We're going to try something that as far as we know, no tech company has done before, which is every day for the next 12, every weekday, we are going to launch or demo some new thing that we built.

據我們所知，還沒有哪家科技公司做過這樣的嘗試，那就是在接下來的 12 個工作日裡，我們每天都要發佈或演示我們打造的新產品。

And we think we've got some great stuff for you starting today.

從今天開始，我們將為您準備一些精彩的節目。

We hope you'll really love it.

我們希望您會愛上它。

And you know, we'll try to make this fun and fast and not take too long, but it'll be a way to show you what we've been working on and a little holiday present from us.

你知道，我們會盡量讓這個過程有趣、快速，不會花太多時間，但這將是向你展示我們一直在努力的一種方式，也是我們送給你的一份小小的節日禮物。

So we'll jump right into this first day.

那麼，我們就直接進入第一天吧。

Today, we actually have two things to launch.

今天，我們實際上有兩件事情要發佈。

The first one is the full version of O1.

第一個是完整版的 O1。

We have been very hard at work.

我們一直在努力工作。

We've listened to your feedback.

我們聽取了您的反饋意見。

You like O1 Preview, but you want it to be smarter and faster and be multimodal and be better at instruction following, a bunch of other things.

你喜歡 O1 預覽版，但希望它更智能、更快、更多模式、更善於跟蹤指令等。

So we've put a lot of work into this.

是以，我們為此付出了很多努力。

And for scientists, engineers, coders, we think they will really love this new model.

對於科學家、工程師和程序員來說，我們認為他們會非常喜歡這種新模式。

I'd like to show you quickly about how it performs.

我想快速向你們展示一下它的性能。

So you can see the jump from GPT-4.0 to O1 Preview across math, competition coding, GPQA, Diamond, and you can see that O1 is a pretty big step forward.

是以，您可以看到從 GPT-4.0 到 O1 Preview 在數學、競賽編碼、GPQA 和 Diamond 方面的飛躍，您可以看到 O1 是一個相當大的進步。

It's also much better in a lot of other ways, but raw intelligence is something that we care about.

它在許多其他方面也更勝一籌，但原始智能是我們所關心的。

Coding performance in particular is an area where people are using the model a lot.

尤其是編碼性能，是人們經常使用該模型的一個領域。

So in just a minute, these guys will demo some things about O1.

稍後，他們將演示有關 O1 的一些情況。

They'll show you how it does at speed, how it does at really hard problems, how it does with multimodality.

他們會向你展示它在速度上的表現，在真正的難題上的表現，以及在多模態上的表現。

But first, I want to talk just for a minute about the second thing we're launching today.

但首先，我想談一下我們今天推出的第二件事。

A lot of people, power users of ChatGPT at this point, they really use it a lot and they want more compute than $20 a month can buy.

很多人都是 ChatGPT 的高級用戶，他們經常使用 ChatGPT，他們需要更多的計算能力，而不是每月 20 美元能買到的。

So we're launching a new tier, ChatGPT Pro.

是以，我們推出了一個新的層級--ChatGPT Pro。

And Pro has unlimited access to our models and also things like advanced voice mode.

專業版可以無限制地使用我們的模型，還可以使用高級語音模式等功能。

It also has a new thing called O1 Pro Mode.

它還有一個新功能，叫做 O1 專業模式。

So O1 is the smartest model in the world now, except for O1 being used in Pro Mode.

是以，除了在專業模式下使用 O1 外，O1 是目前世界上最智能的機型。

And for the hardest problems that people have, O1 Pro Mode lets you do even a little bit better.

對於人們遇到的最棘手的問題，O1 專業模式可以讓您做得更好。

So you can see a competition math.

是以，你可以看到一個競爭數學。

You can see a GPQA Diamond.

您可以看到 GPQA 鑽石。

And these boosts may look small, but in complex workflows where you're really pushing the limits of these models, it's pretty significant.

這些提升看似微不足道，但在複雜的工作流程中，當你真正挑戰這些模型的極限時，其意義就非常重大了。

I'll show you one more thing about the Pro Mode.

我再給你看一件關於專業模式的事情。

So one thing that people really have said they want is reliability.

是以，人們真正想要的是可靠性。

And here, you can see how the reliability of an answer from Pro Mode compares to O1.

在這裡，您可以看到專業模式的答案可靠性與 O1 相比如何。

And this is an even stronger delta.

這是一個更強大的三角洲。

And again, for our Pro users, we've heard a lot about how much people want this.

