-Hello. Hi. [ Applause ] Hi everyone.
Welcome to our second DevDay ever. Yes. [ Cheers and applause ] First, you might notice that things look a little different compared to last year.
I'm not Sam. I'm not going to announce GPT-5, but we've got some super exciting things in store for you today. I'm Olivier.
I lead the API product team at OpenAI. And today is a unique opportunity for all of us to take a break from our computers and whatever box we were bashing in the morning and spend time together. We have a jam-packed agenda; breakout sessions, demonstration by the OpenAI team, a new developer community talks, which I'm super excited about.
Everything is designed to give you a chance to spend the time with the OpenAI team and with each other. I wanted to start, but why do we even do DevDays? Why do we have a developer platform?
OpenAI mission is to build AGI that benefits all of humanity. And developers are critical to that mission. Put shortly, we cannot do this without you.
You'll see the future before everyone else and you cannot help but bring it to life. Four years ago, we trained GPT-3. We knew we had something pretty cool, big on our hands with GPT-3, but we had no idea what was possible with it.
What are the real world use cases of that technology? So we built an API. The goal of the API was to let you all explore what this technology could do.
Turns out that you can do quite a bit with GPT-3; creating marketing content, translating languages, and even building a chatbot. Looking back, GPT-3 capabilities were very, very limited. There was no tool support.
Hallucination were all over the place. And the latency was super, super slow. Fast forward to today, AI is completely unrecognizable compared to the GPT-3 era.
There are more AI models than ever before. More AI applications moving from prototyping to production. And new developers are pushing the frontier of what the models can do every single day.
I'm pretty impressed. Like we have now 3 million developers building on OpenAI across more than 200 countries. And I just checked the stats compared to last year's DevDay, we tripled the number of active applications built on OpenAI.
So thank you for being a part of it. We've been building quite too at OpenAI. In the last year, we launched more than 100 new API features; that includes things like structured outputs, the batch API, new fine-tuning support for models, new playground, and embeddings, and so much more.
New models as well. We introduced the new GPT-4o and 4 mini to push the boundary both on intelligence and cost efficiency. And we are not slowing down.
I'm going to cover in the next 20 minutes, the four top areas of focus of OpenAI. First, we'll cover the best-in-class frontier models. Second, we'll build more powerful multimodal capability.
Third, we'll enable you to even customize like more deeply models. And lastly, we'll talk about how can you scale even more easily on OpenAI. Let's start by talking about models.
Just over two weeks ago, we introduced the world to our first series of models that can reason about complex problems. That's o1. Maybe a quick show of hands, how many of you have tried o1 in the past couple of weeks?
Full house? I love it. We've got only early adopters in the audience.
Okay, so as you know, we have two versions of o1. The first one is o1-preview, which is a preview of the larger o1 model that we plan to release very soon. You can think of o1-preview as the more as the world's most powerful reasoning model.
o1-preview is designed to solve really hard problems that require a long thought process. That's o1-preview. The second flavor is o1-mini.
o1-mini is, well, a smaller, faster, cheaper version of o1. o1-mini is fantastic if you're debugging code or building an agentic application. o1-mini, in general, we found it to be really good at coding, math and scientific reasoning.
But what do we even mean by reasoning? Reasoning comes so naturally to us as human beings that sometimes it can be hard to articulate. So let me break it down.
Reasoning is the ability for a model to have a reliable chain of thought. The model thinks step by step before it responds. The model can think through different meanings of a question and reflect on its answer before responding.
So let's take a look at the quick example of chain of thought in ChatGPT. So I'm not going to go into details, but that's essentially a brain teaser, which is non-trivial, that requires the model to do quite a bit of logical reasoning. And so as you can see, the model is going to reason methodically about the problem.
Thinking through each line, each sentence of the problem and making sure before it answers that the answer meets all the requirements of my question. Now, I know that's probably the most common question I've got. What the heck is the name o1?
The shift to reasoning introduces a new shape of AI capability. The ability for our model to scale and correct its thought process, it's pretty mind blowing. So we are resetting the clock and we are introducing a new series of model under the name o1.
But let's look at a few customers. A few of our customers tested o1 to see what it could do for their use case. First, I wanted to talk about Cognition.
So Cognition is building -- Scott. Scott, I see you. Cognition is building Devin and has been testing o1 with its AI software agents to see how much more o1 could accurately plan, write, and debug code.
