Major ChatGPT Upgrade! | "Canvas" AI Features HANDS ON

OpenAI actually updated the chat GPT interface. Now, this is their first major overhaul since Chat GPT was originally released. And honestly, it's about time. Now, don't get me wrong, the old Chat GPT interface is nice, it's clean, it's easy to interact with. We've got our prompt bar at the bottom. We can upload files, we can send our messages off. But interacting with the large language model in this old interf definitely leaves a little bit to be desired. Let's say, for example, this first paragraph, we want just that rewritten. Well, the only way to do that in this old interface would be to prompt it all over again and say, I like the story, but please rewrite the opening. And then it would have to regenerate the entire thing. Instead of only going up here and changing that first paragraph, we wanted changed. And this is just the first of a lot of micro improvements that you're going to see with this brand new ChatGPT interface. It's a lot of little UI tweaks and nuances that for me are a long time coming and very, very much a warm welcome. So let's dive into the blog post. What is new? They're calling the new version of this UI Canvas and specifically it will be accessed by selecting in the model selection tool chat GPT4O with Canvas. So I don't think this is coming to the brand new Zero models just yet, which is a little disappointing because those are like flagship brand new types of models. Why would your new user interface not work with the brand new models? I don't know. Anyways, here's your first sneak peek at this. The keen amongst you are going to notice that this very, very much looks like the interface that Claude currently uses. So Anthropic is a company that makes their own versions of large language models, a ChatGPT competitor, if you will. And this UI design at first glance is very similar. And example, we're uploading some design notes and we're asking to draft a short blog post from those notes. It then goes, okay, here you go, and then opens up a brand new tab on the side and writes out the blog notes and then just gives you an end reply saying, here you go, here's your new document. This is a lot cleaner, it's a lot easier to manage and work with. Essentially on the side here we have our little chat, which is that macro overview, and then you can click to open each individual piece and get a micro view that would be considered the overall canvas. Right. So in OpenAI's own words. Here they're introducing Canvas. It's a new interface for working with Chat GPT. It makes writing and coding projects a lot easier and more simple. Canvas will open in a separate window. I still have yet to understand exactly what that means. I think they're just talking about this as its own separate window, but allows you and ChatGPT to collaborate on a project more seamlessly. This is still an early, early beta of Canvas, and we'll talk about who gets access soon. But if you're paying for Chat GPT in some capacity, you will get access to this instead of a basic chat conversation. This is creating and refining ideas side by side. And so as they state, Canvas was built with GPT4 omni, which is kind of their, you know, main model. It's been out for a long time. The best version of the GPT4 architecture. And right now, while it's in beta, it's manually selected in the quote unquote model picker, which I don't think is a great right name. I prefer Model Selection tool, Model Selection dropdown, but to each their own. Anyways, yeah, Chad GPT plus, which I'm a part of, but the rollout hasn't happened to me. It's supposed to just take place over the course of this day. So in this video you should see me using this new model because I'm gonna wait until I can actually use it to produce a full video. And of course the team users and enterprise users will all get access. So basically everyone's gonna have access by next week. Once Canvas is out of beta, it should available to Chat GPT free users. So this is basically them explaining to you why the old chat interface kind of sucks. Sure, it's easy to use and works well for a lot of tasks, but it's limited when you want to work on projects that require editing, revisions and deeper work overall, they continue on to tell us that ChatGPT can better understand the context of what we're trying to do in the overarching project and trying to accomplish. At the end of the day, you can highlight specific sections to indicate exactly what you want ChatGPT to focus on, like a copy editor or code reviewer. And again, this is a huge deal for me because it's one of my biggest pet peeves with these basic, bland chat interfaces. A lot of times we're trying to do in depth, deep work and I need to ask very specific things to be changed. Highlighting and allowing to focus on those specific areas is going to be an absolute game changer. For me, it'll be able to give you inline feedback. It'll give you suggestions based on your highlighted selection, but also with the entire project in mind. What's nice is that you can also directly control the project so you can edit the text. And that's another huge deal. In the old interface, you can't just go back in here and edit the text. You'd have to copy it and put it into like a doc or something and then resend it, etc. Now you can just directly, you know, essentially make the whole project inside of that interface. And there's also a menu of shortcuts that you can ask Chat GPT to adjust. So that could be writing length, debugging code, or some other useful actions which we're going to go over in this video. What's nice is they have a saving feature so you can actually restore store previous versions of your work just by using the back button. Now in my mind, guys, one of the main reasons that OpenAI is refreshing the interface for chat GPT is because