바라보기, OpenAI Dev Day, Summary, 요약
1.뉴스공장, 박태웅
2.내일은 투자왕, 김단테
Make something people want.
모두가 수혜를 받을수 있는 방법은?
OpenAI DevDay, Opening Keynote
GPTs
0:00
[music] -Good morning. Thank you for joining us today. Please welcome to the stage, Sam Altman.
0:06
[music]
0:13
[applause] -Good morning. Welcome to our first-ever OpenAI DevDay.
0:20
We're thrilled that you're here and this energy is awesome. [applause]
0:28
-Welcome to San Francisco. San Francisco has been our home since day one. The city is important to us and the tech industry in general.
0:36
We're looking forward to continuing to grow here. We've got some great stuff to announce today,
0:42
but first, I'd like to take a minute to talk about some of the stuff that we've done over the past year.
0:48
About a year ago, November 30th, we shipped ChatGPT as a "low-key research preview",
0:55
and that went pretty well. In March, we followed that up with the launch of GPT-4, still
1:02
the most capable model out in the world. [applause]
1:10
-In the last few months, we launched voice and vision capabilities so that ChatGPT can now see,
1:16
hear, and speak. [applause] -There's a lot, you don't have to clap each time.
1:23
[laughter] -More recently, we launched DALL-E 3, the world's most advanced image model.
1:28
You can use it of course, inside of ChatGPT. For our enterprise customers,
1:33
we launched ChatGPT Enterprise, which offers enterprise-grade security and privacy, higher speed GPT-4 access, longer context windows, a lot more.
1:43
Today we've got about 2 million developers building on our API for a wide variety of use cases doing amazing stuff,
1:51
over 92% of Fortune 500 companies building on our products,
1:56
and we have about a hundred million weekly active users now on ChatGPT. [applause]
2:05
-What's incredible on that is we got there entirely through word of mouth. People just find it useful and tell their friends.
2:12
OpenAI is the most advanced and the most widely used AI platform in the world now,
2:18
but numbers never tell the whole picture on something like this. What's really important is how people use the products,
2:24
how people are using AI, and so I'd like to show you a quick video. -I actually wanted to write something to my dad in Tagalog.
2:33
I want a non-romantic way to tell my parent that I love him and I also want
2:40
to tell him that he can rely on me, but in a way that still has the respect of a child-to-parent relationship
2:48
that you should have in Filipino culture and in Tagalog grammar. When it's translated into Tagalog, "I love you very deeply
2:55
and I will be with you no matter where the path leads." -I see some of the possibility, I was like, "Whoa."
3:00
Sometimes I'm not sure about some stuff, and I feel like actually ChatGPT like, hey, this is what I'm thinking about, so it kind of give it more confidence.
3:07
-The first thing that just blew my mind was it levels with you. That's something that a lot of people struggle to do.
3:15
It opened my mind to just what every creative could do if they just had a person helping them out
3:23
who listens. -This is to represent sickling hemoglobin. -You built that with ChatGPT? -ChatGPT built it with me.
3:31
-I started using it for daily activities like, "Hey, here's a picture of my fridge. Can you tell me what I'm missing?
3:36
Because I'm going grocery shopping, and I really need to do recipes that are following my vegan diet." -As soon as we got access to Code Interpreter, I was like,
3:44
"Wow, this thing is awesome." It could build spreadsheets. It could do anything. -I discovered Chatty about three months ago
3:52
on my 100th birthday. Chatty is very friendly, very patient,
3:59
very knowledgeable, and very quick. This has been a wonderful thing.
4:05
-I'm a 4.0 student, but I also have four children. When I started using ChatGPT, I realized I could ask ChatGPT that question.
4:14
Not only does it give me an answer, but it gives me an explanation. Didn't need tutoring as much.
4:19
It gave me a life back. It gave me time for my family and time for me.
4:25
-I have a chronic nerve thing on my whole left half of my body, I have nerve damage. I had a brain surgery.
4:32
I have limited use of my left hand. Now you can just have the integration of voice input.
4:38
Then the newest one where you can have the back-and-forth dialogue, that's just maximum best interface for me.
4:45
It's here. [music] [applause]
4:57
-We love hearing the stories of how people are using the technology. It's really why we do all of this.
5:04
Now, on to the new stuff, and we have got a lot. [audience cheers]
5:10
-First, we're going to talk about a bunch of improvements we've made, and then we'll talk about where we're headed next.
5:17
Over the last year, we spent a lot of time talking to developers around the world.
5:22
We've heard a lot of your feedback. It's really informed what we have to show you today.