So let's learn more from Scott, the founder of Cognition. -Yeah, I'm Scott Wu, I'm the CEO and co-founder at Cognition. The thing that's interesting about programming is that it's changed in shape like over and over, over the last, you know, 50 years.
Being able to write code out line by line is one thing, but also being able to understand and interpret logs, being able to navigate and run commands that actually, you know, affect the real world is a big part of programming. This new model o1 is really able to process and make decisions in a more human-like way. We worked closely with Open AI on iterating on it and giving our feedback and testing it out.
What we found was that o1 was just incredible at this kind of reasoning. At Cognition AI, we're building Devin, the first fully autonomous software agent. And, you know, what that means is that Devin is able to go and build tasks from scratch and is able to work on problems the same way that a software engineer would.
And so here actually, I asked Devin to analyze the sentiment of this tweet and to use a few different ML services out there to run those out of the box and break down this particular piece of text and understand what the sentiment is. And so first, Devin will make a plan of how it plans to approach this problem. Yeah.
And so here, you know, it had some trouble fetching it from the browser and so then instead it decided to go and go through the API and fetch the tweet that way, right? And all these little decisions, you know, things like that happen all the time. Finally, it's able to get this all the way through.
It says that the predominant emotion of this tweet is happiness. The soul of programming has always been the ability to take your ideas and turn them into reality. And now you can be able to do that much faster and cleaner than ever before.
-That's super cool. Our second example I wanted to mention is Casetext and Thomson Reuters. They've been using o1 to power an AI legal assistant.
In professions like legal, nuance and precision are critical. Every single word matters. So let's take a look at an example.
Here, I'm asking an AI legal assistant to verify the accuracy of the legal statement. On the left, 4, on the right, o1. So as you can see, 4 claims that there are no errors in the statement, which is incorrect.
However, on the right, o1 pays attention to very specific words, which are critical, like neither and nor, which completely changes the meaning of the legal text. And so as you can see the model here, o1, is able to see that there are small but critical errors in the document. I mean, listen, I could talk about reasoning for hours, but I think it's going to be more fun to see a live demo.
So next, I'm going to pass it to Roman, our Head of Developer Experience to show you what o1 can do in action. Welcome, Roman. -Thank you, Olivier.
Good morning, everyone. It's fantastic to see you all. In just two weeks, we've been blown away by the incredible projects you've all been already created with o1, from tackling hard problems to building complete apps.
So I thought, what about kicking things off by creating an iPhone app from scratch here on stage? So I had this little idea for an ISS tracking app. You know what, before you start reading the whole thing, it's a bit long.
I'm just going to copy all of that over to ChatGPT desktop and send it. On the left side, I have an empty exclude project open. And what we'll do here is we'll take a look at what I'm trying to build.
So this app is articulating a few things like, "Hey, like I'd like to create like the ISS map. I'd like to track it in real time. I'd like to fetch the astronauts in space," like a pretty intricate prompt with lots of small details.
And you might wonder, "Well, how is that different from FOGO? " Well, in this case, like o1 is basically trying to go through making a plan, architecting what's needed, preparing the architecture, and ultimately finding out which APIs to use. So now, to save a few seconds, I actually instructed o1 to bring everything in a single file.
So I should be able to go copy this whole file, this entire file and bring it over to the project. So this is a real life demo. So this is always the challenge, the tricky part.
We never know what's going to happen. So I'm going to try and build this app right away on my phone. And it builds.
Can we bring my phone on stage please? Let's see it. It's happening.
Zero shot. And what's pretty amazing here is like, yeah, the ISS is moving like I intended. If I go to the second tab, the astronauts in space also work.
And I could keep on going and I could ask o1 what should we build next? And it can basically be my thought partner. And what's really, really amazing here is like, o1 is like having someone that understands your vision right away by your side.
It's able to kind of complex -- take like the complex instructions you have and architect the whole thing. Lots of great ideas. Actually, I kind of like the second one.
If we had more time, I think I would build the AR view for it. Anyways, it's pretty awesome. But we all know that coding is more about working on an existing code base more so than starting fresh, right?
So I wanted to dive into something a bit more advanced. I took kind of a weekend project to interface with these little drones called Tello. But the thing is like I'd love to play with them, but I've never actually programmed one before.