同樣，對於我們的專業版用戶來說，我們聽說過很多人都非常想要這個功能。

ChatGPT Pro is $200 a month, launches today.

ChatGPT Pro 每月 200 美元，今天推出。

Over the course of these 12 days, we have some other things to add to it that we think you'll also really love.

在這 12 天裡，我們還將添加一些其他內容，相信您也會非常喜歡。

But unlimited model use and this new O1 Pro Mode.

但模型使用不受限制，而且還新增了 O1 專業模式。

So I want to jump right in and we'll show some of those demos that we talked about.

是以，我想直接進入主題，我們將展示一些我們討論過的演示。

And these are some of the guys that helped build O1 with many other people behind them on the team.

這些都是幫助建立 O1 的一些人，在他們身後還有許多其他團隊成員。

Thanks, Sam.

謝謝，山姆。

Hi, I'm Hyungwon.

你好，我是邢原。

I'm Jason.

我叫傑森

And I'm Max.

我是麥克斯

We're all research scientists who worked on building O1.

我們都是研究 O1 的科學家。

O1 is really distinctive because it's the first model we've trained that thinks before it responds.

O1 的與眾不同之處在於，它是我們訓練的第一個先思考後反應的模型。

Meaning it gives much better and often more detailed and more correct responses than other models you might have tried.

這意味著，與您可能嘗試過的其他模型相比，它能提供更好、更詳細、更正確的回答。

O1 is being rolled out today to all Plus and soon-to-be Pro subscribers on ChatGPT, replacing O1 Preview.

O1 今天將在 ChatGPT 上向所有 Plus 和即將成為 Pro 的用戶推出，取代 O1 預覽版。

O1 model is faster and smarter than the O1 Preview model, which we launched in September.

O1 機型比我們在 9 月份推出的 O1 預覽機型更快、更智能。

After the launch, many people asked about the multi-modal input, so we added that.

推出後，很多人詢問了多模式輸入的問題，是以我們增加了這一功能。

So now the O1 model live today is able to reason through both images and text jointly.

是以，現在的 O1 模型可以同時對影像和文本進行推理。

As Sam mentioned, today we're also going to launch a new tier of ChatGPT called ChatGPT Pro.

正如 Sam 所說，今天我們還將推出一個新的 ChatGPT 層級，名為 ChatGPT Pro。

ChatGPT Pro offers unlimited access to our best models like O1, 4.0, and Advanced Voice.

ChatGPT Pro 可以無限制地使用我們的最佳機型，如 O1、4.0 和高級語音。

ChatGPT Pro also has a special way of using O1 called O1 Pro Mode.

ChatGPT Pro 還有一種特殊的 O1 使用方法，稱為 O1 Pro 模式。

With O1 Pro Mode, you can ask the model to use even more compute to think even harder on some of the most difficult problems.

有了 O1 專業模式，您可以要求模型使用更多的計算能力，更努力地思考一些最難的問題。

We think the audience for ChatGPT Pro will be the power users of ChatGPT.

我們認為 ChatGPT Pro 的閱聽人將是 ChatGPT 的高級用戶。

Those who are already pushing the models to the limits of their capabilities on tasks like math, programming, and writing.

那些已經在數學、編程和寫作等任務上將模型的能力推向極限的人。

It's been amazing to see how much people are pushing O1 Preview, how much people who do technical work all day get out of this, and we're really excited to let them push it further.

看到人們對 O1 預覽版的推崇，看到整天從事技術工作的人從中獲得的收益，我們感到非常驚訝，我們很高興能讓他們進一步推動 O1 預覽版的發展。

We also really think that O1 will be much better for everyday use cases, not necessarily just really hard math and programming problems.

我們也確實認為，O1 將更適合日常使用案例，而不一定只是非常困難的數學和編程問題。

In particular, one piece of feedback we received about O1 Preview constantly was that it was way too slow.

特別是，我們不斷收到關於 O1 預覽版的反饋，說它太慢了。

It would think for 10 seconds if you said hi to it, and we fixed that.

如果你跟它打招呼，它會思考 10 秒鐘，我們已經解決了這個問題。

That was really annoying.

這真的很煩人。

It was kind of funny, honestly.

老實說，這有點滑稽。

It really thought, it cared.

它真的想過，它真的在乎。

And so we fixed that.

是以，我們解決了這個問題。

O1 will now think much more intelligently.

現在，O1 可以更聰明地思考問題了。

If you ask it a simple question, it'll respond really quickly.

如果你問它一個簡單的問題，它會很快回答。

And if you ask it a really hard question, it'll think for a really long time.

