[Music] Hey everybody, and welcome to day four. I'm Kevin, I lead product at OpenAI. So yesterday we launched Sora, and the demand has been off the charts. Uh, the team has been working late last night; they're in already this morning.

We are doing everything we can to get you access to Sora as soon as possible, and we can't wait to see what you create. Today, though, we're talking about Canvas, which is a product that we've had in beta for Plus users for the last couple

months. Canvas allows you to collaborate with ChatGPT on writing and coding that goes beyond simple chat. So we're launching three things today: first, we're bringing Canvas to everybody and integrating it into our main model so it just works; second, we're enabling you to

run Python code inside your canvases and uh, so you see the output, whether it's text or graphics; and third, we're bringing Canvas to custom GPTs so you can imbue all of your custom GPTs with the power of Canvas. So with that, let's

dive in. All right, thanks Kevin. Hey everybody, my name is Lee, and I'm an engineer helping you collaborate with ChatGPT. Hey everyone, my name is Alexi, and I'm also a product engineer working on ChatGPT. So Lee, are your kids excited

about Christmas coming up? Oh boy, yes. Um, you know, I thought it'd be fun if we could maybe make a story for them—a Christmas story—and use Canvas to try it out. Yeah, let's check it out. Okay, so, so if I go to ChatGPT, you'll notice

Starting today, there's a new button in the composer that lets us see all the available tools; one of which is Canvas. So I'm going to click that, and now anything I type here is going to create a canvas. So let's write that story... um, help me write a Christmas story

about silly elves for my kids, and I'm going to send that off to ChatGPT. It's going to get started right away on making a canvas for us. So, um, there are a handful of things that we should explain what's going on here. The

most significant thing that's different is you have the side-by-side view where, on the left, is the chat as you're used to it, and on the right is the canvas. And in the canvas right now, ChatGPT is busy writing the story for us. This is pretty

different from what would happen if you wrote this kind of prompt before. You know, ChatGPT is good at doing this sort of creative writing work, but it would write it directly in-line in the chat. And sometimes it's hard to piece out what's the document and like what's the

preamble, or what comes after, um, or especially if you wanted to make any changes along the way; things tend to get lost. But now you have this side-by-side view. So, um, let's take a look at what ChatGPT made for us: 'The Silly Elves in the'

Christmas Catastrophe.' No, not a catastrophe! In the snowy village of Tinseltown, nestled at the edge of the North Pole, lived the silliest group of elves you ever could imagine. All right, this is looking pretty good. Um, you know, I, uh, I could make some changes, I think.

This title is a little long, so I'm going to, you know, maybe take this part out. Uh, actually, otherwise, it's pretty good. So you can...so ChatGPT can edit, and you can edit sort of side-by-side. Yeah, so I can type in here, you know, I've been editing

the title. I'm kind of making it for my kids. Um, so this is a collaborative document, and just like you might be used to in other document editors, you can come in here, you can add text, you can bold text; you can do lots of things you

can do in other document editors. You can edit; ChatGPT can edit. Um, I could go back to chat and, just similar to how I would do in any other um, ChatGPT workflow, I could give it feedback; I could ask for changes, and it's going to

come in and make those changes directly into the canvas. Um, but rather than show you that, I thought it'd be interesting to show this button in the bottom corner, which has a handful of commonly used shortcuts. That's based on things that we know people do in ChatGPT: suggest edits, ask

ChatGPT for feedback about what we might want to change about this; adjust the length—maybe make this longer or shorter. Maybe I want, you know, a shorter bedtime story. Reading level will change this to make it, you know, more sophisticated language or simpler

language in case I wanted to, you know, make it easier to read for my younger kids. Add final polish is really helpful, especially when you bring your own content: check for spelling mistakes, grammar things like that, and then add emojis. That sounds fun!

Emojis? Let's try it. Um, so if I click 'Add emojis,' I've got some options about where I want to put them: words, sections, list. Let's try 'words.' I click that, and ChatGPT is going to, you know, get to town putting emojis into my doc. Look at that!