now they're probably realizing that other competitors, such as Google, for example, already have a really nice interface. Google Docs, Google Slides, and they could pretty seamlessly integrate their Gemini large language models into those products. And it's already a nice interface again to work on larger projects. So this is them, I think, trying to get ahead of the curve a little bit in regards to that and keep people on ChatGPT versus switching over to something like Google Docs. With Gemini built in, Canvas can also automatically open when ChatGPT detects a scenario in which it could be helpful. You can also just straight up prompt it to use a canvas, which is pretty useful. And I like that, you know, it's, it's sort of integrated directly in that way where it's intelligent enough to use the UI tools that it knows you'll probably need for a specific task. See, that's something that I wouldn't expect Google to integrate directly anyways. Here's a little overview of some of the writing shortcuts suggesting edits that is incredibly useful. This allows you to branch off into new ideas, inline suggestions and feedback saying, what if we took this direction with a story, what if we took this direction with a code instead? And it will give you multiple of those as well, sort of allowing you to branch off and go into these deep, I don't know, thinking holes or rabbit holes, which could essentially transform your project in a really creative way. It's utilizing that creative might, that ability for the large language model to Seamlessly and instantly offer other suggestions and feedbacks, providing you with another perspective. These large language models are very powerful, and sometimes a user interface update might just bring that power out that much more. You can also adjust the length. This is classically an issue. Again, with large language models. It'll either be too long or too short for whatever you're trying to do, so that is a welcome and improvement. The reading level is really interesting to me. It adjusts it from kindergarten to graduate school. I think this could very well be helpful for writers. I'm interested to try this one myself though, to see if I can come up with any nuanced scenarios in which I would want to use this. They also have the ability to add some final polish, checking for grammar, clarity, consistency. I already assumed a lot of the writing at least was pretty well polished with Chat GPT, but maybe it can be even better in some ways we didn't know by using this shortcut. And you can also add emojis. And this one's kind of more just gimmicky in my mind. Might be useful for like if you're making a lot of social media posts, something like that. If you do a lot of that kind of work with ChatGPT, I might use it a little bit, but that's, that's in my mind, kind of just like an extra little gimmick. Here is a nice preview for you guys of this new interface and how it's going to look. And again, we're going to test it out in today's video. So yeah, you'll have this button where you can have all of your different little, you know, shortcuts come up, add polish, add emojis, etc, etc. And they're showing the highlighting and you can apply certain things and it's really nice. Yeah, it can just generate or regenerate only snippets that you highlight. That is crazy big. Just change the title, just change this portion and the ability for you to re edit and essentially interact with Chat GPT as needed on a project. Now there's also a lot of coding and it's become a larger and larger part of using large language models as a whole. If you were to hold up a pie chart and say, you know, what percentage of text that is generated by large language models is just code, I think you'd be shocked to see that it's more than you'd think. How many of you would be willing to say that out of all of the generated text that large language models currently produce, at least 10 15% of it is code, maybe even more. Coding is literally the language of computers and it is super powerful. Anyways, it's an iterative process. It can be hard to follow as all of the revisions to your code happen in, you know, the classic chat interface. Coding in the classic chat interface basically sucks. All right, I've tried it. I'm not even a coder and it annoys me. But Canvas apparently makes it easier to track and understand the changes. Chat GPT has its own set of coding shortcuts that are reviewing code, adding logs, adding comments, fixing bugs, porting to different languages. That's all really cool. I will say though, I'm disappointed to see that we don't have it in the new 01, because the new 01 models are really good at coding. They're the best for coding by a long shot, especially for someone who has no idea how to code. Having a model that is that much smarter is that much more useful to me. But all of this does look like it could make it a lot easier for even newbies like me to use code with large language models better. Anyways. We, you know, we have the, the typical inline suggestions to improve code. You can actually now directly insert print statements to help you debug and understand your code, which is pretty cool. You can also add comments to make it easier to understand and easier for the large language model to understand. Fixing bugs, which is, you know, classically a huge part of coding. It's like most of coding. I don't know if coders are gonna be mad at me for saying that. That's probably not even true. Anyways, fixing bugs is something you have to do a lot with these large language models as they code things importing to a language. Not sure how useful this is. Again, I'm not a super experienced coder or anything like that, but if you already have a big repository of, you know, maybe Java code and you wanted to port it to Python or C, this could be pretty cool. And