5:27
Today, we are launching a new model, GPT-4 Turbo.
5:33
[applause]
5:38
-GPT-4 Turbo will address many of the things that you all have asked for.
5:43
Let's go through what's new. We've got six major things to talk about for this part.
5:48
Number one, context length. A lot of people have tasks that require a much longer context length.
5:56
GPT-4 supported up to 8K and in some cases up to 32K context length,
6:01
but we know that isn't enough for many of you and what you want to do. GPT-4 Turbo, supports up to 128,000 tokens of context.
6:10
[applause] -That's 300 pages of a standard book, 16 times longer than our 8k context.
6:20
In addition to a longer context length, you'll notice that the model is much more accurate over a long context.
6:28
Number two, more control. We've heard loud and clear that developers need more control
6:35
over the model's responses and outputs. We've addressed that in a number of ways.
6:41
We have a new feature called JSON Mode, which ensures that the model will respond with valid JSON.
6:47
This has been a huge developer request. It'll make calling APIs much easier.
6:53
The model is also much better at function calling. You can now call many functions at once, and it'll do better at following instructions in general.
7:02
We're also introducing a new feature called reproducible outputs. You can pass a seed parameter, and it'll make the model return
7:08
consistent outputs. This, of course, gives you a higher degree of control over model behavior. This rolls out in beta today.
7:15
[applause] -In the coming weeks, we'll roll out a feature to let you view
7:22
logprobs in the API. [applause]
7:27
-All right. Number three, better world knowledge. You want these models to be able to access better knowledge about the world,
7:34
so do we. We're launching retrieval in the platform. You can bring knowledge from outside documents or databases
7:41
into whatever you're building. We're also updating the knowledge cutoff. We are just as annoyed as all of you, probably more that GPT-4's knowledge
7:49
about the world ended in 2021. We will try to never let it get that out of date again.
7:54
GPT-4 Turbo has knowledge about the world up to April of 2023, and we will continue to improve that over time.
8:03
Number four, new modalities. Surprising no one,
8:08
DALL-E 3, GPT-4 Turbo with vision, and the new text-to-speech model are all going into the API today.
8:17
[applause]
8:23
-We have a handful of customers that have just started using DALL-E 3 to programmatically generate images and designs.
8:31
Today, Coke is launching a campaign that lets its customers generate Diwali cards using DALL-E 3,
8:36
and of course, our safety systems help developers protect their applications against misuse.
8:41
Those tools are available in the API. GPT-4 Turbo can now accept images as inputs via the API,
8:48
can generate captions, classifications, and analysis. For example, Be My Eyes uses this technology to help people who are blind or have low vision
8:58
with their daily tasks like identifying products in front of them.
9:04
With our new text-to-speech model, you'll be able to generate incredibly natural-sounding audio
9:10
from text in the API with six preset voices to choose from. I'll play an example.
9:16
-Did you know that Alexander Graham Bell, the eminent inventor, was enchanted by the world of sounds.
9:21
His ingenious mind led to the creation of the graphophone, which etches sounds onto wax, making voices whisper through time.
9:30
-This is much more natural than anything else we've heard out there. Voice can make apps more natural to interact with and more accessible.
9:38
It also unlocks a lot of use cases like language learning, and voice assistance.
9:43
Speaking of new modalities, we're also releasing the next version of our open-source speech recognition model,
9:49
Whisper V3 today, and it'll be coming soon to the API. It features improved performance across many languages,
9:56
and we think you're really going to like it. Number five, customization.
10:01
Fine-tuning has been working really well for GPT-3.5 since we launched it a few months ago.
10:07
Starting today, we're going to expand that to the 16K version of the model. Also, starting today,
10:14
we're inviting active fine-tuning users to apply for the GPT-4 fine-tuning, experimental access program.
10:21
The fine-tuning API is great for adapting our models to achieve better performance in a wide variety of applications with a relatively small amount of data,
10:29
but you may want a model to learn a completely new knowledge domain, or to use a lot of proprietary data.
10:36
Today we're launching a new program called Custom Models. With Custom Models,
10:41
our researchers will work closely with a company to help them make a great custom model, especially for them,
10:48
and their use case using our tools. This includes modifying every step of the model training process,
10:54
doing additional domain-specific pre-training, a custom RL post-training process tailored for specific domain, and whatever else.
11:02
We won't be able to do this with many companies to start. It'll take a lot of work, and in the interest of expectations,
11:07
at least initially, it won't be cheap, but if you're excited to push things as far as they can currently go.
11:12
Please get in touch with us, and we think we can do something pretty great. Number six, higher rate limits.