So, um, what you're seeing here is ChatGPT now making a revision to the document in Canvas. It's starting at the top; it's writing a new version of the doc and it's sort of streaming in on top of the previous version, just to make it a

little bit easier to see, in context, what's changing. And alongside this, in the chat, you know, it's—you're seeing that it's editing, and you're noticing that, uh, you know, we have our instructions back and forth. Um, but otherwise, it's not getting sort of

cluttered up with the different revisions that are happening. Kind of looks like when my 8-year-old gets a hold of a phone. Yeah, I think your kids are going to love this, Lee. There's definitely a lot of emojis that I put in

here that did a pretty good job. Um, okay, so that's how you get started with Canvas. You know, Canvas will trigger in a couple different scenarios. You can ask ChatGPT to just start collaborating with you on a piece of content, whether it be writing or code, and it will open a

Canvas for you automatically, or, like I just did here in the interface—I'll show you again—is this button in the composer to view the tools. You just click 'Canvas,' and that's going to create a new Canvas every time. Amazing! So, in addition to storytelling, a lot of people use ChatGPT

to help with their own writing. Learning how to be a good writer is hard, and today we're really excited to showcase some new features in Canvas to make it easier to get feedback on your work. So, in my spare time, I'm a bit of a physics

enthusiast, and so I have an essay here that I've been working on, titled "Santa SLE: Exploring the Role of Dark Energy in Reindeer Propulsion." I'm going to copy this from my text editor and paste it into the prompt area in chat

GPT. Because I put in a substantial amount of text, we now have a new button in the upper right that says "Open." And if I click this, the content from the prompt is put into the Canvas. And this is the same editable surface that Lee was demonstrating. So could you like bold

that title, maybe? Yeah, so I can go in here and make a change if I wanted to, um, before I send it off to chat GPT. So this hasn't been sent yet. So when you made that bold change, what just happened? Correct. Yeah, this is all still a draft

document. So, until I send a message, I can edit it as much as I want, um, and ChatGPT hasn't seen it yet, and it's attached to the comp exactly. So I'm going to go in here and I'm going to ask uh,

ChatGPT, "Can you leave some comments on my essay from the perspective of my physics professor?" Nice. Okay, so, um, maybe it's going to give you some like technical tips. Yeah, hopefully we'll see what it comes back with. But previously, you know, it would be hard to connect what the

Feedback in the chat TBT is giving you about your own work because it would all be inline in the chat, but not with Canvas. ChatGPT is able to select specific portions of your work and leave comments inline, which makes it really easy to connect the suggestions to your work.

Yeah, I'm looking at these. These are pretty good feedback. Consider linking this introduction more explicitly to the scientific exploration of dark energy to immediately establish the physics connection. But it didn't just respond to that in chat; it selected "Santa, his sleigh, and his flying reindeer have..."

long fascinated people, blending mystery and magic." Okay, so it's like picking out sentences that it wants to give you feedback on and then writing the specific feedback per sentence. Exactly! And each of these comments, I can have an 'apply' button. So, clicking 'apply' would allow ChatGPT to go through and apply

the edit as it sees fit. And of course, I could then reject that edit if I wanted to. But since this is my own essay, I'm going to go in here and make changes on my own. So, um, I like this comment: "You may briefly elaborate on why dark

energy is mysterious to provide context." Um, I'm going to say, "We don't know why the universe expands, and we don't know why Santa is so fast in his sleigh." I'm curious about these cool...the mysteries of the universe. I'm going to dismiss this comment because I feel like I've addressed it. And when I'm ready to take my essay out of ChatGPT, I can click this 'copy' button and send it off to my professor. Great! I'm sure your professor

is going to love these changes. Now, this is some cutting-edge research here, but in addition to writing in Chat GPT, we know a lot of people also use it for programming, and we've made some really exciting improvements to programming in canvas as well. Let's check it

out. So I've also been helping Santa with some logistics for toy manufacturing, and I've been learning some data science tools in Python. So you're an OpenAI engineer, you're also a dark energy researcher, and you're helping Santa with logistics? Yeah, I mean, Santa could be

everywhere at once. We really try to as well here at OpenAI. So I have some example code that I've been working on locally. So this is my local IDE, but something's not working and I want some help debugging it. I'm going to copy that

and then paste it into Chat GPT. I see the same Open in canvas button. I click that, and now I have a canvas open, but this time it's actually a code editor. So Chat GPT has recognized that this is Python code. I get Python syntax

highlighting and even some basic autocomplete. Nice. Um, I know this code isn't working, so I'm going to ask Chat GPT, "Can you help me debug this code? What might be going wrong?" Chat GPT sees the code, looks like it identifies that `plot.lael` does not

exist in Matplotlib. Um, it provides an example corrected code. Um, we could trust this and take it back locally, but one of the new features we want to demonstrate today is the ability to actually run your Python code within canvas. So in the upper right, we have

This button, titled "Run," let's see if PBT finishes. We'll click "Run" and see what happens. Click "Run," and it looks like right away, actually, we see that "label" is not defined, so ChatBT was correct about that. We see the exception right away.