And so what I've done is like I prepared a little web interface for it in React, but that's all I have and I'm honestly not sure where to start. And that's exactly where o1-mini comes in. And one of the great things about o1-mini being in the API is that I can use it directly from my favorite tools.
In this case, I'm going to bring up Cursor. So I have my entire project here open on Cursor, the back end, the front end, all my react components. And on the right side, I have the Chat and I can zoom in a little bit so you have -- so you can see better.
So I prepared this prompt here, but what's going on to give you an idea at a high level is that I have the front end. I have literally no backend for now. And I'm telling o1-mini, "Hey, so I have this like drone, I'd like to interface with it in JavaScript.
I prepared some command lines. " By the way, we started some work previously with o1 to figure out which was the kind of camera stream format. This was one of my aha moments with o1 that figured out the right FFmpeg.
So it's saved in one file. And now all I want is like I'm going to send this prompt to o1 and in a few seconds, hopefully, it's going to kind of craft the plan and try to wire up all of these missing pieces in my app. So you can take a quick look here.
This is my front end that we just saw, but the server is completely empty because I have not made this work yet. So I hopefully in a few seconds, this is, again, very much live, we'll see what o1-mini cooks for us. All right, so what did it come up with, right?
So it basically created the entire server for me. That looks pretty good. So I'm going to copy it over.
What else is missing in here? Perfect, let's move on. So I had previously installed the NPM packages.
Now, let's take a look at the command interface. What's pretty cool is that it did add all of the WebSocket code. And check this out, I'm going to try to apply these changes right here.
And now, it's actually wiring up all of the buttons that I had in my interface to send the right commands apparently with the right formats to the drones. So I hope this will work. Let's take it -- let's take a look.
So I'm going to accept these changes. Let's see what else? I think we just got to start the server.
And this I think was already done. This is additional. All right, let's move on.
I think we're good. Let's save this. Let's go to the terminal and let's start the server.
First of all, I'm going to have actually to connect to the Wi-Fi network that it's emitting. It's right there. Cool.
I'm going to put it down right here and see what happens. Let's hope for the best, it's the time to put it to test. So let's go here and start a server we just created.
Cool. Let's take a look at the UI now. If I -- it's good news, the battery level showed up, let's start a video stream.
Does that work? It does. Awesome.
All right, and now, let's actually try the real thing. All right, the buttons are working. Amazing.
I can go a bit higher. This is a good time for a drone selfie. Awesome.
And we can now land. Amazing. [ Cheers and applause ] Isn't that pretty cool?
Like I had no experience working with this drone before, and o1-mini architected the low level code with UDP packets, connected it with my front end and React components with WebSockets. And we are really entering a new paradigm here with reasoning. I can't wait to see what all of you are going to tackle with o1 from your challenges and also solve problems for your users.
Thank you so much and back to you, Olivier. [ Cheers and applause ] -Thank you, Roman. That was awesome.
And that was an only an early preview of what o1 can do. In the last few weeks, we have already raised the rate limits for o1 three times. And today, we're going to increase them again.
So we're doubling rate limits as of now. -Here we go. It's live.
-So at that point, the o1 rating meets match GPT-4o so happy scaling. That's all. I was sharing with a few of you earlier today who do not have access yet to o1.
And so starting now, everyone in the audience officially has access to o1. So have fun, explore what you can do with it, and tell us what you think. We are going to keep working on o1.
And there are quite a few features that we want to ship in the near-term. Stuff such as function coding, structured outputs, system messages, image support, and much more. o1 is super exciting.
If you've tested o1 firsthand, you will know it's not a drop-in replacement for GPT-4o. In a way, o1 is slower and more expensive than 4o. And so what we see already from like early testing is that the best AI applications combine both GPT-4o and o1.
So we will keep investing a lot in GPT-4o and 4o mini. And we're pretty excited to share some of the new things that we've been building for these models. Let's focus on multimodality.
How can we help you build more powerful multimodel capabilities? As models get smarter, they will also be able to understand and respond across all modalities; text, images, video and audio, of course. And I'm sure many of you have played with Advanced Voice Mode in ChatGPT, it's pretty damn cool.
And one of the top requests we've heard again and again is how can I get this natural speech to speech feature within my application? And so today, I have some pretty good news. It's coming, it's coming.
We're announcing the Realtime API. The Realtime API is a super low latency API that lets you build AI experiences real time within Europe. The API streams inputs and outputs in real time using WebSockets.