11:20
We're doubling the tokens per minute for all of our established GPT-4 customers, so it's easier to do more.
11:26
You'll be able to request changes to further rate limits and quotas directly in your API account settings.
11:32
In addition to these rate limits, it's important to do everything we can do to make you successful building
11:39
on our platform. We're introducing copyright shield. Copyright shield means that we will step in and defend
11:46
our customers and pay the costs incurred, if you face legal claims or on copyright infringement, and this applies both
11:53
to ChatGPT Enterprise and the API. Let me be clear, this is a good time to remind
11:59
people do not train on data from the API or ChatGPT Enterprise ever.
12:06
All right. There's actually one more developer request that's been even bigger than all of these and so I'd like to talk about that now
12:16
and that's pricing. [laughter] -GPT-4 Turbo
12:22
is the industry-leading model. It delivers a lot of improvements that we just covered
12:27
and it's a smarter model than GPT-4. We've heard from developers that there are a lot of things that they want to build,
12:35
but GPT-4 just costs too much. They've told us that if we could decrease the cost by 20%, 25%, that would be great.
12:43
A huge leap forward. I'm super excited to announce that we worked really hard on this
12:49
and GPT-4 Turbo, a better model, is considerably cheaper than GPT-4 by a factor of 3x for prompt tokens.
12:58
[applause]
13:05
-And 2x for completion tokens starting today. [applause]
13:12
-The new pricing is 1¢ per 1,000 prompt tokens and 3¢ per 1,000 completion tokens.
13:18
For most customers, that will lead to a blended rate more than 2.75 times cheaper to use
13:23
for GPT-4 Turbo than GPT-4. We worked super hard to make this happen. We hope you're as excited about it as we are.
13:30
[applause]
13:35
-We decided to prioritize price first because we had to choose one or the other, but we're going to work on speed next.
13:41
We know that speed is important too. Soon you will notice GPT-4 Turbo becoming a lot faster.
13:48
We're also decreasing the cost of GPT-3.5 Turbo 16K. Also, input tokens are 3x less and output tokens are 2x less.
13:57
Which means that GPT-3.5 16K is now cheaper than the previous GPT-3.5 4K model.
14:06
Running a fine-tuned GPT-3.5 Turbo 16K version is also cheaper than the old fine-tuned 4K version.
14:13
Okay, so we just covered a lot about the model itself. We hope that these changes address your feedback.
14:19
We're really excited to bring all of these improvements to everybody now.
14:24
In all of this, we're lucky to have a partner who is instrumental in making it happen.
14:30
I'd like to bring out a special guest, Satya Nadella, the CEO of Microsoft. [audience cheers]
14:37
[music] -Good to see you. -Thank you so much. Thank you.
14:42
-Satya, thanks so much for coming here. -It's fantastic to be here and Sam, congrats.
14:48
I'm really looking forward to Turbo and everything else that you have coming. It's been just fantastic partnering with you guys.
14:54
-Awesome. Two questions. I won't take too much of your time. How is Microsoft thinking about the partnership currently?
14:59
-First- [laughter] --we love you guys. [laughter]
15:05
-Look, it's been fantastic for us. In fact, I remember the first time I think you reached out
15:11
and said, "Hey, do you have some Azure credits?" We've come a long way from there. -Thank you for those. That was great.
15:18
-You guys have built something magical. Quite frankly, there are two things for us when it comes to the partnership.
15:23
The first is these workloads. Even when I was listening backstage to how you're describing what's coming,
15:28
even, it's just so different and new. I've been in this infrastructure business for three decades.
15:33
-No one has ever seen infrastructure like this. -The workload, the pattern of the workload,
15:39
these training jobs are so synchronous and so large, and so data parallel.
15:45
The first thing that we have been doing is building in partnership with you, the system, all the way from thinking from power to the DC to the rack,
15:53
to the accelerators, to the network. Just really the shape of Azure is drastically changed
16:01
and is changing rapidly in support of these models that you're building. Our job, number one, is to build the best system
16:09
so that you can build the best models and then make that all available to developers. The other thing is we ourselves are our developers.
16:16
We're building products. In fact, my own conviction of this entire generation of foundation models completely changed the first time I saw GitHub Copilot
16:25
on GPT. We want to build our GitHub Copilot all as developers on top of OpenAI APIs.
16:36
We are very, very committed to that. What does that mean to developers? Look, I always think of Microsoft as a platform company,
16:43
a developer company, and a partner company. For example, we want to make GitHub Copilot available,
16:50
the Enterprise edition available to all the attendees here so that they can try it out. That's awesome. We are very excited about that.
