

Harvey can draft and revise legal documents like contracts, identify legal risks and inconsistencies, and prepare negotiation summaries and recommendations.

142

but lawyers still review and approve the final deliverables.

143

Keeping a human in the loop is important in high-stick situations like legal work where accuracy and trust matter a lot.

144

And lastly, we have the most autonomous pattern, which is the agent pattern.

145

This is where AI can take action on behalf of the user and complete tasks end-to-end.

146

A good example is CloudCode, a tool that helps developers write code and build applications.

147

Instead of just answering questions or suggesting code, you can give it a high-level task like build me a weather app, and it can plan the steps,

148

write the code, debug issues, and iterate toward a working solution automatically.

149

In this case, AI moves beyond just assisting the user and starts executing tasks on its own, with little to no human intervention during the process.

150

But building AI products isn't just about making them capable.

151

They also need to be reliable and accurate.

152

Because if AI products give inaccurate answers, make bad decisions, or behave unpredictably,

153

users quickly lose trust and the product… doesn't work.

154

So how do we know whether these AI systems are actually working reliably?

155

This is where evaluations or evals come in.

156

are how you systematically measure whether your AI system is actually performing well in real-world scenarios.

157

You start by creating an eval set, which is a collection of test cases that represent the kind of scenarios your product will see in the real world.

158

Each test case includes an input, like a user question, and an output, which is what a good response should include.

159

So for the customer support chatbot example, you might have an input like what's your refund policy

160

and an output explaining the refund policy like items can be returned within 30 days

161

if they are unworn and still have the original tags attached.

162

Then you run the AI system against the eval set and evaluate how well the responses perform.

163

This can be done manually by comparing the system's output against the expected output in the eval set,

164

or through platforms like Braintrust that help automate these evaluations for you.

165

As you improve the product over time, you keep running the same evals to make sure the quality improves.

166

And as your product evolves, you'll often expand the eval set with new test cases to cover new workflows and edge cases.

167

Within the eval set, you often define a smaller set called a golden eval set.

168

These are your most important test cases, the ones that absolutely need to work in your product.

169

You use these as a benchmark to make sure critical functionality stays reliable.

170

For the customer support chatbot example, a test case in the golden eval set might be an input that states, I think someone hacked my account.

171

And a good response would be correctly guiding the user through securing their account.

172

and resetting their password without exposing sensitive information or bypassing security steps.

173

This would belong in the golden eval set because security-related workflows are high risk and cannot afford to fail in the product.

174

As a PM, you're not aiming for perfection.

175

You're defining what's good enough to ship.

176

So for an MVP version of your AI product, that usually means passing your golden eval set and not aiming to get every eval test case correct.

177

Now let's take a step back and zoom out.

178

Even with strong evaluations and clear quality benchmarks, AI products are never perfect.

179

There are always trade-offs, and as a PM, your job is to decide which trade-offs matter most based on the user experience, business goals, and level of risk your product can tolerate.

180

One of the most common trade-offs is latency versus quality.

181

Higher quality models tend to be slower and more expensive, while faster models often come at the cost of accuracy.

182

So you're constantly balancing how good the answer is with how quickly you can deliver it.

183

The next trade-off is model size versus speed.

184

Larger models are generally more powerful, but they require more compute, which increases both latency and cost.

185

Smaller models are faster and cheaper, but they may not perform as well.

186

For example, even companies that aren't AI native,

187

like Disney, can generate millions of requests and consume billions of tokens per day using tools like Claude or ChatGPT.

188

So even small inefficiencies in model choice can translate into significant costs at scale.

189

Lastly, one of the most important trade-offs to understand is safety versus capability.

190

More capable models can handle increasingly complex tasks, but they also introduce greater risk.

191

As these systems become more powerful and autonomous, stronger guardrails and controls become increasingly important.

192

Here are a few common risks that tend to emerge as a result.

193

The first one is hallucinations, which happens when the model gives confident but incorrect answers.

