All right, we're on day four of the 12 Days of OpenAI, and they've already released DALL-E 2 Pro and Sora for everybody. I've been playing around with that, but now we're on day four and we have some new announcements, some new features

for Canvas. And if you're not familiar with Canvas, it's the feature that allows you to have a separate tab in ChatGPT where you can actually collaborate with ChatGPT inline. So let's watch the livestream together, and I'll give you my

thoughts. Thanks to CodeBuff for sponsoring this video. They are incredible code generation powered by AI, straight from your command line. And I'll tell you more about them in a little bit. 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 are 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.

Yeah, so obviously they did not expect the amount of demand that they got for Sora. So hopefully they get it working better than it was yesterday. Even last night it was working a lot better. You can actually see the progress in a video you create, so definitely excited to

continue to try that out. I'm going to keep posting my Sora videos to Twitter, so if you're not following me, please do @MatthewBurman. Let's keep watching. Today, though, we're talking about Canvas, which is a product that we've had in

Beta for Plus users for the last couple of 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. Okay, that is huge! I have not seen that anywhere else—being able to actually create Python code using ChatGPT and then execute it within the ChatGPT environment, and then allowing it to

essentially iterate on the output. That has a ton of potential. I'm very excited to hear that, and it's cool that they're releasing Canvas to everybody as well, whether it's text or graphics. 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, try it out. Yeah, let's check it out. Okay, 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 tools available, one of which is Canvas. So I'm going to click that. All right, so that's new—being able to actually select the

Tools that ChatGPT uses now. I think one thing they might do—and this is just a guess, I don't have any information—is open up that little drop-down, the tools, to developers to actually build tooling for ChatGPT directly. So you can essentially create custom tools

for ChatGPT to use. Just like this, thank you to the sponsor of this video, CodeBuff! I'm really excited to tell you about this. They are a YC company and are really changing the way that you use AI for code generation. Think of it as a

better code generation experience than Cursor, but straight from your command line, and it couldn't be easier to install, either. I've played around with it, and it is as simple as `npm install`, and you're off to the races. And once you do that, you open up any of your code

repositories that has Git already set up, so here it is, and you simply type `codebuff` and hit Enter. Then it opens up a command-line interface, and you literally just type any changes that you want to make. You can do it across your entire codebase; it has the context for your

entire codebase. Super easy, super nice to use. And here are a few things that you need to know about it: You can build real products with CodeBuff. This is not something where you're building just toy projects; these are full-stack applications. They use more LLM calls,

which means higher-quality code. They're using the most cutting-edge models from both Anthropic and OpenAI. And CodeBuff can automatically document your code's unique quirks and nuances. And CodeBuff does not store your codebase; they just pass individual files along, relevant to

Whatever your prompt is, open AI and Anthropic as needed to generate the code. They were even able to create a Minecraft clone easily. Look at this! So, build your project using natural language with Codebuff. If you use my code, you get an additional 500

credits. And once again, I want to say thank you to Codebuff for sponsoring this video. I'm really proud to be working with them. Check them out: codb.us. 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, and 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 we 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, so um, let's take a look at what ChatGPT made for us: the silly elves. All right. So again, if you're not familiar with canvas, which you probably are

Already, but on the left side is the chat with ChatGPT. This is the conversation you're having with the AI, and then on the right is the output of what you have requested. So, in this instance, he asked for a Christmas story about silly elves.

So you could see on the left side, here's the original prompt; here's the response from ChatGPT, and then this little box is what is appearing on the right side, and this is the actual story. So again, like he said, you can either do creative

writing, you can also do coding, and it's nice to have these separate windows because it definitely cleans up the interface and just makes it easier to interact with ChatGPT directly and then to actually edit and adjust what you want inline. So let's keep watching.