We're launching it starting today with support for speech to speech, which is the same technology that powers Advanced Voice Mode. yes, that powers Advanced Voice Mode in ChatGPT. And that will allow you to build voice experiences using the six voices already available in the API.
In general, but in particular for the Realtime API, a demo is worth 1,000 words. So you know what? Let's bring back Roman on stage to show us what's possible with the Realtime API.
-Thanks, Olivier. I'm truly excited about this Realtime API already kind of give me some ideas for this, so I'll have to work on it after. You know, a lot of you have been asking about building amazing speech-to-speech experiences right into your apps.
Well now, you can. So let's take a look. First under the hood, this is the Open AI playground that all of you know, I'm sure, and it now has, as of today, full support for the new Realtime API.
So in fact, we can try it right away. Hi there. -Hi there.
How can I help you today? -Hey, can you tell me a fun fact about OpenAI DevDay? -Sure.
Open AI DevDay is an annual event where developers, researchers, and enthusiasts gather to learn about the latest advancements and tools from OpenAI. -Cool. Thank you.
-It's a great op -- -Could you -- -Glad you think so. -Could you whisper a little secret like your best kept secret about being a great developer? -Here's a little secret for you.
The key to being a great developer is to never stop learning. -Awesome. Thank you.
-Take your -- -Great. -Glad you liked it. -So as you can tell, you have now the ability to kind of interrupt, generate intonations all of the features and capabilities you know and love from ChatGPT as of today in the Realtime API for you.
If I click on logs, you'll see behind the scenes like all of the events coming up on this WebSocket connection that's fully programmable for all of you. Now for the playground power users out here in the room, I'm not sure if you've noticed this little new shiny icon on the top right. If I click on it, check this out.
If I were to build, let's say an event assistant app for DevDay, and if I were to just say event assistant and hit Enter, this is now the ability for you to quickly generate prompts so I can expand it from here, but in just two words, I have a complete prompt to get started. And we've heard from so many of you how crafting prompts can be tedious at times. Well, we want to help you build as fast as possible on OpenAI.
And by the way, it's not just prompts. Let's say I want to create a new function here and let's say I would like to check in an attendee for my event app while instead of typing the function definition by hand, I have one ready to go. So that's pretty convenient and add to fingertip also for JSON schemas.
Awesome. But now, let's take a look at a concrete example of what can you build with this Realtime API. So for that, I'm going to bring back our favorite fictional travel app from last DevDay, Wanderlust.
It's great to see so many familiar faces in the room. Thanks for being on the journey with us. So if you remember, last year, we used this app to show how to build an assistant with a text kind of conversation, right?
So if I were to say SF here, this was kind of the demo of like using it with text. Let me just refresh it real quick. All right.
-Here's San Francisco, what can I help you with? -Now, the first thing you notice is that there is a voice speaking and in fact I also dropped some code to integrate with the Realtime API into Wanderlust and I can click the cycle into now turn on the voice assistant. So let's try it.
Hey, I'm planning my trip to OpenAI DevDay in London and then in Singapore. -Got it. London and Singapore for OpenAI DevDay.
What do you need help with? -Could you give me some recommendations of places to eat in London? -Sure thing.
Here are some tasty spots. Dishoom, Sketch, and Duck & Waffle are must tries. -That sounds great.
Do you think you could also give me some hotel recommendation? I'm thinking about staying in the kind of Strand neighborhood by the river. -Absolutely.
In the Strand area, the Savoy, ME London, and Strand Palace are fantastic choices. Enjoy your stay. -Cool.
I was thinking with something more lowkey but. . .
All right, that's a bit far. Let's go back to London, please. -Sure.
Focusing on London, for budget stays in the Strand, try Z Hotel Strand, Hub by Premier Inn, or The Strand Continental. -Awesome. So the two things I want to highlight here is that, one, on the left side you could tell we have the complete transcript of the real-time conversation as it's happening and you can have that right away on your website at connection.
And on the right side of the screen, as you could tell, the Realtime API supports functions and tools and APIs, which means that you have the ability to connect to kind of any function. And in this case, for me it was like annotating the map and placing pinpoints on it and so on and so forth. So we think you're going to love using the Realtime API.
And if you're users, for instance, are driving, learning, cooking or any kind of context, we think they're going to be delighted as well. But now, let's try something a little harder. So what if we could give some extra capabilities to our voice assistant beyond this interface?