194

For example, Ontario tested AI medical transcription tools

195

and found that some systems made up incorrect patient information and treatment suggestions that doctors never actually said.

196

In high-stakes industries like healthcare, mistakes like this can have serious consequences.

197

The next issue is bias, which occurs when the model reflects patterns from its training data in ways that might be unfair and inaccurate it.

198

For example, Amazon once built an AI hiring tool

199

that became biased against women because it was trained on historical hiring data from a male-dominated industry.

200

Next, we have data leakage and privacy risks.

201

As models gain access to more internal or sensitive data to become more useful,

202

they also increase the risk of exposing that information in their outputs.

203

For example, example, Samsung engineers accidentally leaked confidential source code by pasting it into ChatGPT,

204

raising concerns about sensitive information being shared outside the company.

205

And lastly, we have prompt injection,

206

which is when a user tries to manipulate the model into ignoring its instructions or doing something unintended.

207

For example, a Chevrolet dealership's AI chatbot was manipulated by users into ignoring its instructions

208

and agree to sell a car for only one dollar.

209

If safeguards aren't properly implemented,

210

the model might follow those instructions and expose information or behave in ways you didn't intend.

211

On top of the technical trade-offs in these systems, there are also user-side challenges that can block adoption of AI products.

212

First, we have trust and reliability.

213

Users often don't trust AI systems to be correct, especially when the outputs can vary or occasionally be wrong.

214

For example, if you ask ChatGPT to help you with your math homework and it gets the answer wrong, even once, you'll stop trusting it.

215

Next, we have privacy and compliance concerns.

216

Users and companies are hesitant to share sensitive data due to security and regulatory risks.

217

For example, employees avoid putting internal documents into tools like ChatGPT if they're unsure where the data goes.

218

The next user-side challenge is workflow disruption.

219

AI often changes how people work and users resist tools that don't fit naturally into their existing processes.

220

For example, if an AI feature doesn't integrate directly into the user's workflow and requires extra effort to use,

221

it often gets ignored.

222

And finally, AI systems may lack explainability, which erodes user adoption.

223

If users don't understand how or why an AI system made a decision, it becomes much harder to trust and adopt.

224

For example, if you're using Amazon's AI Shopping Assistant and it recommends a product without explaining why it was recommended,

225

users are much less likely to trust the suggestion.

226

So now that we've covered AI fundamentals, how AI products are built, and the trade-offs and risks involved,

227

the next question is, what do you do with all of this information as a PM?

228

Your job as a PM isn't to just add AI everywhere, it's to decide where AI meaningfully improves the product experience.

229

That means thinking about what workflows can be automated, where AI can assist versus take action, and where human control is still important.

230

At the same time, you should also be using AI in your own day-to-day workflows.

231

For example, using AI to help draft product requirement docs or summarize other docs, analyzing data or user feedback,

232

responding to common questions or triaging feature requests, or quickly prototyping ideas before involving engineering.

233

The goal isn't to replace your role, it's to amplify it.

234

Use AI as an accelerator for execution so you can spend more time on what actually matters – strategy, decision making, and creative thinking.

235

And that's where AI really becomes game-changing.

236

And that's it for this video.

237

I hope it was helpful in giving you a breakdown on

238

all the AI concepts you'll need to know as a product manager.

239

Hopefully this video made the space feel a little bit less overwhelming

240

and gave you something you can actually apply to your day-to-day workflows or projects as well.

241

Thanks so much for watching and I'll see you in the next one.

ดาวน์โหลดแอป

ทุกอย่างที่ต้องใช้เพื่อพูดได้คล่อง

AI ให้คะแนนการออกเสียงให้คะแนนทุกประโยค

ฝึก IPAเชี่ยวชาญทุกเสียง

คำศัพท์สร้างคลังคำของคุณ

เกมคำศัพท์เล่นไปเรียนไป

สแกน QR

Download on theApp Store Get it onGoogle Play

ช่องเดียวกัน

How to Answer Situational Questions for Product (PM) Behavioral Interview | Framework + Solution

Interview EnglishUniversity StudentsWorking Adults

How to Answer “Tell me About Yourself” in Product Manager (PM) Interviews (Framework + Example)

Interview EnglishUniversity StudentsWorking Adults

FOR YOU

วิดีโอแนะนำ

105 ประโยค

Google’s AI Course for Beginners (in 10 minutes)!