here's a little, you know, snapshot image of what that looks like. Obviously the code is going to be over in the side and again, you're talking to it in that same exact layout. This, to me, is probably one of the most interesting insights out of this entire blog post. And it's the fact that they actually trained the model to become a collaborator. I thought that they just would have made an interface and some basic prompt tweaks for GPT4O and it would work with Canvas, but it looks like they've actually retrained it and this actually ties into knowing when to open a canvas, for example, making targeted edits or knowing when to fully rewrite things. Apparently they needed to train it to help it better understand the broader context of the project, to provide more precise feedback and suggestions. These are the core behaviors, obviously triggering the canvas for writing or coding. Again, these are actually two separate things generating more diverse content types. Like to see that because I think everyone's kind of sick of that generic bland. Let's delve into this style that ChatGPT typically writes in, targeted edits, which I definitely understand. Back in the GPT three days, they actually had separate models for doing this, rewriting documents and inline critiques. It's really interesting that these last two, again, it's like these new core behaviors that it had to be trained to learn. To train it, they actually did some evaluations. They use synthetic data generation techniques like distilling outputs from the new O1 preview to actually train the model for the core behaviors, which is nice because they don't really have to rely as much on human generated data. Then they actually show us a performance graph, which I wasn't expecting for a blog post about a UI change. And this apparently is supposed to represent how they improved correctly, triggering the Canvas decision boundary, reaching 83 and 94% respectively, compared to a baseline 0 shot GPT4O with prompted instructions. So this is supposed to, I think, be the success rate in triggering the canvas decision. I don't know if I'm just like missing something here, but to me this graph doesn't seem especially prominent or important. Again, I might be missing something though, so let me know in the comments if I'm just being stupid. This is more interesting to me. This is a full rewrite of the canvas versus a targeted edit. You'll see that the targeted edits here perform better than a baseline prompted GPT4O by 18%, which is definitely good to see. Now this one is for training the model to generate high quality comments. This is human evaluations assessing the canvas comment quality and accuracy. Their canvas model outperforms the 0shotGPT4O with prompted instructions by 30% in accuracy and 16% in quality. So a little bit of a bump. They end the blog post by saying Canvas is a new approach in the first major update to ChatGPT's visual interface since they launched it two years ago. Now it is wild that that was two years ago, man. To me it feels like Chat GPT launched like eight months ago or something. Anyways, guys, as I mentioned, it's a rollout over the course of this day. So I think this video is probably going to be going up tomorrow because currently I don't have access. So I'm going to have to come back to my ChatGPT once it has the new canvas and then we can explore and actually get hands on this thing and see what it's all about. For me, this could take hours and hours of time, but for you it should be immediate. And welcome back. See, I told you guys it wouldn't take that long. Anyways, here we are in chat GPT I now have access. If you go to the drop down where you can change the different model, you'll see it's right under GPT4O and it's just called GPT4O with canvas. And if you click that, I can say, open me a brand new blank canvas. Let's try that. And then it just opens it right up. And to actually get into Canvas mode, I think I just click on it maybe. Oh, there we go. Okay, that worked. I have to click on the blank Canvas title and then we can just type anything we want in here. It's essentially again like a Google Doc or something. Draft me a quick idea list for ways in which I can dynamically interact with my YouTube viewers during a live stream. You know, this is like something I might actually do for my line of work. Will it know to put this right in the canvas? Okay, it does. And it just like appears there like it's revealing a ghostly artifact or something. Wow. Okay, cool. Live polls, Q and A sessions. Like in this example. I think this example that we're doing right now is ideation. Right, let's play into that. Let viewers guide the content or story of the stream by making decisions for you through chat votes. Ask ChatGPT. Expand on this. Could I connect it to my Discord server? And what is it going to do now? Okay, so it just highlighted and changed just that portion for me. Connect this to your server. Setting up a dedicated voting channel where viewers can vote on the next action. I see. Okay, cool. That actually worked really smoothly. Oh, and it auto recognizes each new line here on the side. You can see with the ui. If I zoom in here, you guys can see if I scroll down, it actually can recognize each of these as their own separate pieces. Because I think of this little dot here. The fact that it's a bulleted list. That's smooth, that's nice. But you know what, let's just delete all this. Can it just continue something like. Can I say Message to chatgpt complete whatever text is below. We turn it into essentially a completion model. My name is Gorg the Caveman. What if I just click now add final polish. Oh, okay. So now it's just trying to rewrite the whole thing with the YouTube live streams again. I see, I see. All right, let's highlight this instead. Ask Chat GPT and just say follow the