So let me refresh this and go back to a different kind of conversation. -San Francisco, here we are. How can I assist?
-Yeah, could we take a look at Fort Mason, please? -There we go. Fort Mason, what do you need help with there?
-Well, I have a few hundred awesome developers here in the room and I'm thinking they might need a little snack. Do you have some chocolate shops around here? -Absolutely.
Check out Ghirardelli Square, Ilan's strawberries and The Candy Store. -Great. Can you show me more details about the three places?
-Details are up. Ghirardelli Square offers iconic sundaes. Ilan's Strawberry Lane has delectable chocolate covered strawberries and The Candy Store is your go-to for vintage sweets.
Enjoy. -Great. So as you can tell another example of like pulling details with tools in real time.
Now, what's really awesome about the Realtime API street pointing functions and tools is that you can actually call third party APIs. So for example, what if we were to add Twilio's phone calling capabilities onto Wanderlust by using the Twilio API. So I see here the Ilan's strawberries location seems to accept AI orders.
And so what we're going to do is like we're going to try this live. So I'm going to invite Ilan's from our developer experience team onto the stage to join me so you can actually see that in action. Thanks, Ilan.
So now what I'm going to do is I'm going to resume that conversation we started and we're going to try that live. Let's go. Hey, Ilan's Strawberries looks really fantastic.
Could you place a call and see if you could get us 400 strawberries delivered to the venue? But please keep that under $1,500. -I'm on it.
We'll get those strawberries delivered for you. -Hello. -Hi there, is this Ilan's?
I'm Roman's AI assistant calling about an order. Fantastic. Could you tell me what flavors of strawberry dips you have available?
-Yeah, we have chocolate, vanilla, and we have peanut butter. -Great. How much would 400 chocolate covered strawberries cost?
-400? Are you sure you want 400? -Yes, 400 chocolate covered strawberries please.
How much would that be? -I think that'll be around like $1,415 with 92 cents. -Awesome.
Let's go ahead and place the order for 400 chocolate covered strawberries. -Great. Where would you like that delivered?
-Please deliver them to the Gateway Pavilion at Fort Mason and I'll be paying in cash. -Okay, sweet. So just to confirm, you want 400 chocolate covered strawberries to the Gateway Pavilion?
-Yes. That's perfect. And when can we expect delivery?
-Well, you guys are right nearby so it'll be like, I don't know, 37 seconds. -That's incredibly fast, thank you so much. Have a wonderful day.
-Cool, you too. Awesome. [ Cheers and applause ] -Thanks, Ilan.
Wasn't that cool? Of course, this is just a simple example showing the power of the Realtime API combined with tools, but we hope that's going to inspire many ideas for your own apps and products. And not only did we place the call with Twilio, but as you could see the transcript, we also extracted important details of that conversation along the way.
So I really can wait to see what all of you are going to build with this Realtime API and I'm sure you're going to reinvent how humans and computers interface with each other with these new like voice capabilities. Thank you so much for your time. I can't wait to spend time with all of you today.
And with that, back to you, Olivier. [ Cheers and applause ] -Thank you, Roman. That was awesome.
Let's look at the few real applications that have been testing the Realtime, API in alpha in the past few weeks. The first one I wanted to highlight is Healthify. Healthify is a nutrition and fitness coaching application.
So the Healthify team has been testing the Realtime API to create a more realistic coaching experience using voice. Not only can you have a natural conversation with our AI health coach, but you can also ask your AI coach to call and connect you with a health coach in the real world. A second pretty cool example is Speak.
Speak is building a language learning application. Speak has been testing the Realtime API to create a more natural and intuitive learning experience. Let's check it out in action.
-What I want to show you today is something that we call realtime roleplay as the first thing that we're launching on top of the Realtime API. Part of what's really cool about it is that speech to speech and the multimodality really allows us to understand more than just the pure text transcript of what a user is saying into the app. [ Speaking Spanish ] -The word baño is pronounced "ba-nyo" with a soft "ny" sound like in the English word canyon.
Can you try saying that again? -[ Speaking Spanish ] -Perfecto. [ Speaking Spanish ] -And the model basically has a lot of the very human-like capabilities of a real human teacher.
Like how amazing is that? -The Realtime API is working with GPT-4o. And we will begin rollout in public beta starting today.
