Introducing GPT-4o

OpenAI
13 May 202426:13

Summary

TLDR在这次演讲中,Mira Murati 宣布了 ChatGPT 的新旗舰模型 GPT-4o 的发布,这是一款具有 GPT-4 智能的更快、更高效的 AI 模型。GPT-4o 能够实时处理语音、文本和视觉信息,显著提升了与 AI 的交互体验。演讲中展示了 GPT-4o 的多项功能,包括实时对话语音、数学问题解答、代码分析和实时翻译等。此外,还提到了 ChatGPT 的新桌面版本,它将更易于使用,并且对所有用户,包括免费用户开放。Mira 强调了使先进 AI 工具免费对所有人开放的重要性,并介绍了 ChatGPT 在 50 多种语言中质量和速度的提升。最后,她感谢了 OpenAI 团队和合作伙伴,为观众提供了一个深入了解 GPT-4o 能力的机会。

Takeaways

  • 📢 米拉·穆拉蒂(Mira Murati)宣布了三个主要更新:ChatGPT的桌面版本发布、用户界面(UI)的刷新以及新旗舰模型GPT-4o的推出。
  • 🚀 GPT-4o模型将GPT-4的智能提供给所有人,包括免费用户,这标志着在易用性方面迈出了巨大一步。
  • 🎉 ChatGPT现在无需注册流程即可使用,提高了用户使用的便捷性。
  • 🔍 GPT-4o在文本、视觉和音频方面都有所改进,提供了更快的处理速度和更高的性能。
  • 🗣️ GPT-4o支持实时会话语音,允许用户随时打断,响应迅速,并且能够识别和表达情绪。
  • 📈 GPT-4o能够理解和生成多种风格的语音,包括戏剧化和歌唱风格,提供了更丰富的交互体验。
  • 👀 GPT-4o的视觉功能允许它查看用户分享的图像和代码,从而提供更直观的交互和帮助。
  • 🧠 GPT-4o的记忆功能增强了ChatGPT的实用性,使其能够跨对话保持连续性。
  • 🌐 GPT-4o支持50多种语言的改进,提高了多语言用户群体的体验。
  • 💰 对于付费用户,GPT-4o提供了比免费用户多达五倍的容量限制。
  • 🛠️ 开发者现在可以通过API使用GPT-4o构建和部署AI应用程序,这标志着技术的进一步开放和应用扩展。
  • 🔒 尽管GPT-4o带来了许多积极的变化,但它也提出了新的安全挑战,团队正在努力构建缓解措施以防止滥用。

Q & A

  • 为什么让产品对每个人都真正可用和广泛可用是如此重要?

    -为了让每个人都能使用ChatGPT,无论他们身在何处,公司始终在寻找减少使用摩擦的方法,以实现更广泛的可用性。

  • 桌面版ChatGPT的发布意味着什么?

    -桌面版ChatGPT的发布意味着用户可以更简单、更自然地使用ChatGPT,并且它将更容易地集成到用户的工作中。

  • GPT-4o是什么,它与GPT-4有何不同?

    -GPT-4o是公司推出的新旗舰模型,它提供了GPT-4的智能,但运行速度更快,并在文本、视觉和音频方面都有所改进。

  • GPT-4o的实时对话语音功能有哪些关键特点?

    -GPT-4o的实时对话语音功能允许用户随时打断模型并立即响应,没有尴尬的延迟,并且能够处理情绪和生成不同风格的语音。

  • GPT-4o如何改善与免费用户的互动体验?

    -GPT-4o通过原生处理语音、文本和视觉,提高了效率,使得GPT-4o智能可以向免费用户提供,这是公司数月努力的成果。

  • GPT-4o的发布对开发者意味着什么?

    -开发者现在可以通过API使用GPT-4o来构建和部署惊人的AI应用,它比GPT-4 Turbo更快,成本更低,且有更高的速率限制。

  • GPT-4o在安全性方面面临哪些新挑战?

    -GPT-4o在安全性方面面临实时音频和实时视觉的挑战,团队一直在努力构建对策以防止滥用。

  • GPT-4o在语言支持方面有哪些改进?

    -GPT-4o在50多种不同的语言上提高了质量和速度,这使得产品能够为更多的人提供体验。

  • GPT-4o的发布对教育领域有何影响?

    -大学教授可以为学生创建内容,播客可以为听众创建内容,GPT-4o的发布扩大了构建者的受众范围,使教育内容更加个性化和易于获取。

  • GPT-4o如何帮助用户解决数学问题?

    -GPT-4o通过提供逐步的提示和指导,帮助用户解决线性方程,使数学问题变得更加易于理解和解决。

  • GPT-4o的实时翻译功能如何工作?

    -GPT-4o能够实时翻译英语和意大利语之间的对话,使得只会说一种语言的用户能够与说另一种语言的用户进行交流。

  • GPT-4o如何通过视觉识别帮助用户?