above instructions. I know we're kind of of getting out of the way that OpenAI expects you to use this thing, and it's still trying to just continue with what we were doing before. Interesting. Okay, so it looks to me, guys, like you can't use it in that way. Good to know. Write me a list of paradoxes. We'll do the classic grandfather paradox. Oh, nice. The list is automatically updating. They definitely put some effort into making this feel seamless and clean. We'll now highlight just that one instead. Expand in great detail. It will now highlight, rewrite. Might be wondering why I'm doing this. The purpose is to try this. The reading level, current reading level. Or we could go all the way down to kindergarten. Let's see what happens. And it rewrote it for a kindergartner. Imagine someone goes back in time and tries to stop their grandpa from having kids. If they did that, they wouldn't be born. If they were never born, they couldn't go back in time to do it. It's like a big circle that doesn't make sen. Yeah, that's kindergarten level. Everything stays the same. Different worlds. Okay. Okay. Now let's highlight it again and we'll change the reading level to graduate school. And now more or less, it's like writing a research paper, right? Um, maybe not. This is kind of like a typical ChatGPT answer. Actually. It's pretty well detailed, though. Cool. Cool. Uh, let's try now to add emojis. Oh, there we go. You know what? I actually kind of like the way that it does this because it allows you to skip around a little bit easier, maybe throughout the test text because you can make visual connections to the language, I guess. Yeah, I mean, like. Like I said, this could be good for social media for sure. What happens if we re emojify it? Add more emojis. Okay. I don't know if it's necessarily adding any extra emojis. Doesn't seem to do much. Okay, we'll restore the original version that is graduate level, add some final polish, and I think this is just going to do like, bolding, etc. Etc. Right. Let's see what this does. Okay. Yeah. No, it just kind of reformats everything more or less. So it doesn't necessarily rewrite it, it just makes it look cleaner and nicer. I see. All right, let's do a new canvas. Now let's try working with images. I'm uploading some new screenshots of my brand new Discord server, which you should join by the way, if you want to stay up to date in all things AI. We'll say open a canvas, organize and highlight each of the brand new features from this Discord server. Okay, so it's taken a look at the images and now it's going to give us a breakdown of the Discord server. Like if I highlight this and then go into the talk the ChatGPT. Okay. It just deselects that. How do I get to the suggestions? That's sort of what I was interested in. Like this maybe Editor, explain. Give me some alternative suggestions to improve to make Creation Sharing more interactive. Okay. And now it's going to rewrite everything. Apparently it says that it's added several suggestions to make the Creation sharing zone more interactive. Okay, I see, I see. It just added these down here. Interesting. All right. I wish it popped out with its own little suggestion. Sidebar. Whenever you highlighted something, I think that would be pretty cool. But I understand how that could be a little bit expensive. Adjust the length of this all the way down to the shortest possible. How does that work? Is it still going to keep the information dense? Okay, not really. I see that's a little bit more streamlined though. At least for something like this. That actually works out well, I think. Can we make the reading level kindergarten now for this? How is it gonna change this to be more kindergarten level? It literally just does it like as if a child had written it. We have some cool bots to play with intros, say hello and tell everyone about yourself. I mean, yeah, though. I mean like a kid could actually read this though. That's I guess the whole point, right? Oh, okay guys, I think I'm kind of a knucklehead because I didn't realize that the first button down here is Suggest Edits. I just thought that was like the logo. I don't know what's wrong with me. Alright, Suggest Edits. There we go. Now it pulls up on the side. This is what I was looking for. This was like of the main features I was excited about. Okay, consider specifying an example interaction to make this description more engaging. You might want to clarify if there are any guidelines. Oh, okay. So it's it's like correcting the editing of the actual wording and then you can just simply, like, apply things. Okay, I am getting a better, more clear picture of what this interface is supposed to. Or how it's supposed to work. Work. So this was as if, you know, I'm trying to describe the Discord server for someone. That's the context. That's the background context that chat GPT is going with. Who is Matt Vid Pro? Okay. It gave me a basic breakdown. Open a new canvas and write a bio for this person. There we go. So now, since we're doing real writing, suggest me some edits for my own bio. Wow, that's crazy how fast it actually does all this, though. This is cool. This is cool. I like the editing is definitely the coolest part here where it actually gives you suggestions. You can apply them, specifying a unique aspect of the channel that makes it stand out from other AI focused content at a specific type of content that appeals to newcomers. Okay. Wow. And then we can just apply all of these. But looks like you can only change one at once. Something like that. There it goes. It does it pretty quick, though. Also, it's probably just lying and hallucinating about all of this stuff about me too. So this wouldn't necessarily be useful in the real world because I'd have to double check and make sure I'm actually doing all this stuff. Write me a