如果你問它一個非常難的問題，它會想很久。

We ran a pretty detailed suite of human evaluations for this model, and what we found was that it made major mistakes about 34% less often than O1 Preview while thinking fully about 50% faster.

我們對該模型進行了一套相當詳細的人工評估，結果發現，該模型犯重大錯誤的頻率比 O1 預覽少了約 34%，而完全思考的速度卻快了約 50%。

And we think this will be a really, really noticeable difference for all of you.

我們認為，這將給大家帶來非常、非常明顯的變化。

So I really enjoy just talking to these models.

所以我非常喜歡和這些模特哈拉。

I'm a big history buff, and I'll show you a really quick demo of, for example, a sort of question that I might ask one of these models.

我是個歷史迷，我給你們演示一個非常快的例子，比如，我可能會向這些模型中的一個提問。

So right here, on the left, I have O1.

所以在這裡，左邊是 O1。

On the right, I have O1 Preview, and I'm just asking it a really simple history question.

右邊是 O1 預覽，我只是問它一個非常簡單的歷史問題。

List the Roman emperors of the 2nd century.

列舉 2 世紀的羅馬皇帝。

Tell me about their dates, what they did.

告訴我他們的約會，他們做了什麼。

Not hard, but, you know, GPT-40 actually gets this wrong a reasonable fraction of the time.

這並不難，但是，你知道，GPT-40 實際上有相當一部分時間都會出錯。

And so I've asked O1 this.

所以我問了 O1 這個問題。

I've asked O1 Preview this.

我已經問過 O1 預覽版了。

I tested this offline a few times, and I found that O1, on average, responded about 60% faster than O1 Preview.

我離線測試了幾次，發現 O1 的響應速度平均比 O1 預覽版快 60%。

This could be a little bit variable, because right now we're in the process of swapping all our GPUs from O1 Preview to O1.

這可能會有些變化，因為我們現在正在將所有 GPU 從 O1 預覽版切換到 O1。

So actually, O1 thought for about 14 seconds.

實際上，O1 想了大約 14 秒鐘。

O1 Preview, still going.

O1 預覽，仍在繼續。

There's a lot of Roman emperors.

有很多羅馬皇帝。

There's a lot of Roman emperors.

有很多羅馬皇帝。

Yeah, 40 actually gets this wrong a lot of the time.

是的，40 其實經常搞錯。

There are a lot of folks who rolled for, like, 6 days, 12 days, a month, and it sometimes forgets those.

有很多人滾動了 6 天、12 天或一個月，它有時會忘記這些。

Can you do them all from memory, including the 6-day people?

你能憑記憶把它們都做出來嗎，包括 "6 天人"？

No.

不

Yep, so here we go.

是的，我們開始吧。

O1 thought for about 14 seconds.

O1 想了大約 14 秒鐘。

O1 Preview thought for about 33 seconds.

O1 預覽版思考了大約 33 秒鐘。

These should both be faster once we finish deploying, but we want this to go live right now.

一旦我們完成部署，這兩項都會更快，但我們希望現在就上線。

Exactly.

沒錯。

So, yeah, we think you'll really enjoy talking to this model.

所以，是的，我們認為你會非常喜歡與這位模特交談。

We found that it gave great responses.

我們發現它給出了很好的迴應。

It thought much faster.

它想得更快。

It should just be a much better user experience for everyone.

這將為每個人帶來更好的用戶體驗。

So one other feature we know that people really wanted for everyday use cases that we've had requested a lot is multimodal inputs and image understanding, and Hyungwon is going to talk about that now.

我們知道，在日常使用案例中，人們非常需要的另一項功能是多模態輸入和影像理解。

Yep.

是的。

To illustrate the multimodal input and reasoning, I created this toy problem with some hand-drawn diagrams and so on.

為了說明多模態輸入和推理，我製作了這個玩具問題，其中包括一些手繪圖表等。

So here it is.

就這樣吧。

It's hard to see, so I already took a photo of this, and so let's look at this photo in a laptop.

很難看到，所以我已經拍了一張照片，讓我們用筆記本電腦看看這張照片。

So once you upload the image into the chat GPT, you can click on it to see the zoomed-in version.

是以，將圖片上傳到哈拉 GPT 後，就可以點擊圖片查看放大版本。

So this is a system of a data center in space.

是以，這是一個太空數據中心繫統。

So maybe in the future we might want to train AI models in the space.

所以，也許將來我們會想在這個空間裡訓練人工智能模型。

I think we should do that, but the power number looks a little low.

我認為我們應該這樣做，但功率數字看起來有點低。

One gigawatt. Okay.