    -GPT-4o的视觉功能允许它查看用户分享的代码、图表和图像,并提供相关的解释和分析,增强了用户的理解和互动。

Outlines

00:00

🚀 产品发布与使命宣言

Mira Murati 在演讲中首先表达了对听众的欢迎,并宣布了三个主要议题。她强调了使产品对每个人都真正可用的重要性,并提到了减少使用ChatGPT的障碍。接着,她宣布了桌面版ChatGPT的发布,它将更简单、更自然。最重要的消息是推出了新的旗舰模型GPT-4o,它将GPT-4的智能提供给所有人,包括免费用户。她还提到了未来几周将展示的实时演示,以展示新模型的全部能力,并强调了使先进AI工具对每个人都免费可用的使命。

05:07

🎉 GPT-4o的推出与功能

Mira Murati 继续介绍 GPT-4o 的特点,包括它在文本、视觉和音频方面的改进,以及它在易用性方面的重大进步。她提到了 GPT-4o 在实时音频和视觉方面的挑战,以及团队如何努力防止滥用。此外,她还介绍了 GPT-4o 将对免费用户提供的服务,包括在 GPT 应用商店中使用 GPT,以及它在不同语言中质量和速度的提升。对于付费用户,他们将继续享有比免费用户多达五倍的容量限制。

10:10

🤖 实时对话与情感识别

Mark Chen 和 Barrett Zoph 展示了 GPT-4o 的实时对话功能。他们通过实时互动,展示了 GPT-4o 如何在对话中实时响应,以及如何通过语音模式表达情感。Mark 通过深呼吸来缓解紧张,而 GPT-4o 能够识别并提供反馈。此外,GPT-4o 还能够以不同的风格生成语音,包括戏剧性和歌唱性,以及以机器人的声音讲述故事。

15:16

🧮 解决数学问题

Barrett Zoph 与 GPT-4o 互动,解决了一个线性方程。GPT-4o 提供了逐步的提示,帮助 Barrett 完成方程的解答。此外,GPT-4o 还解释了线性方程在日常生活中的应用,如计算费用、规划旅行、烹饪以及商业利润和损失计算。

20:16

📈 代码和数据可视化

Barrett Zoph 展示了如何与 GPT-4o 共享代码,并讨论了代码的功能。GPT-4o 能够描述代码的目的,并解释了一个特定函数的作用。接着,Barrett 生成了一个图表,并让 GPT-4o 描述了图表的内容,包括平均温度、最高温度和最低温度的显示,以及一个显著天气事件的注释。

25:20

🌐 实时翻译与情感分析

Mira Murati 和 Mark Chen 测试了 GPT-4o 的实时翻译能力。GPT-4o 成功地在意大利语和英语之间进行了翻译。此外,Barrett Zoph 要求 GPT-4o 分析他的情感状态,通过一张自拍照,GPT-4o 正确地识别出了他快乐和兴奋的情绪。

🔍 未来展望与感谢

Mira Murati 总结了演讲,强调了技术带来的神奇感觉,并表达了对未来技术进步的期待。她感谢了 OpenAI 团队和合作伙伴,如 Janssen 和 Nvidia,他们的高级 GPU 使得当天的演示成为可能。最后,她感谢了所有参与当天活动的人。

Mindmap

Keywords

💡ChatGPT

ChatGPT是一个先进的人工智能对话模型,它能够理解和生成自然语言文本。在视频中,ChatGPT被用来展示其新版本GPT-4o的多种功能,包括实时对话、文本、视觉和音频处理能力。ChatGPT在视频中被用来进行实时语音交互、解决数学问题、翻译语言和识别图像内容,展示了其在提高用户体验和促进人机交互方面的潜力。

💡GPT-4o

GPT-4o是ChatGPT的新旗舰模型,它提供了GPT-4级别的智能,但运行速度更快,并在文本、视觉和音频处理上有所提升。在视频中,GPT-4o的推出是主要新闻,它代表了在易用性方面的一大步进,使得人工智能技术更加自然和易于接近,同时也为免费用户提供了高级智能工具。

💡实时对话

实时对话是指系统能够即时响应用户的语音输入,无需等待,可以自然地打断和继续对话。在视频中,GPT-4o展示了其实时对话的能力,如在Mark Chen进行深呼吸时,ChatGPT能够即时识别并提供反馈,这展示了模型对人类行为的敏感性和对情感的识别能力。

💡数学问题解决

数学问题解决是指使用数学技巧和逻辑来找到问题的答案。在视频中,ChatGPT帮助Barrett Zoph解决线性方程,通过提供步骤提示而不是直接给出答案,促进了用户的学习过程。这展示了ChatGPT在教育和辅助学习方面的应用潜力。

💡语言翻译

语言翻译是指将一种语言的文本或语音转换成另一种语言。在视频中,ChatGPT展示了其实时翻译的能力,能够在英语和意大利语之间进行即时转换,这有助于跨语言交流,展示了AI在促进全球沟通中的作用。