Business EnglishUniversity StudentsWorking Adults

213 ประโยค

The Technical, Business and UX Skills Every Product Manager Needs

Business EnglishUniversity StudentsWorking Adults

120 ประโยค

Learn Data Modeling in 8 minutes: Dimensional Data Modeling, Data Vault, and One Big Table

Business EnglishUniversity StudentsWorking Adults

222 ประโยค

คำศัพท์และข้อสังเกตด้านการพูดสำหรับบทเรียนนี้

วิดีโอนี้มี 241 ประโยค และ 3769 คำ สำหรับฝึกพูดตาม ช่วงที่มีเสียงพูดยาว 24:00 ผู้พูดพูดด้วยความเร็วเป็นธรรมชาติ ประมาณ 157 คำต่อนาที ใกล้เคียงกับบทสนทนาในชีวิตประจำวัน มีเพียง 81% ของคำที่อยู่ใน 3,000 คำที่ใช้บ่อยที่สุดในภาษาอังกฤษ คำศัพท์จึงค่อนข้างยาก

คำศัพท์สำคัญในวิดีโอนี้

คำที่ควรเรียนรู้ 15 คำจากวิดีโอ พร้อมคำอ่านและความหมาย:

คำศัพท์	คำอ่าน	ความหมาย
output คำนาม	/ˈaʊtpʊt/	เอาท์พุต
vector คำนาม	/ˈvɛktɚ/	เวกเตอร์
label คำนาม	/ˈleɪ.bəl/	ฉลาก
database คำนาม	/ˈdeɪtəˌbeɪs/	ฐานข้อมูล
tune คำนาม	/tjuːn/	ทำนอง
behave คำกริยา	/bɪˈheɪv/	ประพฤติ, ทำตัว
ignore คำกริยา	/ɪɡˈnɔː/	เมิน
connect คำกริยา	/kəˈnɛkt/	เชื่อม, ต่อ
expose คำกริยา	/ɪkˈspoʊz/	เผย, เปิดเผย
compare คำกริยา	/kəmˈpɛɚ/	เปรียบเทียบ
hire คำกริยา	/ˈhaɪɹ/	จ้าง
privacy คำนาม	/ˈpɹaɪ.və.si/	ความเป็นส่วนตัว
loop คำนาม	/luːp/	บ่วง
reliable คำคุณศัพท์	/ɹɪˈlaɪəbl̩/	ไว้ใจได้, เชื่อถือได้
suggestion คำนาม	/səˈd͡ʒɛst.ʃən/	ข้อเสนอ, ข้อเสนอแนะ

การออกเสียงที่ควรระวัง

ผู้พูดใช้รูปย่อและรูปลดเสียง 40 ครั้ง เช่น you're, isn't, don't ให้พูดแบบสั้นตามที่ได้ยิน

เสียง “th”: clothing /ˈkloʊðɪŋ/, mathematics /mæθ(.ə)ˈmæt.ɪks/
เสียง “sh” และ “zh”: instruction /ɪnˈstɹʌkʃən/, recommendation /ˌɹɛkəmɛnˈdeɪʃən/, suggestion /səˈd͡ʒɛst.ʃən/, adoption /əˈdɑp.ʃən/, associate /əˈsoʊʃi.ət/
คำยาว — ลงเสียงหนักให้ถูกพยางค์: autonomous /ɔˈtɑnəməs/, recommendation /ˌɹɛkəmɛnˈdeɪʃən/, accuracy /ˈæk.jɚ.ə.si/, developer /dɪˈvɛləpɚ/, capability /ˌkeɪ.pəˈbɪl.ɪ.ti/