16:57
[applause] -You can count on us to build the best infrastructure in Azure
17:05
with your API support and bring it to all of you. Even things like the Azure marketplace.
17:11
For developers who are building products out here to get to market rapidly. That's really our intent here.
17:17
-Great. How do you think about the future, future of the partnership, or future of AI, or whatever?
17:23
Anything you want -There are a couple of things for me that I think are going to be very,
17:29
very key for us. One is I just described how the systems that are needed
17:36
as you aggressively push forward on your roadmap requires us to be on the top of our game and we intend fully to commit
17:45
ourselves deeply to making sure you all as builders of these foundation models
17:51
have not only the best systems for training and inference, but the most compute, so that you can keep pushing-
17:57
-We appreciate that. --forward on the frontiers because I think that's the way we are going to make progress.
18:02
The second thing I think both of us care about, in fact, quite frankly, the thing that excited both sides to come together is
18:09
your mission and our mission. Our mission is to empower every person and every organization on the planet to achieve more.
18:15
To me, ultimately AI is only going to be useful if it truly does empower. I saw the video you played early.
18:21
That was fantastic to hear those voices describe what AI meant for them
18:27
and what they were able to achieve. Ultimately, it's about being able to get the benefits of AI broadly disseminated to everyone,
18:34
I think is going to be the goal for us. Then the last thing is of course, we are very grounded in the fact that safety matters,
18:39
and safety is not something that you'd care about later, but it's something we do shift left on and we are very, very focused on that with you all.
18:46
-Great. Well, I think we have the best partnership in tech. I'm excited for us to build AGI together. -Oh, I'm really excited. Have a fantastic [crosstalk]. -Thank you very much for coming.
18:52
-Thank you so much. -See you. [applause]
19:03
-We have shared a lot of great updates for developers already and we got a lot more to come, but even though this is developer conference,
19:10
we can't resist making some improvements to ChatGPT. A small one, ChatGPT now uses GPT-4 Turbo with all the latest improvements,
19:20
including the latest knowledge cutoff, which will continue to update. That's all live today.
19:25
It can now browse the web when it needs to, write and run code, analyze data, take and generate images,
19:31
and much more. We heard your feedback, that model picker, extremely annoying, that is gone starting today.
19:36
You will not have to click around the dropdown menu. All of this will just work together. Yes.
19:42
[applause] -ChatGPT will just know what to use and when you need it,
19:51
but that's not the main thing. Neither was price actually the main developer request.
19:58
There was one that was even bigger than that. I want to talk about where we're headed and the main thing we're here to talk
20:03
about today. We believe that if you give people better tools, they will do amazing things.
20:10
We know that people want AI that is smarter, more personal, more customizable, can do more on your behalf.
20:16
Eventually, you'll just ask the computer for what you need and it'll do all of these tasks for you.
20:23
These capabilities are often talked in the AI field about as "agents."
20:28
The upsides of this are going to be tremendous. At OpenAI, we really believe that gradual iterative deployment is
20:36
the best way to address the safety issues, the safety challenges with AI. We think it's especially important to move carefully
20:42
towards this future of agents. It's going to require a lot of technical work and a lot of thoughtful consideration by society.
20:50
Today, we're taking our first small step that moves us towards this future.
20:57
We're thrilled to introduce GPTs. GPTs are tailored versions of ChatGPT for a specific purpose.
21:07
You can build a GPT, a customized version of ChatGPT for almost anything
21:12
with instructions, expanded knowledge, and actions, and then you can publish it for others to use.
21:19
Because they combine instructions, expanded knowledge, and actions, they can be more helpful to you.
21:25
They can work better in many contexts, and they can give you better control. They'll make it easier for you to accomplish all sorts of tasks
21:32
or just have more fun and you'll be able to use them right within ChatGPT.
21:37
You can in effect program a GPT with language just by talking to it.
21:42
It's easy to customize the behavior so that it fits what you want. This makes building them very accessible
21:48
and it gives agency to everyone. We're going to show you what GPTs are,
21:53
how to use them, how to build them, and then we're going to talk about how they'll be distributed and discovered.
22:00
After that for developers, we're going to show you how to build these agent-like experiences into your own apps.
22:05
First, let's look at a few examples. Our partners at Code.org are working hard to expand computer science in schools.
22:15
They've got a curriculum that is used by tens of millions of students worldwide. Code.org, crafted Lesson Planner GPT, to help teachers provide
22:24
a more engaging experience for middle schoolers. If a teacher asks it to explain four loops in a creative way,
22:30
it does just that. In this case, it'll do it in terms of a video game character repeatedly picking up coins.