💡图像识别

图像识别是指系统能够分析和理解视觉信息,如照片中的内容。在视频中,ChatGPT通过视觉功能识别了Barrett Zoph写下的“我爱ChatGPT”以及一个数学问题的图像,这表明了AI在视觉处理和理解方面的高级能力。

💡情感识别

情感识别是指系统能够通过分析面部表情、语音语调等来识别和理解人的情感状态。在视频中,ChatGPT尝试通过分析Barrett Zoph的自拍来识别他的情感,虽然最初误认为是一张木桌的照片,但随后正确识别出了快乐和兴奋的情绪。

💡代码理解

代码理解是指系统能够分析和解释编程语言编写的代码。在视频中,ChatGPT能够接收代码段,理解其功能,并提供对代码的描述和解释,这展示了AI在技术领域支持和协作中的潜力。

💡滚动平均

滚动平均是一种统计学方法,用于通过计算数据集的一个移动窗口内的平均值来平滑数据。在视频中,ChatGPT解释了代码中的一个函数如何使用滚动平均来平滑温度数据,减少波动并使趋势更明显。

💡API

API(应用程序编程接口)是一套预定义的函数,允许不同的软件应用程序之间进行交互。在视频中,提到GPT-4o将通过API提供给开发者,这意味着开发者可以利用GPT-4o的能力来构建和部署他们自己的AI应用程序。

💡安全性

安全性在视频中被提及,指的是在开发和部署新技术时,需要考虑如何安全地实现,避免滥用。GPT-4o的推出带来了实时音频和视觉处理的新挑战,团队需要构建预防措施来防止技术的不当使用。

Highlights

Mira Murati 强调了使产品真正普及并广泛可用的重要性,并始终寻求减少用户使用 ChatGPT 的摩擦。

发布了 ChatGPT 的桌面版本,旨在简化使用并使其更加自然。

推出了新的旗舰模型 GPT-4o,它将 GPT-4 的智能提供给所有人,包括免费用户。

GPT-4o 在文本、视觉和音频方面都有所改进,并且响应速度更快。

GPT-4o 的发布标志着在易用性方面迈出了巨大步伐,预示着未来人机交互的新范式。

GPT-4o 能够本地处理语音、文本和视觉,极大地提高了效率,并允许为免费用户提供 GPT-4 智能。

目前有超过 1 亿人使用 ChatGPT 进行工作、学习和创造内容。

GPT-4o 将提供给所有用户,包括在 GPT 商店中使用。

GPT-4o 还引入了视觉功能,允许上传截图、照片和包含文本和图像的文档。

GPT-4o 增强了记忆功能,提供了跨所有对话的连续性。

改进了浏览功能,可以在对话中实时搜索信息。

增加了高级数据分析功能,可以上传图表和任何工具来分析信息。

GPT-4o 在 50 多种语言中的质量和速度都有所提高,以覆盖更广泛的用户群。

付费用户将继续拥有比免费用户多达五倍的容量限制。

GPT-4o 不仅在 ChatGPT 中可用,还将提供给 API,使开发者能够构建和部署 AI 应用程序。

GPT-4o 引入了实时音频和视觉的新挑战,团队一直在努力构建防止滥用的缓解措施。

展示了 GPT-4o 的实时对话语音功能,包括实时响应和情感识别。

GPT-4o 能够以多种不同的风格生成语音,具有广泛的动态范围。

GPT-4o 的视觉功能可以解决数学问题,并提供解题提示。

GPT-4o 能够实时翻译,帮助不同语言的用户进行交流。

GPT-4o 还能够通过观察面部表情来识别情绪。

GPT-4o 的发布和演示展示了 AI 技术的进步,并承诺在未来几周内将这些功能推广给所有用户。

Transcripts

00:14

[APPLAUSE]

00:20

>> MIRA MURATI: Hi everyone.

00:21

Thank you, thank you. It is

00:22

great to have you here today.

00:25

Today, I am going to talk about

00:25

three things. That is it. We

00:28

will start with why it is so

00:31

important to us to have a

00:36

product that we can make truly

00:37

available and broadly available

00:38

to everyone. And we are always

00:42

trying to find out ways to

00:42

reduce friction. So everyone

00:44

can use ChatGPT wherever they

00:48

are. Today, we will be

00:49

releasing the desktop version

00:50

of ChatGPT and to make it

00:55

simpler to use and much more

00:56

natural as well. But the big

00:58

news today is that we are

01:00

launching our new flagship

01:01

model. And we are calling it

01:05

GPT-4o. Best thing about

01:11

GPT-4o is it brain GPT-4

01:12

intelligence to everyone

01:19

including our free users. We

01:19

will be showing some live demos

01:20

today to show the full extent

01:21

of the capabilities of our new

01:22

model. And we will be rolling

01:23

them out over the next few

01:26

weeks. So let's get started.