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, 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. Um, actually, otherwise, this is 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. So this is really interesting and kind of gives a peek at how OpenAI is thinking about their tooling, and it's not just a model provider, and I think that's a really interesting way to view

OpenAI, just like Sora, they're not just giving you this text-to-video model; they're actually giving you the tooling to be creative as well. So, as these tools evolve, they are not only going to be a model provider, but they're going to

be an AI tooling company as well, so giving you all these creative tools to create different types of media and different types of code and whatever you need. But the point is, and something that I've been saying for a while, OpenAI can't just be a model provider; models

are becoming commoditized. So they have to have the hardware, which they do; they have to provide the models, because they do; but now they're also providing the tooling, which I think is a really good decision on their part. But as a developer building on top of OpenAI, you should

probably be considering the platform risk as well of building on top of OpenAI. And so if you're considering, 'Hey, should I use an OpenAI model or should I use an open-source model like Llama?', that should be one of the many factors that you consider, which is platform risk,

meaning, is OpenAI going to build what you're building on top of? 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; you 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 would 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 chat.

TBT: Suggest edit; ask ChatGPT for feedback about what we might want to change. So all of these features have been available; these are not new features, and I've covered them in previous videos, so I'm going to skip over this part a little bit. 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.

and ChatGPT. Because I put in a substantial amount of text, we now have a new button in the upper right that says "Open in Canvas." If I click this... okay, that's really cool! So rather than having to explicitly say, "Hey, I'm going to do

this thing in Canvas," or "Can you pop open Canvas?" and then pasting it in, just by pasting it in, there's the button that's going to appear, and you can open up Canvas and start collaborating directly on that large piece of text. 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 ChatGPT. And by the way, that simple thing that he did

the bolding of the title—that's not AI; that is a simple feature in a word editor, which, again, that's where it seems ChatGPT is heading. 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 composer. Exactly. So I'm going to go in here and I'm going to ask, uh, ChatGPT:

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 ChatGPT is giving you to your own work because it would all be in line in the chat, but not with Canvas. ChatGPT is able to select specific portions of your work and leave comments in line, which makes it all

right. That's really cool. I think this isn't a new feature. Correct me in the comments if I'm wrong, but I remember this feature already being there—at least being able to have comments in line—but this is really cool. You're collaborating with AI on this document.

And it's like Google Docs where it just leaves a comment, and obviously you can click apply or reject. All right, I'm going to skip ahead just a little bit because this is too much in the weeds. Let's get to the next feature. All right, next,

they're going to show off some new features with code editing, and that's more what I'm interested in. So let's check it out. But in addition to writing in ChatGPT, 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. Got it. So I have some example code that I've been working on locally. 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 ChatGPT. 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

ChatGPT has recognized that this is Python code. I get Python syntax highlighting and even some basic autocomplete. All right, so already you can see again where OpenAI is going with their features. Now, if I were a cursor, I'd be a little bit nervous, and even VS

Code: I'd be a little bit nervous because it seems like they're bringing all of this functionality into native ChatGPT. If we look in the upper right corner, we can see "Run". So we talked about that at the beginning of the video; you're going to be able to actually

execute Python code right from ChatGPT. This is a huge unlock for developers, and so now you have this native code editor. Obviously, it's not the most sophisticated yet, but you have this native code editor directly in ChatGPT, and you can work with ChatGPT on the

left side to edit your code as necessary. Let's keep watching. I know this code isn't working, so I'm going to ask ChatGPT, "Can you help me debug this code? What might be going wrong?" ChatGPT sees the code, looks like it identifies that `plot.label` does not

exist in `matplotlib`. Um, it provides an example of 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 the canvas. So in the upper right, we have

this button titled "Run". Let's see if ChatGPT finishes. We'll click "Run" and see what happens. Click "Run", and it looks like right away, actually, we see that... all right, that's so cool! A little console pops up from the bottom; it feels very

familiar. And although all of this looks very basic, there's a lot going on in the background. Let's keep watching. That `label` is not defined, so ChatGPT was correct about that. We see the exception, and right away, anything in the console is shown to ChatGPT, so ChatGPT sees this

Exception, and it actually leaves a comment on it. It looks like this: 'corroborates what 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.

Uh, this is exactly the kind of error I get all the time—um, just like forgetting the API for something. 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

ChatGPT fix it for you. So when I click 'Fix bug', we show ChatGPT the exception and your code, and it's able to edit the canvas to make that correction. That's amazing! Again, this feels just like Cursor. This is essentially Cursor, but directly in ChatGPT.