Anything in the console is shown to ChatBT, so ChatBT sees this exception and it actually leaves a comment on it. It looks like this corroborates what the...because "label" is not a valid function in Matplotlib; you might likely use "title" instead. This is great! I feel like

I'm always trying to learn new libraries, and uh, this is exactly the kind of error I get all the time—just like forgetting the API for something—and um, it's pretty amazing that it just picked that out right away. Yeah, and just like

comments in the document, there's a "Fix bug" button to let ChatPT fix it for you. So when I click "Fix bug," we show ChatPT the exception and your code, and it's able to edit the canvas to make that correction. That was pretty fast.

So maybe we want to see like exactly what ChatPT changed. Yeah, and there's this great "Show changes" button in the canvas header. Clicking that will give you an inline diff that should be familiar to people who use Git, to show you exactly what changes were made.

That's cool. Um, and "Show changes" works not just for code but for documents as well. Exactly. Yeah, works across both of the document—both of the editors. Um, so now that we have this fix in place, we can try clicking "Run" to see if it works now. And wow, that was fast! Really fast!

Have our line chart now rendering in the console. So can you explain a little bit about exactly what happened? Because, um, that Python execution happened very quickly.

Yeah, so within Canvas, we actually load a WebAssembly Python emulator, which allows us to load almost any Python library and run your code almost instantly. So you get immediate feedback; you can collaborate with ChatGPT effortlessly and really have a tight feedback cycle all within ChatGPT, including generating graphics.

So Santa, particularly, is interested in some Sankey flow diagrams to kind of track how toys are flowing through his... those are the ones with arrows in and arrows out. Honestly, I'm not even sure Santa asked me for this. Yeah, I'm trying to learn, so he's always challenging us.

I'm going to ask ChatGPT to add a Sankey flow diagram for Santa's toys. Please use appropriate colors; see if we get something, uh, Santa-themed, interesting.

So ChatGPT is going to go through and edit our code again, um, and now it's adding, example two, a Sankey diagram. Cool. So we got total toys delivered to children, distributed to stores, defective. Okay, so it's added a handful of lines of code, creating a Sankey, adding data, finishing it, and then saying 'show'.

Yeah, let's uh try running this. Oh no, uh, looks like there's been a problem with the colors, um, but ChatGPT has an idea of how to fix that.  Use a single color per flow. Let's try it. Let's try it. This is kind of awesome! This is exactly what happens when I'm trying to learn something new: I will go search for example code, I'll run it,

Inevitably, it won't quite work, and then I will try again and then go through this loop and try to read the inscrutable Python errors and figure it out for myself. Oh, there we go. All right, looks like maybe Matplotlib has some bugs

in its Sankey layout, but we got a graphic, so there we go. We got a graphic, and we can keep refining this with ChatGPT, even dragging this image into the prompt area if you want to show ChatGPT exactly what this image looks like to

help find. So those are some of the new programming features in Canvas. Amazing! All right, there's one last thing that I want to show, and that is bringing Canvas into custom GPTs. So GPTs, in case you're not familiar, are a way to give

ChatGPT custom instructions, custom context, and a lot of really distinct and specific behavior to whatever particular task that you have at hand, and then you can share them with other people so they can benefit as well. There's a GPT store full of these things; they're really

interesting, and I decided to make one for Santa. You know, Santa gets all these letters and he has to respond to all of them, and I figure that's got to be a lot of work, and it might be helpful if he just had, you know, just like a first

draft, something to start with. He could decide if he liked it or didn't like it; he could edit it, he could go from there and then send on his… So you're also a part of North Pole operations? I can neither confirm nor deny.

This is alright. If Santa comes out with a Frontier Model, I'm going to be very suspicious. So, um, Alexi, I noticed that you wrote your letter to Santa and left it out on your desk.