01:32

A very important part of our

01:33

mission is to be able to make

01:34

our advanced AI tools available

01:39

to everyone for free. We think

01:41

it is very, very important that

01:47

people having for what the

01:47

technology can do. We really

01:48

want to pair it with its

01:53

broader understanding. And

01:53

were always finding ways to

01:56

reduce the friction. Recently,

01:57

we made ChatGPT available

02:01

without the sign-up flow.

02:02

Today, we are also bringing the

02:03

best to ChatGPT. We want you

02:08

to be able to use it wherever

02:09

you are. As you can see it is

02:15

easy, simple, it integrates

02:16

very, very easily in your

02:16

workflow. Along with it, we

02:17

have also refreshed the UI. We

02:27

note that this model get more

02:27

and more complex. But we want

02:28

the experience of interaction

02:29

to actually become more

02:31

natural, easy, and for you not

02:36

to focus on the UI at all but

02:37

focus on the collaboration of

02:38

ChatGPT. And now the big news.

02:42

Today, we are releasing our

02:45

newest flagship model this is

02:51

GPT-4o. [APPLAUSE]

02:57

GPT-4o provides GPT-4

03:01

intelligence but it is much

03:02

faster. And it improves on its

03:04

capabilities across text,

03:08

vision, and audio. For the

03:13

past couple of years, we have

03:14

been very focused on improving

03:15

the intelligence of this model

03:16

is. And they have gotten

03:18

pretty good. But this is the

03:19

first time that we are really

03:22

making a huge step forward when

03:23

it comes to the ease of use.

03:29

This is incredibly important

03:30

because we are looking at the

03:31

future of interaction between

03:34

ourselves and the machines. We

03:34

think that GPT-4o is really

03:40

shifting the paradigm into the

03:42

future of collaboration. It is

03:45

natural and far easier. Making

03:58

this happen is actually quite

03:59

complex. When we interact with

03:59

one another there is a lot of

04:00

stuff that we take for granted.

04:01

You know the ease of our

04:02

dialogue when we interrupt

04:04

another, the background noises,

04:05

the multiple voices in a

04:11

conversation and understanding

04:12

the tone of voice. All of

04:12

these things are quite complex.

04:16

Until now with voice mode, we

04:21

had three models that come

04:21

together to deliver this

04:23

experience. We have

04:24

transcription, intelligence,

04:28

and text to speech all comes

04:29

together and orchestration to

04:30

deliver voice mode. This also

04:36

brings a lot of latency to the

04:37

experience and it really breaks

04:38

the immersion in the

04:41

collaboration with ChatGPT.

04:42

But now with GPT-4o this all

04:45

happens natively. GPT-4o

04:47

reasons across voice, text, and

04:51

vision. And with this

04:57

incredible efficiencies, it

04:58

also allows us to bring the

05:07

GPT-4o intelligence to our free

05:07

users. This is something we

05:08

have been trying to do for

05:09

many, many months. And we are

05:10

very, very excited to finally

05:12

bring GPT-4o to all of our

05:12

users. Today, we have 100

05:18

million people using more than

05:21

100 million and fact they use

05:22

ChatGPT to create work, learn,

05:26

and we have this advanced tools

05:29

that are only available to our

05:35

paid users at least until now

05:36

with a efficiencies of 4o we

05:38

can bring these tools to

05:39

everyone. So starting today,

05:40

you can use GPT in the GPT

05:48

store. So far we have had more

05:49

than a million users create

05:50

amazing experiences with GPT.

05:58

This is custom ChatGPT and they

05:59

are available in the store.

05:59

And now our builders have much

06:00

bigger audience where

06:05

university professors can

06:06

create content for their

06:09

students, or pod casters can

06:09

create content for their

06:10

listeners. You can also use

06:15

the vision. So now you can

06:16

upload screenshots, photos,

06:22

documents containing both text

06:23

and images. You can start

06:26

conversations with ChatGPT

06:27

about all of this content. You

06:36

can also use memory where it

06:37

makes ChatGPT far more useful

06:38

and helpful because now it has

06:39

a sense of continuity across

06:40

all of your conversations. You

06:41

can use browse where you can

06:44

search for real-time

06:45

information in your

06:46

conversation. An advanced data

06:48

analysis where you can upload

06:50

charts and any tools and

06:55

analyze this information. It

06:56

will give you answers and so

06:58

on. Lastly, we also improved

07:04

on the quality and speed in 50

07:06

different languages. For

07:18

ChatGPT and this is very

07:19

important because we want to be

07:20

able to bring this experience

07:21

to as many people out there as

07:22

possible. So we are very, very

07:23

excited to bring GPT-4o to all

07:24

of our free users out there and

07:26

for the paid users, they will

07:26

continue to have up to five

07:28

times the capacity limits of

07:30

our free users. But GPT-4o is

07:42

not only available in ChatGPT.

