📺 视频信息

Title: The Inside Story of ChatGPT’s Astonishing Potential (ChatGPT 惊人潜力的幕后故事)
Speaker: Greg Brockman
Accent: General American (通用美式发音)
Duration: 30:09 (Total)

🎧 语音现象标注说明 (Legend)

为了让你更直观地看到发音技巧，我使用以下符号进行标注：

📜 语音现象标注全文 (Annotated Transcript)

(Legend Review: Bold=Stress; Italics**=Weak forms; ‿=Link; (t)/(d)=Stop; /d/=Flap T; //=Pause)**

OpenAI // was founded // seven years ‿ ago.
(years ago s与a连读)
It's been ‿ a busy few months.
(been a n与a连读)
We released Chat G P T, // we released G P T-4, // and we've seen the world // really ‿ embrace these tools.
(really embrace y与e连读)
But I want ‿ to take you back // a little bit // to the early days.
(want to 弱读为 wanna; little bit t浊化)
Back to // twenty-seventeen.
We had just // developed ‿ a system // that could predict // the next character // in Amazon ‿ reviews.
(developed a d与a连读; predict ct失爆)
And if you look ‿ at what this model learned, // it learned ‿ a lot ‿ of syntax.
(look at k与a连读; learned a d与a连读; lot of t浊化)
It learned // where the commas go // and how to spell words.
But also // something miraculous happened.
It learned // sentiment.
It learned ‿ to tell you // if ‿ a review // was positive // or negative.
(learned to d失爆; if a f与a连读)
And in fact, // this one neuron // in the model, // if you forced ‿ it // to be positive // or negative, // it would actually // generate // the corresponding text.
(forced it d与i连读)
And this // was our first glimpse // into what // would become // Chat G P T.
It showed ‿ us // that the key // to AI // was ‿ in scaling ‿ up // these systems.
(showed us d与u连读; was in s与i连读; scaling up ng与u连读)
So we built // G P T-1.
And then // G P T-2.
And then // G P T-3.
And with each step, // the capabilities // got // better.
(got t失爆; better tt浊化)
And now, // we are starting ‿ to see // technology // that can do tasks // that humans // find difficult.
For ‿ example, // in the G P T-4 launch, // we showed ‿ it // solving // complex // tax problems.
(showed it d与i连读)
And people // started ‿ to ask, // "Wait, // is this // real?"
(Wait t失爆)
Is this really // going ‿ to change // everything?
(going to 弱读为 gonna)
And I think // that the answer // is yes.
But I also think // that it looks // different // than what people // expected.
It looks // less like // a giant // overmind // and more like // a personal // assistant.
An always-available // expert // that is by your side // to help you // do your best work.
(help you p与y连读)
And so to show you // what I mean, // I wanna show you // a live demo.
(want to 弱读为 wanna)
Let's // take ‿ a look.
(take a k与a连读)

(Demo begins on screen)

So, // this is // Chat G P T.
And what I wanna do // is show you // something // that we haven't released yet.
(haven't t失爆)
Something // that brings ‿ it // even closer // to that vision.
(brings it s与i连读)
We call ‿ them // plug-ins.
(call them l与th连读/同化)
And the idea // is that instead ‿ of just // talking ‿ to the model, // you can actually // have the model // use tools.
(instead of d与o连读; talking to g失爆)
Just like you do.
So, // let's say // that I want ‿ to cook // a nice meal.
And maybe // I want ‿ to be // a little bit // more ‿ ambitious // than normal.
(more ambitious r与a连读)
I might ask // Chat G P T, // "Hey, // help ‿ me come ‿ up // with ‿ a great // meal plan."
(help me p与m同化; come up m与u连读; with a th与a连读)
But not just that.
I want ‿ it ‿ to make // a shopping list // for me.
(want it t与i连读; it to t失爆)
And I wanna have ‿ it // sent // to Instacart // so I can get // the ingredients // delivered.
(have it v与i连读; get t失爆)
So let's see // what it does.
So as you can see, // it starts ‿ off // by using // the Instacart plug-in.
(starts off s与o连读)
And it's asking // questions // about // what I want ‿ to eat.
And I can say, // "You know, // I wanna do // something // vegetarian."
And now, // it's going ‿ to actually // use its knowledge // to come ‿ up // with ‿ a recipe.
(going to 弱读为 gonna)
It suggests // chickpea salad // sandwiches.
Which // actually sounds // pretty good.
(pretty tt浊化)
But I want ‿ to make // one change.
I want ‿ to see // what it looks ‿ like.
(looks like s与l连读)
So I can use // a different plug-in.
This one // is DALL-E.
It generates // images.
And so I ask ‿ it, // "Hey, // generate // an image // of this sandwich."
(ask it k与i连读; image of dg与o连读)
And what you see // is that Chat G P T // doesn't just // pass the text // to DALL-E.
It actually // writes // a prompt.
It says, // "Hey, // Chat G P T, // describe // this sandwich // in vivid detail // so that DALL-E // can draw ‿ it."
(draw it w与i连读)
And there // you go.
That // looks // delicious.
(looks s失爆)
And now, // back // to Instacart.
It adds // the items // to my cart.
And all I have ‿ to do // is click // a link.
(have to v弱化为f)
And I'm ready // to order.