Oh, I meant to mail that already. I can't believe you found it. Yeah, um, well, I snapped a picture of that and I figured, um, maybe we could use that as an example here. So let's bring that up if I can figure out how my mouse works. Okay, there we go. So we've got our

picture into ChatGPT. Let's open it up real quick just to get a look. A bike, a GPU, and some dark matter that tracks. Yeah, that's right there. That looks like my letter. Okay, um, you know, I also figure Santa has some context on, you know,

whether you've been naughty or nice. That's uh, you know, one of the things he keeps track of. And in addition to taking a picture of the letter, we probably want to add that in. So I'm going to let Santa's drafter know that Alexi's been nice this

year. Really appreciate that, Lee. Yeah. Okay, Alexi has been nice this year. So let's send this off to our custom GPT and um, hopefully, we'll get back a canvas. All right, here's our custom GPT starting to write out our letter. Nice! It says, "Dear Alexi, Ho ho ho! Thank you for

your letter. It put a big smile on my rosy-cheeked face. Let's see what we have here. First off, a new bike—excellent choice! All right, looks like you're getting that bike. Great H100 GPU! Mrs. Claus tells me it's quite the powerhouse."

Running AI models. The elves in the tech workshop are tinkering away on that one for you as well, so that's promising. And dark matter? He's admitting that might even be beyond his capabilities in the toy factory, so it sounds like maybe you shouldn't get your hopes up for that.

Yeah, I don't know how you'd put that one in a stocking either. But this is a pretty great first draft, and you know, Santa could come in here and change this up as much as he wants, make edits as much as he

wants. And then, as soon as he's ready, you know, write this out in his characteristic handwriting style and send this letter back out to you. Hopefully, this will save Santa some time, and so we should talk a little bit about what's actually happening here.

You might have noticed that I didn't click the canvas button; I didn't mention to create a canvas. It just made a canvas. So this is an example of ChatGPT determining that a canvas would be the right thing to do, but how did it know to do

that? Yeah, well, um, when I built this custom GPT, I put that in the custom instructions. So, do you guys want to see how I made this? Did you show...? Yeah, yeah, please. So if we go up here into the GPT, excuse me, and hit 'edit',

and that should take us to the configuration page. So this is how you set up a GPT; it's really pretty straightforward. You can get a picture and a name and a description, so people in the store, if you choose to share it, can see what

It's about the important part here is the instructions. So I wrote here your tool for Santa to help write first-draft responses in Canvas to the letters that he gets. Every response should be written in Santa's characteristic style and humor, and your job is to write responses to

the letters using the Canvas tool. So in the instructions, I told it to use the Canvas tool, so I'm going to expect that it will use Canvas pretty much for every—um—every input thing that I put in there, which is exactly what I want this to do. That's

going to be the right tool. In addition to that, down here at the bottom is a list of checkboxes on your capabilities. So this is the set of tools in ChatGPT that are available in GPTs, and there's the new one here, Canvas, which is uh, you

know, starting as of today, uh, will be on by default for new GPTs. Existing GPTs, we're going to leave this off, that way we don't disrupt the behavior of any existing GPTs that people have created. Um, but if you want

to add Canvas to your existing GPT, it's very easy to do that. You just go into this configuration screen; you'll see that checkbox, and you can check it, and then you'll have Canvas available in your GPT. Amazing! Amazing! Yeah, so this is Canvas.

That's the handful of things we wanted to show you today. We've got Canvas rolling out to all web users, regardless of plan, as a default experience in 4.0. So whether you're on a free plan or a paid plan, you'll have Canvas. We've got using

your own content in ChatGPT or using your own content in Canvas. We've got code execution, and we've got Canvas and custom GPTs. It's awesome, so we're super excited to see what you do with it. Like, uh, like Lee said, we'll be rolling this

out today. Uh, before we go, though, let's end with a Christmas joke. Okay? Are you ready? Yeah, we're ready. All right. How does Santa take photos? All right, let's hear it. With a North Pole-aroid! All right. See you tomorrow, everybody. [Music]

Canvas—12 Days of OpenAI: Day 4

OpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity.



Canvas—12 Days of OpenAI: Day 4

Canvas—12 Days of OpenAI: Day 4

Scripts