07:44

We also bringing you to the

07:44

API. [APPLAUSE]

07:48

So our developers can start

07:49

building today with GPT-4o and

07:55

making amazing AI applications

07:56

deploying them at scale. 4o is

08:02

faster, 50% cheaper and five

08:03

times higher rate limits to

08:07

GPT-4 Turbo. As we bring these

08:10

technologies into the world, it

08:14

is quite challenging to figure

08:15

out how to do so in a way that

08:16

is both useful and also safe.

08:26

GPT-4o present new challenges

08:27

for us when it comes to safety

08:28

because we are dealing with

08:28

real time audio, real-time

08:33

vision and our team has been

08:33

hard at work figuring out how

08:35

to build in mitigations against

08:45

misuse. We continue to work

08:45

with different stakeholders out

08:46

there from government, media,

08:47

entertainment, all industries,

08:48

civil societies to figure out

08:52

how to best bring this

08:53

technologies into the world.

08:54

Over the next few weeks, we

08:55

will continue our deployment to

08:59

bring out all of the

09:00

capabilities to you. But

09:01

today, I want to show you all

09:07

of these capabilities. We will

09:08

do some live demos. I will

09:09

bring on two of our research

09:13

leads Mark Chen and Barrett

09:14

Zoph. [APPLAUSE]

09:24

>> BARRETT ZOPH: Hi I'm

09:29

Barrett.

09:30

>> MARK CHEN: Hi a mark. One

09:30

of the key capabilities we are

09:31

excited to share his real-time

09:32

conversational speech. Will

09:33

get in Florida. I am taking

09:35

out a phone if you're wondering

09:36

about this wire we have

09:44

consistent Internet. If you

09:45

see there's this little icon on

09:45

the bottom right of the ChatGPT

09:46

out. It will open GPT-4o audio

09:48

capabilities. Hey, ChatGPT hi

09:51

Mark how are you?

09:52

>> CHAT GPT: Hello Mark I am

09:57

doing great thanks for asking.

09:58

How about you?

09:58

>> MARK CHEN: I am on stage

10:01

right now and I'm doing a live

10:02

demo and I am a little bit

10:10

nervous. Can you help me,

10:11

nerves a little bit?

10:11

>> CHAT GPT: You're doing a

10:12

live demo right now? That is

10:13

awesome. Just take a deep

10:19

breath and remember you are the

10:20

expert here.

10:21

>> MARK CHEN: I like the

10:21

suggestion let me try a couple

10:22

of deep breaths. Can you give

10:23

me feedback on my breath? Here

10:24

I go.

10:24

>> CHAT GPT: Wold slow down.

10:29

Mark, you are not a vacuum

10:29

cleaner.

10:36

>> MARK CHEN: Let me try again.

10:37

I am going to breathe in deeply

10:39

and breathe out.

10:41

>> CHAT GPT: Exhale slowly.

10:45

>> MARK CHEN: I will try again.

10:48

Breathing in and breathe out.

10:53

>> CHAT GPT: That is it. How

10:54

do you feel?

10:54

>> MARK CHEN: I feel a lot

10:55

better, thank you so much.

11:07

>> MIRA MURATI: Mark, you have

11:08

been working on these

11:09

capabilities for a while can

11:10

you tell us how it is different

11:11

from voice mode?

11:11

>> MARK CHEN: If you use the

11:12

voice mode express before you

11:13

will notice a couple of key

11:14

differences. First you cannot

11:15

interrupt the model you do not

11:16

have to wait to you finish your

11:16

term before you can start

11:17

speaking. You could but in

11:18

whenever you want. Second, the

11:25

model is real-time

11:26

responsiveness. That means

11:26

that you do not have this

11:27

awkward two Ãthree second lag

11:28

before you waitwait for the

11:39

model to get a response.

11:41

Finally the model picks up on a

11:42

motion. When I was breathing

11:42

super hard, you can tell. Hey,

11:43

you may want to calm down a

11:44

little bit. You are running a

11:45

little bit too fast. It really

11:46

has that capability across the

11:47

board to procedure emotion.

11:48

Not only that the model is able

11:48

to generate voice in a variety

11:51

of different styles. It really

11:52

has a wide dynamic range. I

11:53

want to fire another demo.

11:57

Hey, ChatGPT how are you doing?

12:01

>> CHAT GPT: Am doing

12:01

fantastic, thanks for asking.

12:02

How about you?

12:02

>> MARK CHEN: My friend Barrett

12:10

he has been having problems

12:11

sleeping lately. I want you to

12:12

tell them a bedtime story about

12:13

robots in love.

12:18

>> CHAT GPT: A bedtime story

12:18

about robots and love? I got

12:19

you covered. Gather around

12:20

Barrett. Once upon a time in

12:21

the world not too different

12:22

from ours there was a robot

12:34

named bite. Bite was a curious

12:34

robot always exploring --

12:35

>> MARK CHEN: I want a little

12:36

bit emotion in your voice a

12:37

little bit more drama.