And eventually, I'm going to guess they're going to allow for entire directories of code to be uploaded. And then, what's the difference between that and VS Code or Cursor, whatever you're using? And so it really becomes clear the direction that OpenAI is headed: they

are becoming a tooling company—a software tooling company, but AI-native. That was pretty fast. So maybe we want to see like exactly what ChatGPT 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! We really fast have our line chart now rendering in the console. So can you explain a little bit about exactly what happened, because uh...so cool, I think.

Right now, it only supports Python, but there's no reason it can't support other languages in a containerized environment. So very safe for OpenAI Python execution; it happens 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.  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 where there's like 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...um...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. Um, I like that. ChatGPT wrote the code with an error and then already fixed it. And that's also the power of agents being able to have multiple versions of AI work with each other, fix each other's issues.

And it's all doing that in real time, so very cool. Use a single color per flow. Let's try it. Let's try it. This is kind of awesome because 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, uh, and then I will try again and then, uh, go through this loop and try to read the inscrutable Python errors, 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, 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

refine. So those are some of the new programming features in Canvas. 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. By the way, do you use custom GPTs? I have a handful that I use pretty frequently, but not much beyond that. And if you have cool custom GPTs, drop them in the comments below or just describe what they do, because I'd love to learn more.

about how to use custom GPTs, custom context, and a lot of really distinct and specific behavior for whatever particular task 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 uh, 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 uh...

So you're also a part of North Pole operations? Uh, I can neither confirm nor deny

this. All right, if Santa comes out with a Frontier Model, I'm going to be very suspicious. So um, Alexi, I happen to notice 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 figure um, maybe we could use that as an example here. 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 Lexy 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 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... Santa is admitting that might even be beyond his capabilities in the toy factory. So sounds like maybe you shouldn't get your hopes up for that one. Yeah, I don't know how you put that one

on the stocking either. So um, but this is a pretty great first draft, and you know, Santa could come in here and, you know, change this up as much as he wants, make edits. Um, all right, so that's it. In custom GPTs, you can now use

canvas. Very cool! So those are the announcements for day four of OpenAI's 12 Days of OpenAI. Very cool updates for Canvas. Obviously, not as big of an announcement as DALL-E 2 or Sora, but still very cool quality-of-life updates from OpenAI. And once again, I want to

Say thank you, thank you to CodeBuff for sponsoring this video. I'm really proud to be working with them. Check them out at cbu.edu

ChatGPT Canvas Gets Major Updates - Code Execution!

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

Matthew Berman

12일간의 OpenAI: 4일차 공지 사항

새로운 캔버스 기능이 발표되었습니다.

이제 모든 사용자가 캔버스를 사용할 수 있으며 기본 모델에 통합되었습니다.

캔버스 내에서 Python 코드 실행이 활성화되었습니다.

사용자 지정 GPT와 캔버스 통합.

캔버스 데모: 크리스마스 스토리 만들기

캔버스를 사용하여 크리스마스 스토리 만들기.

ChatGPT에서 사용 가능한 도구를 선택할 수 있는 버튼이 추가되었습니다.

나란히 보기: 왼쪽은 채팅, 오른쪽은 캔버스.

캔버스를 사용하면 창의적인 글쓰기와 코딩을 모두 할 수 있습니다.

Canvas를 사용하여 에세이에 대한 피드백을 받습니다.

"큰 텍스트 입력을 위한 '캔버스에서 열기' 버튼.

ChatGPT는 에세이에 인라인 댓글을 제공합니다.

코드 편집 및 디버깅을 위한 캔버스

Canvas의 프로그래밍 기능이 개선되었습니다.

구문 강조 표시 및 자동 완성 기능을 갖춘 Python 코드 편집기.

ChatGPT에서 파이썬 코드를 직접 실행할 수 있습니다.

ChatGPT로 Python 코드 디버깅하기.

빠른 코드 실행을 위한 웹어셈블리 파이썬 에뮬레이터.

사용자 지정 GPT에 캔버스 기능 제공.

예시: 예: 산타가 편지에 응답할 수 있도록 사용자 지정 GPT 만들기.

사용자 지정 GPT는 캔버스를 사용하여 산타에게 보내는 편지의 초안을 생성합니다.