story about a lemon character who is a total badass. It actually said badass lemon character. Okay. So it wrote me this little story. It looks like the main character is Lenny. What I'm interested in is if we can basically just change, like, one small aspect, change the character's name to something more interesting. Okay. Oh, and it can just edit it. All right, this is what I wanted to see. So it is able to go through and find each point where his name is mentioned and then just only change the name. So it actually didn't do a full rewrite. Very interesting. All right, suggest some edits now. Oh, cool. So this is obviously something the AI entirely wrote itself and now itself, improving a little bit more context on why this character is different from the other lemons to. Okay, so, yeah, apply that. Apply the context. A bolder, rebellious spirit. Cool. Will it give me different edit suggestions each time or are they, like, essentially tied to whatever is written? Oh, it definitely is different each time. So it's gonna give you different ones each time. You could essentially just keep having it edit and overdo and rewrite everything. You could have it sort of self operate in that sense, like apply automatic edits. That would be cool if it could automatically edit itself over time and then build something. Hmm. The gears are turning. The gears are turning. I could see the future of. Of how this technology could be app. This is pretty cool. This is definitely cutting edge stuff. Yeah. Let's do a brief anecdote or specific example. I think one problem for me so far at least is like for example, could I maybe just highlight this portion and then suggest edits on just that? Okay. Yes, you can. Okay, so that's good. So that's. That's one thing about this tool that bugs me a little bit. It's not entirely clear that you can just highlight just this section and then go in here and say only add emoji to this section, for example. Right, that's true. You can do that. But I wish it was a little bit more clear because when you just click this button and say, you know, adjust length or suggest edits to everything, it can be a little bit overwhelming. Cool. I'm definitely gonna have some fun with this interface. Alright, one last thing I want to try. I am going to copy the transcript of my latest video and we're gonna paste it in chat GPT. Did it get the whole thing? I think it did. All right, sweet. This is something very specific that I have not not been able to have chat GPT do for me. Create me some copy paste timestamps for this video, please. Does this new ChatGPT canvas have the ability to do it? Okay, this looks pretty decent. Yeah. You know what? It's. It's pretty decent. It just needs some editing obviously because the transcription is far from perfect. That is a lot better than it used to be. Again, this is something that ChatGPT personally used to really struggle with. It's a task that I've had it tried to do before. Not exactly copy paste though. We could just add emojis to this. Might actually be a useful thing for the emojis. Oh my God. Now it's doing all of them. I just highlighted this section though. I only wanted this one to be emojis. I guess it's gonna rewrite the whole thing though. Interesting. That might be a part of the training. Right. Where maybe if I highlight something, doesn't necessarily know that I only want to do it to that portion. It's kind of just guessing. I really don't know. It's hard. OpenAI isn't very revealing about the technical details behind all this. It's cool though. It's. It's different man. I I if Google or another company tried to implement something like this into docs or whatever their version of an editor is, it's not going to have those specific nuances like for example suggesting edits like this is next level stuff. Being able to highlight these portions and just immediately apply genuinely useful edits. It's a way better interface. I am going to try it for coding at some point. Probably not going to cover it in this first look type of video, but either way this, this is awesome. Like I said, it's a warm welcome, much needed Interface update for ChatGPT and I think they've done some really interesting OpenAI esque things with it. They've gone the extra mile, they've retrained the model to actually work with the interface and it seems very seamless and genuinely useful for improving your writing or code. I think something cool would be for them to, you know, in the dropdown menu where you can add emojis, add final polish, etc. It'd be cool if you could actually make your own shortcuts for example, that could be applied in there where it's like, you know, a quick button to humanize the writing, let's say make it look a little bit less AI generated. Or it could be something like goofify, like whatever you want it to do. Custom edits or custom quick edits that can be added to a little bar on the side. I think that would be really cool. Maybe generate the logos for those with Dall E3. I don't just just an idea. Open AI anyways, let me know what you guys think about this. I'd love to have a conversation down in the comments below. It's pretty dumb. Darn cool. It's cool for sure. I think it's probably in some ways better than the CLAUDE interface. I know we didn't get too deep into that in today's video, but the CLAUDE interface can actually run some of the code that CLAUDE generates and that's maybe a next step on the coding front that I'd like to see from OpenAI. They've done something similar with Codex in the past to that, so maybe it's not a huge leap that they would do something like that. Huge thanks to everyone for watching. Check out my Discord server. Check out my other videos. Should have hopefully a news roundup coming soon. Sometime pretty soon. Thank you so much for watching and I'll see you in the next video. Goodbye.