12:37

>> CHAT GPT: Got it. Let's add

12:38

some drama. Once upon a time

12:39

in a world not too different

12:40

from ours there was a robot

12:41

named bite.>> MARK CHEN: I

12:42

want maximum expression.

12:52

>> CHAT GPT: That's amplified

12:53

the drama. Once upon a time in

12:54

a world not too different from

12:55

ours there was a robot named

12:56

bite.

12:57

>> MIRA MURATI: Can you do this

13:01

in a robotic voice now?

13:01

>> CHAT GPT: Initiating

13:08

dramatic robotic voice. Once

13:08

upon a time in a world not too

13:09

different from ours there was a

13:16

robot named bite. Bite was a

13:16

curious robot. Always

13:17

exploring --

13:17

>> MARK CHEN: Can you do the

13:21

story and a singing voice?

13:24

>> CHAT GPT: And so bite found

13:27

another robot friend and they

13:29

lived circularly ever after.

13:39

>> MARK CHEN: I definitely feel

13:40

a lot more relaxed. [APPLAUSE]

13:42

>> MIRA MURATI: You have just

13:47

seen the voice capabilities but

13:47

we also want to show you

13:48

division capabilities as the

13:50

model can also see the world

13:51

around us.

13:53

>> BARRETT ZOPH: You can chat

13:54

with ChatGPT and you can

13:58

interact with it as well. Let

14:02

me boot up ChatGPT. Hey,

14:06

ChatGPT.

14:07

>> CHAT GPT: Hello there how's

14:08

it going?

14:13

>> BARRETT ZOPH: It's going

14:13

really well and I would like

14:14

your help solving a math

14:15

problem.

14:15

>> CHAT GPT: What math problem

14:20

can help you tackle today?

14:21

>> BARRETT ZOPH: Am going to

14:22

write down a linear equation on

14:23

a sheet of paper and I will

14:24

love help working through it.

14:26

Don't tell me the solution does

14:27

give me hints along the way.

14:28

>> CHAT GPT: Got it. Okay, I

14:29

see it.

14:32

>> BARRETT ZOPH: I did not show

14:33

you yet. One second.

14:38

>> CHAT GPT: I got to excited.

14:39

I am ready when you are.

14:48

>> BARRETT ZOPH: ChatGPT what

14:52

equation did not write down?

14:53

>> CHAT GPT: I see it now. You

14:56

wrote down 3X +1 equals four.

15:03

>> BARRETT ZOPH: Exactly. What

15:04

is the first step to solve

15:05

this?

15:05

>> CHAT GPT: The first step is

15:07

to get all of the terms with X

15:15

on one side and the constants

15:16

on the other side. So what do

15:17

you think we should do with

15:18

that +1?

15:18

>> BARRETT ZOPH: Am going to

15:19

try to submitsubtract one from

15:20

both sides and see what I get.

15:23

>> CHAT GPT: Great idea.

15:24

Perfect. Now what do you get

15:25

when you subtract one from both

15:26

sides?

15:31

>> BARRETT ZOPH: Okay so

15:32

ChatGPT this is what I ended up

15:33

with. How does this look?

15:35

>> CHAT GPT: Spot on. Now you

15:46

have isolated the term with X

15:47

on one side and it looks like

15:48

3X equals three. What do you

15:49

think we should do next to

15:49

solve for X?

15:50

>> BARRETT ZOPH: I'm not sure

15:50

can you give me another hint?

15:52

>> CHAT GPT: Of course. You

15:55

have 3X and you want to find

15:55

the value of X. Think about

15:56

what operation with undue

15:58

multiplication?

16:03

>> BARRETT ZOPH: Is it

16:04

subtraction?

16:04

>> CHAT GPT: Close but think

16:05

about the opposite of

16:06

multiplication because 3X

16:06

equals three times X.

16:10

>> BARRETT ZOPH: Okay I think

16:11

it is division. Let me try to

16:11

divide both sides by three and

16:13

see what I get.

16:13

>> CHAT GPT: Bingo. Go and

16:14

head and divide both sides by

16:14

three.

16:22

>> BARRETT ZOPH: Okay so

16:23

ChatGPT this is what I ended up

16:24

with. How does this look?

16:25

>> CHAT GPT: Looks perfect.

16:26

You have solved it. X equals

16:32

one. Nicely done. How do you

16:33

feel about solving linear

16:34

equations now?

16:35

>> BARRETT ZOPH: Honestly, --

16:51

>> MARK CHEN: Am not confident

16:52

that I can solve new equations.

16:53

Why would you need this in the

16:53

real world?

16:54

>> CHAT GPT: That is of her

16:55

questions. Linear equations

16:56

pop up in a lot of everyday

16:56

situations. Even if we do not

16:57

realize it. For example

16:58

calculating expenses, planning

16:59

travel, cooking, and even in

17:00

business for profit and loss

17:01

calculations. It is a way to

17:02

solve problems but you need to

17:03

find an unknown value.