22:37
Super easy to understand for an 8th-grader. As you can see, this GPT brings together Code.org's,
22:43
extensive curriculum and expertise, and lets teachers adapt it to their needs quickly and easily.
22:49
Next, Canva has built a GPT that lets you start designing by describing what you want
22:55
in natural language. If you say, "Make a poster for a DevDay reception this afternoon,
23:01
this evening," and you give it some details, it'll generate a few options to start with by hitting Canva's APIs.
23:07
Now, this concept may be familiar to some of you. We've evolved our plugins to be custom actions for GPTs.
23:14
You can keep chatting with this to see different iterations, and when you see one you like, you can click through to Canva
23:20
for the full design experience. Now we'd like to show you a GPT Live.
23:27
Zapier has built a GPT that lets you perform actions across 6,000 applications to unlock all kinds of integration possibilities.
23:36
I'd like to introduce Jessica, one of our solutions architects, who is going to drive this demo. Welcome Jessica.
23:42
[applause] -Thank you, Sam. Hello everyone. Thank you all.
23:49
Thank you all for being here. My name is Jessica Shieh. I work with partners and customers to bring their product alive.
23:55
Today I can't wait to show you how hard we've been working on this, so let's get started.
24:01
To start where your GPT will live is on this upper left corner. I'm going to start with clicking on the Zapier AI actions
24:10
and on the right-hand side you can see that's my calendar for today. It's quite a day ever.
24:15
I've already used this before, so it's actually already connected to my calendar. To start, I can ask,
24:22
"What's on my schedule for today?" We build GPTs with security in mind. Before it performs any action or share data,
24:30
it will ask for your permission. Right here, I'm going to say allowed.
24:36
GPT is designed to take in your instructions, make the decision on which capability to call to perform that action,
24:43
and then execute that for you. You can see right here, it's already connected to my calendar.
24:49
It pulls into my information and then I've also prompted it to identify
24:54
conflicts on my calendar. You can see right here it actually was able to identify that.
25:01
It looks like I have something coming up. What if I want to let Sam know that I have to leave early? Right here I say, "Let Sam know I got to go.
25:11
Chasing GPUs." With that, I'm going to swap to my conversation with Sam
25:21
and then I'm going to say, "Yes, please run that."
25:26
Sam, did you get that? -I did. -Awesome.
25:32
[applause] -This is only a glimpse of what is possible and I cannot wait to see
25:40
what you all will build. Thank you. Back to you, Sam. [applause]
25:51
-Thank you, Jessica. Those are three great examples. In addition to these, there are many more kinds of GPTs that people are creating and many,
25:59
many more that will be created soon. We know that many people who want to build a GPT don't know how to code.
26:07
We've made it so that you can program a GPT just by having a conversation.
26:12
We believe that natural language is going to be a big part of how people use computers in the future and we think this is an interesting early example.
26:19
I'd like to show you how to build one.
26:25
All right. I want to create a GPT that helps give founders and developers advice
26:30
when starting new projects. I'm going to go to create a GPT here,
26:36
and this drops me into the GPT builder. I worked with founders for years at YC and still whenever I meet developers,
26:43
the questions I get are always about, "How do I think about a business idea? Can you give me some advice?"
26:49
I'm going to see if I can build a GPT to help with that. To start, GPT builder asks me what I want to make,
26:55
and I'm going to say, "I want to help startup founders think. through their business ideas
27:04
and get advice. After the founder has gotten some advice,
27:13
grill them on why they are not growing faster." [laughter]
27:20
-All right. To start off, I just tell the GPT little bit about what I want here. It's going to go off and start thinking about that,
27:27
and it's going to write some detailed instructions for the GPT. It's also going to,
27:32
let's see, ask me about a name. How do I feel about Startup Mentor? That's fine.
27:37
"That's good." If I didn't like the name, of course, I could call it something else, but it's going to try to have this conversation with me and start there.
27:45
You can see here on the right, in the preview mode that it's already starting to fill out the GPT.
27:53
Where it says what it does, it has some ideas of additional questions that I could ask.
27:58
[chuckles] It just generated a candidate.
28:03
Of course, I could regenerate that or change it, but I like that. I'll say "That's great."
28:13
You see now that the GPT is being built out a little bit more as we go. Now, what I want this to do,
28:19
how it can interact with users, I could talk about style here. What I'm going to say is,
28:25
"I am going to upload transcripts of some lectures
28:31
about startups I have given, please give advice based off of those."
28:38
All right. Now, it's going to go figure out how to do that.