一千兆瓦 好的

But the general idea, I think.

但我想，大致意思是這樣的。

Rookie numbers.

新秀數據

Yeah, rookie numbers.

是啊，菜鳥數據。

Yeah.

是啊

So we have a sun right here taking power on this solar panel, and then there's a small data center here.

是以，我們在這裡有一個太陽，用太陽能電池板發電，然後這裡有一個小型數據中心。

That's exactly what they look like.

這正是他們的樣子。

Yeah.

是啊

GPU racks.

GPU 架。

And then pump.

然後抽水。

Nice pump here.

這泵不錯。

And one interesting thing about operation in space is that on Earth we can do air cooling, water cooling to cool down the GPUs, but in space there's nothing there, so we have to radiate this heat into the deep space.

在太空中運行的一個有趣之處是，在地球上我們可以通過空氣冷卻、水冷卻來冷卻 GPU，但在太空中什麼都沒有，所以我們必須把熱量輻射到深空。

And that's why we need this giant radiator cooling panel.

這就是為什麼我們需要這個巨大的散熱器冷卻板。

And this problem is about finding the lower bound estimate of the cooling panel area required to operate this one gigawatt data center.

而這個問題就是要找出運行這個千兆瓦級數據中心所需的冷卻板面積的下限估計值。

Probably going to be very big.

可能會非常大。

Yeah.

是啊

Let's see how big it is.

讓我們看看它有多大。

Let's see.

讓我們看看

So that's the problem.

這就是問題所在。

I'm going to this prompt, and yeah, this is essentially asking for that.

我要去看這個提示，是的，這基本上就是在要求我這麼做。

So let me hit go, and the model will think for seconds.

所以，讓我按 "開始 "鍵，模特就會思考幾秒鐘。

By the way, most people don't know.

順便說一句，大多數人都不知道。

I've been working with Hyungwon for a long time.

我和 Hyungwon 合作了很長時間。

Hyungwon actually has a Ph.D. in thermodynamics, which is totally unrelated to AI, and you always joke that you haven't been able to use your Ph.D. work in your job until today.

實際上，Hyungwon 擁有熱力學博士學位，這與人工智能完全無關，而你總是開玩笑說，直到今天你還沒能在工作中用到你的博士學位成果。

So you can trust Hyungwon on this analysis.

所以你可以相信 Hyungwon 的分析。

Finally, finally.

終於，終於

Thanks for hyping up.

謝謝你的誇獎。

Now I really have to get this right.

現在，我真的要把這件事做好了。

Okay, so the model finished thinking.

好了，模特思考完畢。

Only 10 seconds.

只有 10 秒鐘。

It's a simple problem.

問題很簡單。

So let's see how the model did it.

讓我們看看模型是如何做到的。

So power input.

輸入功率

So first of all, this one gigawatt, that was only drawn in the paper.

是以，首先，這個千兆瓦只是在文件中畫出來的。

So the model was able to pick that up nicely, and then radiative heat transfer only.

是以，模型能夠很好地捕捉到這一點，然後只進行輻射傳熱。

That's the thing I mentioned.

這就是我提到的事情。

So in space, nothing else, and then some simplifying choices.

是以，在太空中，沒有其他東西，只有一些簡化的選擇。

And one critical thing is that I intentionally made this problem underspecified, meaning that the critical parameter is the temperature of the cooling panel.

最關鍵的一點是，我故意讓這個問題不夠明確，也就是說，關鍵參數是冷卻板的溫度。

I left it out so that we can test out the model's ability to handle ambiguity and so on.

我把它留出來，是為了測試模型處理模糊性等問題的能力。

So the model was able to recognize that this is actually an unspecified but important parameter, and it actually picked the right range of temperature, which is about the room temperature, and with that, it continues to the analysis and does a whole bunch of things and then found out the area, which is 2.42 million square meters.

是以，模型能夠識別這實際上是一個未指定但很重要的參數，它實際上選擇了正確的溫度範圍，也就是室溫，然後繼續進行分析，做了一大堆事情，然後找出了面積，也就是 242 萬平方米。

Just to get a sense of how big this is, this is about 2% of the land area of San Francisco.

為了讓大家瞭解這裡有多大，這裡的面積約佔舊金山陸地面積的 2%。

This is huge.

這是個大問題。

Not bad.

還不錯。

Not bad, yeah.

還不錯

Oh, okay.

哦，好吧

Yeah, so I guess this is reasonable.