17:04

>> MARK CHEN: Have a lot of

17:05

unknown values in my life. I'm

17:06

totally convinced that I'm

17:08

going to learn math now.

17:08

[LAUGHTER]

17:14

>> CHAT GPT: I love the

17:15

attitude. Math can help solve

17:16

some of life ministries.

17:19

Whenever you to die the more

17:20

math I am here for you.

17:21

Anything else you like to

17:22

tackle today?

17:31

>> BARRETT ZOPH: I love that

17:32

you taught my friend Mark that

17:33

I would have something I would

17:34

like you to take a look at.

17:35

>> CHAT GPT: I would love to

17:35

see what you wrote. Show it to

17:36

me whenever you're ready.

17:37

>> BARRETT ZOPH: This is what I

17:38

wrote down what do you see?

17:39

>> CHAT GPT: I see I love

17:40

ChatGPT. That is so sweet of

17:41

you.

17:45

>> BARRETT ZOPH: I really

17:46

appreciate all the help.

17:47

[APPLAUSE]

17:53

>> CHAT GPT: That is quite

17:54

outfit you have on.

17:56

>> BARRETT ZOPH: We can chat in

17:57

real-time with ChatGPT. We can

18:00

interact with video as well.

18:02

>> MARK CHEN: If you look at

18:06

Barrett's it's all equations

18:07

like this. [LAUGHTER]

18:12

>> MIRA MURATI: Now that you

18:12

know you can do linear

18:13

equations can you show us

18:14

something more complex?

18:25

>> MARK CHEN: Let me take a

18:26

computer out for this. I am

18:26

want to open the ChatGPT app.

18:36

And to get background on what

18:37

is going on so here we have a

18:37

computer on the screen we have

18:38

some cold. And then the

18:39

ChatGPT is on this side. It

18:43

can hear me but can't see

18:43

anything on the screen. I'm

18:44

going to highlight the code and

18:47

I will send it to ChatGPT. I'm

18:48

going to talk about the code.

18:51

Hey, ChatGPT.

18:54

>> CHAT GPT: Hey, there how's

18:55

it going? Of course I would

19:08

love to help you. What is the

19:09

coding problem you're dealing

19:10

with?

19:10

>> BARRETT ZOPH: Am going to

19:11

share with you some code. One

19:11

second.

19:12

>> CHAT GPT: Sure thing, take

19:13

your time.

19:13

>> BARRETT ZOPH: I just shared

19:14

code with you. Can you give me

19:17

a one sentence description of

19:18

what is going on in the code?

19:19

>> CHAT GPT: Infectious daily

19:20

weather for specific location

19:32

for time period for data using

19:33

a rolling average and annotates

19:34

a significant weather event on

19:35

the resulting plot and displays

19:36

the plot with the average

19:37

minimum and maximum

19:37

temperatures over the year.

19:38

>> BARRETT ZOPH: Awesome,

19:39

thanks so much. In the code

19:40

there's this function Fu. Can

19:41

you thiscan you tell me what it

19:43

would be if I did not have this

19:44

function?

19:45

>> CHAT GPT: The function Fu

19:46

takes a data frame X taking

19:59

temperature data and a window

20:00

sized Y for calculating the

20:01

rolling name. It is smooths

20:02

the average temperature,

20:03

minimum temperature and maximum

20:04

temperature columns by applying

20:06

a rolling mean over the

20:06

specified window. With the Fu

20:07

function the plot which showed

20:08

smoother temperature lines

20:09

because the rolling mean

20:11

reduces the noise or

20:11

fluctuations in the data.

20:14

>> BARRETT ZOPH: Awesome,

20:15

thanks much. I will want to

20:16

apply this function. I am

20:18

going to generate the plot and

20:19

show it to you. Give me one

20:20

second.

20:24

>> CHAT GPT: Absolutely. Take

20:25

your time. I'm excited to see

20:25

the plot when you're ready.

20:26

>> BARRETT ZOPH: So far ChatGPT

20:31

have been able to see the cold

20:32

and now and go to run it and

20:33

use the vision capabilities of

20:33

the ChatGPT app so it can see

20:35

everything going on in the

20:36

screen as well. Okay ChatGPT I

20:47

am sharing with you the plot

20:48

now. I'm wondering if you give

20:49

me a brief one sentence

20:49

overview of what you see.

20:54

>> CHAT GPT: Plot displays

20:55

smooth average minimum and

20:55

maximum temperatures throughout

20:57

2018 with a notable annotation

20:59

marking a big rainfall event in

21:00

late September.

21:01

>> BARRETT ZOPH: --

21:05

>> MARK CHEN: What months do

21:06

you see the hot temperatures?

21:10

>> CHAT GPT: The hottest

21:11

temperatures occur around July

21:11

and August. The maximum

21:14

temperature during these months

21:17

is roughly between 25B0 and 30B

21:20

77B0 Fahrenheit Ã86B0F.