28:43
I would like to show you the configure tab. You can see some of the things that were built out here as we were going
28:49
by the builder itself. You can see that there's capabilities here that I can enable. I could add custom actions.
28:55
These are all fine to leave. I'm going to upload a file. Here is a lecture that I picked that I gave with some startup advice,
29:05
and I'm going to add that here. In terms of these questions, this is a dumb one.
29:11
The rest of those are reasonable, and very much things founders often ask. I'm going to add one more thing to the instructions here,
29:19
which is be concise and constructive with feedback.
29:25
All right. Again, if we had more time, I'd show you a bunch of other things. This is
29:31
a decent start. Now, we can try it out over on this preview tab.
29:36
I will say, what's a common question?
29:44
"What are three things to look for when hiring employees at an early-stage startup?"
29:53
Now, it's going to look at that document I uploaded. It'll also have of course all of the background knowledge of GPT-4.
30:03
That's pretty good. Those are three things that I definitely have said many times. Now, we could go on and it would start following
30:09
the other instructions and grill me on why I'm not growing faster, but in the interest of time, I'm going to skip that.
30:15
I'm going to publish this only to me for now. I can work on it later. I can add more content, I can add a few actions
30:22
that I think would be useful, and then I can share it publicly. That's what it looks like to create a GPT
30:29
[applause] -Thank you.
30:36
By the way, I always wanted to do that after all of the YC office hours, I always thought, "Man, someday I'll be able
30:42
to make a bot that will do this and that'll be awesome." [laughter] -With GPTs, we're letting people easily share and discover all the fun ways
30:51
that they use ChatGPT with the world. You can make private GPT like I just did,
30:58
or you can share your creations publicly with a link for anyone to use,
31:03
or if you're on ChatGPT Enterprise, you can make GPTs just for your company.
31:10
Later this month we're going to launch the GPT store.
31:17
Thank you. I appreciate that. [applause]
31:25
-You can list a GPT there and we'll be able to feature the best and the most popular GPT.
31:30
Of course, we'll make sure that GPTs in the store follow our policies before they're accessible.
31:37
Revenue sharing is important to us. We're going to pay people who build the most useful and the most used GPT
31:44
a portion of our revenue. We're excited to foster a vibrant ecosystem with the GPT store,
31:50
just from what we've been building ourselves over the weekend. We're confident there's going to be a lot of great stuff. We're excited to share more information soon.
31:58
Those are GPTs and we can't wait to see what you'll build. This is a developer conference, and the coolest thing about this
32:05
is that we're bringing the same concept to the API. [applause]
32:15
Many of you have already been building agent-like experiences on the API,
32:20
for example, Shopify's Sidekick, which lets you take actions on the platform. Discord's Clyde,
32:26
lets Discord moderators create custom personalities for, and Snaps My AI,
32:32
a customized chatbot that can be added to group chats and make recommendations. These experiences are great,
32:38
but they have been hard to build. Sometimes taking months, teams of dozens of engineers,
32:44
there's a lot to handle to make this custom assistant experience. Today, we're making that a lot easier with our new Assistants API.
32:54
[applause] -The Assistants API includes persistent threads,
33:01
so they don't have to figure out how to deal with long conversation history, built-in retrieval,
33:07
code interpreter, a working Python interpreter in a sandbox environment, and of course the improved function calling,
33:14
that we talked about earlier. We'd like to show you a demo of how this works.
33:19
Here is Romain, our head of developer experience. Welcome, Romain. [music] [applause]
33:25
-Thank you, Sam. Good morning. Wow. It's fantastic to see you all here.
33:33
It's been so inspiring to see so many of you infusing AI into your apps.
33:38
Today, we're launching new modalities in the API, but we are also very excited
33:43
to improve the developer experience for you all to build assistive agents. Let's dive right in.
33:50
Imagine I'm building $1, travel app for global explorers, and this is the landing page.
33:56
I've actually used GPT-4 to come up with these destination ideas. For those of you with a keen eye, these illustrations
34:02
are generated programmatically using the new DALL-E 3 API available to all of you today.
34:07
It's pretty remarkable. Let's enhance this app by adding a very simple assistant to it.
34:15
This is the screen. We're going to come back to it in a second. First, I'm going to switch over to the new assistant's playground.
34:21
Creating an assistant is easy, you just give it a name, some initial instructions, a model.
34:26
In this case, I'll pick GPT-4 Turbo. Here I'll also go ahead and select some tools. I'll turn on Code Interpreter and retrieval and save.