Now // I wanna show you // something // a little bit different.
(want to 弱读为 wanna; little bit t浊化)
We call ‿ this // the code ‿ interpreter.
(call this l与th连读)
And this // is an A I // that can write // code // and run ‿ it // for you.
(run it n与i连读)
So, // I uploaded // a file here.
It contains // data // about // the last thirty years // of E G O T winners.
(Emmy, Grammy, Oscar, Tony)
That's // Emmy, // Grammy, // Oscar, // Tony.
And I wanna ask // something // interesting // about this data.
I ask, // "Who // is // the O G // E G O T?"
(OG = Original Gangster/Originator)
Who is // the person // who won // all four // first?
And what you see // is that the model // writes // Python code.
(writes Python s与P组合，s稍停)
It loads // the data file.
It parses // the dates.
It sorts ‿ them.
(sorts them s与th同化)
And it finds // the person // who won // all four // first.
Richard Rodgers.
Creator // of the Sound ‿ of Music.
(Sound of d与o连读)
Pretty cool.
(Pretty tt浊化)
But let's // try // something // a little bit // more ‿ involved.
(more involved r与i连读)
Let's ask ‿ it // to make // a graph.
(ask it k与i连读)
I want ‿ to see // how the age // of winners // has changed // over time.
(age of dg与o连读)
So, // once ‿ again, // it writes // Python code.
(once again s与a连读)
It calculates // the age.
It plots ‿ it.
(plots it s与i连读)
And here // we have ‿ it.
(have it v与i连读)
You can see // the average age // of winners // over time.
And it looks ‿ like // it's going ‿ up // a little bit.
(looks like s与l连读; going up ng与u连读)
Now, // this // is just // the tip // of the iceberg.
Be cause // the model // isn't just // writing code.
It can also // use // that code // to interact // with the world.
It can send // emails.
It can make // reservations.
It can buy // products.
And this // is really // where we see // the potential.
//
Now, // I want ‿ to talk // a little bit // about // how we built ‿ this.
(built this t失爆)
How // did we get // from G P T-3, // which just predicted // the next word, // to Chat G P T?
Well, // the secret // ingredient // is // people.
We use // a technique // called // Reinforcement Learning // from Human Feedback.
(Reinforcement t失爆)
Or R L H F.
The idea // is very simple.
We train // the model // by having // humans // rate // its outputs.
(rate its t与i连读)
So, // if the model // says // something // rude, // or factually // incorrect, // a human // flags ‿ it.
(flags it z与i连读)
And the model // learns, // "Okay, // don't do // that // again."
(that again t浊化)
And over // time, // through millions // of interactions, // the model // aligns // with human // values.
This is // crucial // for safety.
Be cause // as these models // get // smarter, // we need ‿ to make sure // they are also // getting // safer.
(need to d与t同化; getting tt浊化)
Imagine // if we had // a super intelligent // system // that didn't // care // about // truth // or harm.
That // would be // a disaster.
So, // safety // isn't // an afterthought.
It's baked // into // the process.
(baked into d与i连读)
And this // is why // we release // iteratively.
Why we put // Chat G P T // out // there // as ‿ a research // preview.
(as a z与a连读)
We need // real-world // feedback // to find // the edge cases.
(edge cases dg与c组合)
To find // the things // we missed // in the lab.
//
I think // we are at // a very // exciting // moment // in history.
We are building // tools // that can amplify // human // potential.
Just like the steam engine // amplified // our physical strength, // A I // can amplify // our intellectual // capabilities.
It can help ‿ us // solve // hard problems.
(help us p与u连读)
It can help ‿ us // be // more // creative.
But // it requires // all ‿ of ‿ us // to participate**.
(all of us 连读)
To guide // this technology // in the right // direction.
Thank you.

💡 高级词汇与地道表达 (Vocabulary & Expressions)

🗣️ 练习建议 (Tips)

1. Technical Acronyms (首字母缩写):
   在演讲中，演讲者大量使用了缩写，如 GPT, AI, EGOT, RLHF。
   • 技巧: 缩写的每一个字母通常都需要重读，并且如果有元音开头的字母（如 A, E, I, O, R, S, F, L, M, N, X），前一个词尾可能会与之连读。
   • Example: "an A I" (an 与 A 连读); "G P T" (T 发音清晰有力)。
2. Conversational vs. Technical Tone (语调转换):
   Greg Brockman 在演示 Demo 时（例如点餐、生成图片），语调比较轻松、口语化，使用了 "wanna", "cool", "pretty good" 等词。但在解释原理（RLHF, Safety）时，语调变得严肃、缓慢、清晰。
   • 练习: 试着模仿他在 "Pretty cool" 时的轻快感，以及在 "Safety is not an afterthought" 时的坚定感。
3. The "Stop T" (失爆/不完全爆破):
   这是美式英语听力的难点。注意单词结尾的 t，如果后面接辅音，通常不发音。
   • Example: "jus(t) like", "go(t) better", "wai(t)".
   • 注意: 不要把 t 发得很重，只要舌头点到齿龈位置停顿一下即可。
4. Interactive Phrasing (交互式表达):
   演讲中充满了 "So, let's take a look", "And what you see is...", "Here we have it"。这些是做 Presentation 时的黄金句型，非常值得背诵积累，用于引导观众视线。