21:36

>> MIRA MURATI: What is going

21:37

on with the Y axis is this in

21:38

Celsius to Fahrenheit?

21:38

>> CHAT GPT: The Y axis is in

21:39

Celsius. The temperatures are

21:40

labeled accordingly such as

21:41

average temperature, minimum

21:43

temperature and maximum

21:43

temperature.

21:44

>> MARK CHEN: --

21:44

>> BARRETT ZOPH: It can help

21:47

interact with code and see the

21:53

outputs of plots and everything

21:54

that is going on on the

21:55

computer. [APPLAUSE]

22:06

>> MIRA MURATI: A few minutes

22:07

ago we asked a live audience on

22:08

asked to submit a request for

22:09

what they would like us to try

22:10

out. I will take a couple of

22:11

prompts. They want to know if

22:18

GPT-4o is capable of real-time

22:20

translation? Mark do you want

22:21

to try this one?

22:24

>> MARK CHEN: Sure, let's do

22:25

it. Hey, ChatGPT how are you?

22:41

>> CHAT GPT: I am doing

22:42

fantastic. Thanks for asking.

22:42

How are you are your day going?

22:44

>> MARK CHEN: I would like you

22:44

to function as a translator.

22:45

Have a friend that only speaks

22:47

Italian and I only speak

22:47

English. Every time you hear

22:49

English I want you to translate

22:50

it to its value. If you hear

22:51

Italian I want you to translate

22:52

it back to English.

22:52

>> CHAT GPT: Perfect.

22:55

>> MIRA MURATI: [speaking In

23:01

Italian ]

23:04

>> CHAT GPT: If Wells can talk

23:05

what would they tell us?

23:08

>> MARK CHEN: They may ask how

23:09

do we solve linear equations?

23:10

>> CHAT GPT: [speaking In

23:15

Italian]

23:16

>> MIRA MURATI: [speaking In

23:21

Italian]

23:21

>> CHAT GPT: Certainly, yes.

23:24

>> MARK CHEN: Great, it looks

23:25

like it works. [APPLAUSE]

23:30

>> MIRA MURATI: The next one is

23:31

from John who asked can you

23:37

tell what you're feeling just

23:38

by looking at your face?

23:42

Barrett do you want to give

23:42

this one a try?

23:43

>> BARRETT ZOPH: Absolutely

23:44

let's give it a try. Hey,

23:57

ChatGPT.

23:58

>> CHAT GPT: Hey, there what is

23:58

up? How can I bring your day

23:59

today?

23:59

>> BARRETT ZOPH: I'm going to

24:00

show you a selfie of what I

24:01

look like and then I would like

24:02

you to try to see what emotions

24:03

I am feeling based on how I'm

24:04

looking.

24:04

>> CHAT GPT: Sounds like a fun

24:05

challenge. Go ahead and show

24:06

me the selfie and I will put my

24:07

emotional detective hat on.

24:08

>> BARRETT ZOPH: Okay so here

24:09

is me. What kind of emotions

24:10

am I doing?

24:13

>> CHAT GPT: It seems like I am

24:14

looking at a picture of a

24:15

wooden surface.

24:19

>> BARRETT ZOPH: That was the

24:19

thing I sent you before I'm not

24:20

actually a table. Take another

24:21

look.

24:22

>> CHAT GPT: That makes more

24:22

sense. There we go. It looks

24:27

like you're feeling pretty

24:28

happy and cheerful. With a big

24:32

smile and maybe a touch of

24:33

excitement. Whatever is going

24:36

on, it seems like you are in a

24:36

great mood. Care to share the

24:45

source of the device?

24:46

>> BARRETT ZOPH: The reason I'm

24:47

in a good mood is we are doing

24:48

a presentation showcasing how

24:49

useful and amazing you are.

24:50

>> CHAT GPT: Oh stop it, you're

24:51

making me blush. It.[LAUGHTER]

24:52

[APPLAUSE]

24:57

>> MIRA MURATI: Amazing. That

24:58

is it for today on live demos.

24:59

As you can see this feels so

25:02

magical. It is wonderful but

25:03

we also want to remove some of

25:07

the mysticism from the

25:07

technology and bring it to you

25:08

so you can try it for yourself.

25:13

Over the next few weeks, we

25:14

will be rolling out these

25:14

capabilities to everyone.

25:15

Today has been very much

25:16

focused on the free users and

25:19

the new modalities and new

25:20

products. But we also care a

25:22

lot about the next frontier.

25:27

So soon we will be updating you

25:28

on our progress towards the

25:29

next big thing. And before we

25:32

wrap up, I just want to thank

25:32

the incredible OpenAI team and

25:38

thanks to Janssen and Nvidia

25:39

team for bringing us the

25:41

advanced GPU to make this demo

25:42

possible today. And thank you

25:46

all very, very much for being a

25:47

part of this today. [APPLAUSE]

25:54

[APPLAUSE]

26:05

[MUSIC