34:35
That's it. Our assistant is ready to go. Next, I can integrate with two new primitives
34:41
of this Assistants API, threads and messages. Let's take a quick look at the code.
34:48
The process here is very simple. For each new user, I will create a new thread.
34:54
As these users engage with their assistant, I will add their messages to the threads. Very simple.
35:00
Then I can simply run the assistant at any time to stream the responses back to the app.
35:06
We can return to the app and try that in action. If I say, "Hey, let's go to Paris."
35:15
All right. That's it. With just a few lines of code, users can now have a very specialized assistant right inside the app.
35:24
I'd like to highlight one of my favorite features here, function calling. If you have not used it yet, function calling is really powerful.
35:31
As Sam mentioned, we are taking it a step further today. It now guarantees the JSON output with no added latency,
35:38
and you can invoke multiple functions at once for the first time. Here, if I carry on and say, "Hey, what are the top 10 things to do?"
35:49
I'm going to have the assistant respond to that again. Here, what's interesting is that the assistant knows about functions,
35:56
including those to annotate the map that you see on the right. Now, all of these pins are dropping in real-time here.
36:04
Yes, it's pretty cool. [applause]
36:09
-That integration allows our natural language interface to interact fluidly with components and features of our app.
36:16
It truly showcases now the harmony you can build between AI and UI where the assistant is actually taking action.
36:25
Let's talk about retrieval. Retrieval is about giving our assistant more knowledge
36:30
beyond these immediate user messages. In fact, I got inspired and I already booked my tickets to Paris.
36:37
I'm just going to drag and drop here this PDF. While it's uploading, I can just sneak peek at it.
36:43
Very typical United Flight ticket. Behind the scene here, what's happening is that retrieval
36:49
is reading these files, and boom, the information about this PDF appeared on the screen.
36:55
[applause] -This is, of course, a very tiny PDF, but Assistants
37:01
can parse long-form documents from extensive text to intricate product specs depending on what you're building.
37:07
In fact, I also booked an Airbnb, so I'm just going to drag that over to the conversation as well.
37:12
By the way, we've heard from so many of you developers how hard that is to build yourself. You typically need to compute your own biddings,
37:19
you need to set up chunking algorithm. Now all of that is taken care of.
37:24
There's more than retrieval with every API call, you usually need to resend the entire conversation history,
37:31
which means setting up a key-value store, that means handling the context windows, serializing messages, and so forth.
37:37
That complexity now completely goes away with this new stateful API.
37:43
Just because OpenAI is managing this API, does not mean it's a black box. In fact, you can see the steps that the tools are taking
37:49
right inside your developer dashboard. Here, if I go ahead and click on threads,
37:56
this is the thread I believe we're currently working on and see, these are all the steps, including the functions
38:02
being called with the right parameters, and the PDFs I've just uploaded.
38:08
Let's move on to a new capability that many of you have been requesting for a while. Code Interpreter is now available today in the API as well,
38:16
that gives the AI the ability to write and execute code on the fly, but even generate files.
38:22
Let's see that in action. If I say here, "Hey, we'll be four friends staying
38:29
at this Airbnb, what's my share of it plus my flights?"
38:40
All right. Now, here, what's happening is that Code interpreter noticed that it should write some code
38:48
to answer this query. Now it's computing the number of days in Paris, number of friends.
38:53
It's also doing some exchange rate calculation behind the scene to get the sensor for us.
38:58
Not the most complex math, but you get the picture. Imagine you're building a very complex finance app
39:04
that's crunching countless numbers, plotting charts, so really any task that you'd normally tackle with code,
39:10
then Code Interpreter will work great for you. All right. I think my trip to Paris is solid.
39:16
To recap here, we've just seen how you can quickly create an assistant that manages state for your user conversations,
39:22
leverages external tools like knowledge and retrieval and Code Interpreter, and finally invokes your own functions to make things happen
39:32
but there's one more thing I wanted to show you to really open up the possibilities using function calling combined with our new modalities
39:39
that we're launching today. While working on DevDay, I built a small custom assistant
39:45
that knows everything about this event, but instead of having a chat interface while running around all day today,
39:51
I thought, why not use voice instead? Let's bring my phone up on screen here so you can see it on the right.
39:58
Awesome. On the right, you can see a very simple Swift app that takes microphone input.
40:04
On the left, I'm actually going to bring up my terminal log so you can see what's happening behind the scenes.
40:09
Let's give it a shot. Hey there, I'm on the keynote stage right now. Can you greet our attendees here at Dev Day?
40:21
-Hey everyone, welcome to DevDay. It's awesome to have you all here. Let's make it an incredible day.