วิธีฝึกกับวิดีโอนี้

ฟังวิดีโอให้จบหนึ่งรอบโดยยังไม่ต้องพูด แล้วจดคำที่ยังไม่รู้จัก
เริ่มที่ความเร็ว 0.75× พูดตามทีละประโยค แล้วกลับไปใช้ความเร็วปกติเมื่อเริ่มคล่อง
อัดเสียงตัวเองแล้วเทียบกับต้นฉบับ โดยสังเกตคำอย่าง output, vector, label เป็นพิเศษ

ไวยากรณ์ในวิดีโอนี้

โครงสร้างที่ผู้พูดใช้บ่อยที่สุด พร้อมคำพูดจริงจากวิดีโอ:

โครงสร้าง	ในวิดีโอ
Passive voice be + กริยาช่อง 3 — เน้นสิ่งที่เกิดขึ้น ไม่ใช่ผู้กระทำ	are powered · are supervised · is trained
Relative clauses who / which + อนุประโยค — ข้อมูลเพิ่มเติมเกี่ยวกับคนหรือสิ่งของ	LLMs, which is · content, which is · learn, which are
Present perfect have/has + กริยาช่อง 3 — เหตุการณ์ในอดีตที่ยังเกี่ยวข้องกับปัจจุบัน	have become · you've converted · we've covered

เทคนิค Shadowing คืออะไร?

Shadowing เป็นเทคนิคการเรียนรู้ภาษาที่ได้รับการรับรองทางวิทยาศาสตร์ พัฒนาขึ้นสำหรับการฝึกนักแปลมืออาชีพ วิธีการนี้เรียบง่ายแต่ทรงพลัง: คุณฟังเสียงภาษาอังกฤษจากเจ้าของภาษาและพูดตามทันที — เหมือนเงาที่ตามผู้พูดด้วยช่วงเวลาห่าง 1-2 วินาที การวิจัยแสดงว่าเทคนิคนี้ปรับปรุงความแม่นยำในการออกเสียง ทำนองเสียง จังหวะ การเชื่อมเสียง การฟังเข้าใจ และความคล่องแคล่วในการพูดได้อย่างมีนัยสำคัญ

เทคนิค shadowing: อ่านคู่มือฉบับเต็มทีละขั้นตอน →

ฝึกพูดภาษาอังกฤษด้วยเทคนิค Shadowing จากวิดีโอ: AI Concepts Every Product Manager Should Know in 25 Minutes | Crash Course

เลือกโหมดฝึก

ดาวน์โหลดแอป

How to Answer Situational Questions for Product (PM) Behavioral Interview | Framework + Solution

How to Answer “Tell me About Yourself” in Product Manager (PM) Interviews (Framework + Example)

Google’s AI Course for Beginners (in 10 minutes)!

The Technical, Business and UX Skills Every Product Manager Needs

Learn Data Modeling in 8 minutes: Dimensional Data Modeling, Data Vault, and One Big Table

The Multi-Agent Architecture That Actually Ships — Luke Alvoeiro, Factory

The 7 phases of AI-driven development

9 Mental Models You Can Use to Think Like a Genius

How AI is Changing Data Scientist's Workflow

Give Me 18 Minutes and I’ll Make you Dangerously Smart (with AI)

Supervised Learning: Crash Course AI #2

MCP vs API: The protocol every developer needs to know

Understand AI in 14 minutes – with Anthropic's Chloe Lubinski [ARC 2026]

Using AI to read research articles

คำศัพท์และข้อสังเกตด้านการพูดสำหรับบทเรียนนี้

คำศัพท์สำคัญในวิดีโอนี้

การออกเสียงที่ควรระวัง

วิธีฝึกกับวิดีโอนี้

ไวยากรณ์ในวิดีโอนี้

เทคนิค Shadowing คืออะไร?