是的，所以我想這是合理的。

I'll skip through the rest of the details, but I think the model did a great job making nice, consistent assumptions that make the required area as little as possible.

其他細節我就不多說了，但我認為模型做得很好，做出了很好的、一致的假設，使所需面積儘可能小。

And so, yeah, so this is the demonstration of the multimodal reasoning.

這就是多模態推理的演示。

And this is a simple problem, but O1 is actually very strong, and on standard benchmarks like MMU and MathVista, O1 actually has the state-of-the-art performance.

這是個簡單的問題，但 O1 實際上非常強大，在 MMU 和 MathVista 等標準基準測試中，O1 實際上擁有最先進的性能。

Now Jason will showcase the Pro mode.

現在，傑森將展示專業模式。

Great.

好極了

So I want to give a short demo of ChatGPT-O1 Pro mode.

是以，我想簡單演示一下 ChatGPT-O1 Pro 模式。

People will find O1 Pro mode the most useful for, say, hard math, science, or programming problems.

人們會發現，O1 Pro 模式對數學、科學或編程等難題最有用。

So here I have a pretty challenging chemistry problem that O1 preview gets usually incorrect, and so I will let the model start thinking.

是以，我這裡有一個相當有挑戰性的化學問題，O1 預覽通常會出錯，所以我會讓模型開始思考。

One thing we've learned with these models is that for these very challenging problems, the model can think for up to a few minutes.

我們從這些模型中學到的一點是，對於這些極具挑戰性的問題，模型可以思考長達幾分鐘。

I think for this problem, the model usually thinks anywhere from one minute to up to three minutes.

我認為，對於這個問題，模型通常會思考一分鐘到三分鐘不等。

And so we have to provide some entertainment for people while the model is thinking.

是以，我們必須在模型思考時為人們提供一些娛樂。

So I'll describe the problem a little bit, and then if the model is still thinking when I'm done, I've prepared a dad joke for us to fill the rest of the time.

所以，我會稍微描述一下這個問題，然後如果我說完了，模型還在思考，我就會為我們準備一個爸爸的笑話來填補剩下的時間。

I hope it thinks for a long time.

我希望它能思考很久。

You can see the problem asks for a protein that fits a very specific set of criteria.

你可以看到，問題要求蛋白質符合一套非常具體的標準。

So there are six criteria, and the challenge is each of them asks for pretty chemistry domain-specific knowledge that the model would have to recall.

是以有六項標準，而挑戰在於每項標準都要求模型必須調用特定領域的化學知識。

And the other thing to know about this problem is that none of these criteria actually give away what the correct answer is.

關於這個問題，還需要知道的一點是，這些標準實際上都沒有給出正確答案。

So for any given criteria, there could be dozens of proteins that might fit that criteria, and so the model has to think through all the candidates and then check if they fit all the criteria.

是以，對於任何給定的標準，可能有幾十種蛋白質符合該標準，是以模型必須考慮所有候選蛋白質，然後檢查它們是否符合所有標準。

Okay, so you can see the model actually was faster this time, so it finished in 53 seconds.

好了，你可以看到模型這次實際上更快了，所以它只用了 53 秒就完成了。

You can click and see some of the thought process that the model went through to get the answer.

您可以點擊查看模型得出答案的一些思考過程。

You can see it's thinking about different candidates like neuroligin initially.

你可以看到它正在考慮不同的候選者，比如最初的神經利金。

And then it arrives at the correct answer, which is retinochisin, which is great.

然後它得出了正確答案，即視黃醇，這很好。

Okay, so to summarize, we saw from Max that O1 is smarter and faster than O1 preview.

好了，總結一下，我們從 Max 身上看到了 O1 比 O1 預覽版更聰明、更快。

We saw from Hyungwon that O1 can now reason over both text and images.

我們從 Hyungwon 身上看到，O1 現在可以對文本和影像進行推理。

And then finally, we saw with ChatterBot Pro mode, you can use O1 to reason about the hardest science and math problems.

最後，我們看到在 ChatterBot Pro 模式下，您可以使用 O1 來推理最難的科學和數學問題。

Yep, there's more to come for the ChatterBot Pro tier.

沒錯，ChatterBot Pro 層級還有更多驚喜。

We're working on even more compute-intensive tasks to power longer and bigger tasks for those who want to push the model even further.

我們正在開發更多的計算密集型任務，為那些希望進一步推動該機型的用戶提供更長、更大的任務支持。

And we're still working on adding tools to the O1 model, such as web browsing, file uploads, and things like that.

我們仍在努力為 O1 模型添加工具，如網頁瀏覽、文件上傳等。