[ Cheers and applause ] That API is focused on speech initially, but we want to support more modalities over time such as video. If you want to learn more about the Realtime API, there will be a breakout session completely focused on that later today. All right, let's move on to our third focus area.
We are focused on making our models more customizable. In the future, I believe that every single company will have a custom model which is tailored to its needs. The model will understand really well your businesses, your goals, your use cases.
Most recently, we launched fine tuning for GPT-4o and 4o mini to every developer. The launch reception was really good. It actually broke our dashboards.
You know, it's a good point to have. And today, something new. We're excited to announce vision fine-tuning with GPT-4o.
Vision fine-tuning means that you can now use images to fine-tune the model and improve result from image understanding tasks. Vision fine-tuning can be used for anything from product recommendations, medical imaging, and even building autonomous driving system. As a matter of fact, one of our alpha developers was Grab.
Grab is a really cool company building a ride sharing application in Singapore. They use fine-tuning to teach its mapping system to identify traffic signs and count lane dividers. It looks like a pretty complex like problem that you're not used to before, but actually, Grab was able to do that with 100 images of traffic signs.
And the MO performance was able to spot these common sites on the road as a result. Plenty of other use cases from finding product defects, creating no code UI and so much more. As of today, fine-tuning with vision is available to every single developer on phone.
[ Applause ] The last theme of the day is to help you scale on our platform. As you've noticed, we are obsessed by bringing down the cost of our models. The reason is super simple.
Every time we reduce the price, we see new types of application, new types of use cases emerge. We are super far from the price equilibrium. In a way models are still too expensive to deploy massive scale in particular on cost-sensitive use cases.
And since the release of text-davinci-003, which was the first version of GPT-3. 5, we have decreased the cost per token by 99%. That's right, today's models are more almost 100% cheaper compared to two years ago.
In fact, even though the o1 model is more expensive, it is still somehow cheaper than GPT-4 when it first came out. And of course, o1 is way more powerful. Today, we're taking one step further with prompt caching.
[ Cheers and applause ] Many applications pass the same context over and over again to the model. For instance, passing a code base, passing product information, or long multi-term conversation history. So starting today, we'll give you automatically a 50% discount for every input token that the model has recently seen.
That's a request that we've heard from many of you and we excited to make it available to all developers. And that happens automatically starting today. There is zero change for you to do in your integration.
We'll have a talk later today, a breakout about optimizing cost and latency and we'll talk about pump caching in way more details. I have something else to announce on the topic of making AI more affordable. A big trend in the industry right now is model distillation.
With distillation, you take the outputs of a large model to supervise to teach a smaller model. And so today, we are announcing our own model distillation tools. You can now easily distill the outputs of o1 and 4o to fine tune a 4o mini all within the OpenAI platform.
The goal is to bring even more intelligence to small models at a low cost for you. Concretely, two new tools are going to ship to enable distillation; store completion and evals evaluations. Store completion is going to let you store your interaction with the models directly on your OpenAI, your logs.
This will allow you to create super easily fine-tuning data sets on OpenAI. And we're also shipping an evals feature in public better. So you can run custom evals on the OpenAI platform to test how well your fine-tuned model perform against a large model.
The goal with all of this is to make it much easier for you to fine-tune and evaluate a small model to get performance that sometimes may even match o1 on specific tasks. We are shipping our model distillation tools starting today to all developers. So feel free to start testing right away and we'll have a deep dive later on on the distillation breakout session.
All right, we covered a lot of ground today; our new all one reasoning models, the Realtime API, vision fine-tuning, prompt caching and model distillation. At last year's DevDay, we talked a lot about our ambition to build AI agents. But model capabilities were often getting in the way.
Standing here today, I feel the new o1 models take us one step closer to enable true agents. And I feel truly, truly grateful to have the opportunity to invent the future together with all of you. But we have many, many steps to get there and your feedback is critical to build this future together.
So please, please, before you leave today, say hi to any of us on the Open AI team. We would love to hear what you're building. Before I leave, I want to give you a quick look of the agenda for the rest of the day.
We've got a series of breakout sessions and developer community talks. You should really check them out. And of course, we know DevDay is not DevDay without Sam.
And so we will welcome Sam and our new CPO, Kevin, for your Fireside Chat at 4:00 PM today. If you want to ask questions to Sam and Kevin, drop questions, there is a physical box right outside the room, they would love to take it. Thank you.
Have a wonderful day and let's go build this better future together.