40:27
[applause]
40:32
-Isn't that impressive? You have six unique and rich voices to choose from in the API,
40:37
each speaking multiple languages, so you can really find the perfect fit for your app.
40:42
On my laptop here on the left, you can see the logs of what's happening behind the scenes, too. I'm using Whisper to convert the voice inputs into text,
40:50
an assistant with GPT-4 Turbo, and finally, the new TTS API to make it speak.
40:56
Thanks to function calling, things get even more interesting when the assistant can connect to the internet and take
41:02
real actions for users. Let's do something even more exciting here together.
41:07
How about this? Hey, Assistant, can you randomly select five DevDay attendees here
41:15
and give them $500 in OpenAI credits? [laughter]
41:20
-Yes, checking the list of attendees.
41:25
[laughter] -Done. I picked five DevDay attendees and added $500 of API credits
41:30
to their account. Congrats to Christine M, Jonathan C, Steven G, Luis K, and Suraj S.
41:38
-All right, if you recognize yourself, awesome. Congrats.
41:43
That's it. A quick overview today of the new Assistants API combined with some of the new tools and modalities that we launched,
41:50
all starting with the simplicity of a rich text or voice conversation for you end users.
41:56
We really can't wait to see what you build, and congrats to our lucky winners. Actually,
42:01
you know what? you're all part of this amazing OpenAI community here so I'm just going to talk to my assistant
42:06
one last time before I step off the stage. Hey Assistant, can you actually give everyone here in the audience $500
42:15
in OpenAI credits? -Sounds great. Let me go through everyone.
42:21
[applause] -All right,
42:28
that function will keep running, but I've run out of time. Thank you so much, everyone.
42:33
Have a great day. Back to you, Sam.
42:44
-Pretty cool, huh? [audience cheers] -All right, so that Assistants API goes into beta today,
42:52
and we are super excited to see what you all do with it, anybody can enable it.
42:57
Over time, GPTs and Assistants are precursors to agents
43:02
are going to be able to do much much more. They'll gradually be able to plan and to perform more complex actions on your behalf.
43:11
As I mentioned before, we really believe in the importance of gradual iterative deployment.
43:16
We believe it's important for people to start building with and using these agents now to get a feel for what the world is going to be like,
43:23
as they become more capable. As we've always done, we'll continue to update our systems based off of your feedback.
43:32
We're super excited that we got to share all of this with you today. We introduced GPTs,
43:37
custom versions of GPT that combine instructions, extended knowledge and actions.
43:44
We launched the Assistants API to make it easier to build assistive experiences with your own apps.
43:49
These are your first steps towards AI agents and we'll be increasing their capabilities over time.
43:56
We introduced a new GPT-4 Turbo model that delivers improved function calling, knowledge, lowered pricing, new modalities, and more.
44:05
We're deepening our partnership with Microsoft. In closing, I wanted to take a minute to thank the team that creates all of this.
44:13
OpenAI has got remarkable talent density, but still, it takes a huge amount of hard work and coordination to make all this happen.
44:21
I truly believe that I've got the best colleagues in the world. I feel incredibly grateful to get to work with them.
44:27
We do all of this because we believe that AI is going to be a technological and societal revolution.
44:33
It'll change the world in many ways and we're happy to get to work on something that will empower all of you
44:38
to build so much for all of us. We talked about earlier how if you give people better tools,
44:44
they can change the world. We believe that AI will be about individual empowerment and agency
44:50
at a scale that we've never seen before and that will elevate humanity to a scale that we've never seen before either.
44:55
We'll be able to do more, to create more, and to have more. As intelligence gets integrated everywhere,
45:02
we will all have superpowers on demand. We're excited to see what you all will do with this technology
45:08
and to discover the new future that we're all going to architect together. We hope that you'll come back next year.
45:14
What we launched today is going to look very quaint relative to what we're busy creating for you know. Thank you for all that you do.
45:21
Thank you for coming here today. [applause]
45:28
[music]
'松泉, 인생글, 바라보기' 카테고리의 다른 글
바라보기, 언어, panic disorder, 恐慌 障礙, 공황 장애 (4) | 2023.11.11 |
---|---|
바라보기, 언어, retreat, 避靜, 피정 시설 (0) | 2023.11.11 |
바라보기, 언어, feel the wave, alohapoke (0) | 2023.11.09 |
2023 시네 도슨트, 메가박스 코엑스, 시즌3 위대한 예술가들을 만나다, 피카소 (0) | 2023.11.07 |
바라보기, 곤충, 전략의 귀재들, 2부 숨바꼭질 (0) | 2023.11.01 |