# GPT-5 Full Review & 10 Mind-Blowing Use Cases

## Метаданные

- **Канал:** Varun Mayya
- **YouTube:** https://www.youtube.com/watch?v=U5wStfgp4do
- **Дата:** 07.08.2025
- **Длительность:** 38:39
- **Просмотры:** 481,823
- **Источник:** https://ekstraktznaniy.ru/video/11960

## Описание

Learn more about Code Rabbit here: https://coderabbit.link/varun

OpenAI has just launched its most awaited model yet: GPT-5. And it’s not just one step closer to AGI, but has almost entirely automated a lot of things using just simple prompts.

In this video, we put the model to the test and build all sorts of things. From building Twitter clones, drag-and-drop CRMs, to one-shotting 2D Spider-Man game, Marvel-themed Tetris!

And then we go beyond building by testing its writing style mimicry, prompting it with deep philosophical questions, and even running deep research to see how well it combines its advanced reasoning with pulling niche internet sources.

Spoiler alert: it outperforms even human researchers in some cases… and it’s blazing fast!

Watch it till the end for some amazing use-cases.

🚨Note: If our content inspires you and you’d like to be part of the team behind it, we’re hiring a Senior Motion Graphics Designer:

Apply now:  https://forms.gle/DhKxYneiurfGbV4P8

00:00 - 

## Транскрипт

### Intro: Lets Build with GPT-5 []

Ladies and gentlemen, OpenAI has launched GPD5. Now, we've had early access to GPD5 for a while now, and we've tried a lot of different prompts, and let me tell you something about the model before I show you what it can do. I think it's definitely one of the smartest models out there, and it can oneshot a lot of things. It's able to oneshot entire front ends. It's able to build lots of apps on the fly. It can make games, like simple 2D games on the browser. It can make them pretty well. It's even able to think about things like physics in 2D games and stuff like that. And it is blazing fast. But I'm going to stop talking here and straight up show you what you can actually do with GPD5. So, we're going to show you a bunch of use cases live and you can decide for yourself what the model can do and what it can't. And remember, 3 years ago, less than 3 years ago, when GPD 3. 5 first came out, it really did a pathetic job of coding. We'd built an agenting tool called Autocore Pro back then, and since then it has evolved quite a bit. So let's get straight into

### Twitter App in One Shot [0:56]

the action. There are two versions of the model. One is GBD5 and the other one is called GBD5 thinking. We're going to use GBD5 thinking which thinks a lot more before it outputs tokens. So, I'm going to go straight up and say make a Twitter app uh with modern UI black and white um with all buttons functional in the front end and use pretty icons for the default profile pictures. That's it. And it's going to do its job. Like I told you, GB5 is blazing fast. So you can see what's happening on the screen uh in real time. We're going to fast forward a little bit after the first one. But for the first one, I want you to see, do you remember how slowly it used to output tokens? And to be very honest, this doesn't even look like WEB coding. This is just Waiting. But to be fair, it is building a full Twitter app. So let's see what it comes up with. As you can see, earlier models really struggled with long context outputs. It is now really, it's already close to about 900 lines of code and still going. Awesome. As you can see, now there's a preview button that's appeared on the top right. I'm going to click the preview button. You can preview your output straight in the browser in chat GPT. And wow, this looks very much like Twitter. So, I can actually say, "Hi, what's up? " You can see those clever icons as well from the profile pictures and you can see it's here. Now this is all happening in the front end. But let's click on some of the other icons. You can see that explore now shows you an explore feed. There's notifications. There's messages. There's DMs which is looks like an exact copy of Twitter DMs, right? There's bookmarks, there's profile. This is really cool. And I can actually click on following and follow and whatnot. I can reply to this single prompt. Now I want to tell you one thing right which is that a lot of this is happening on the front end. Now the question is there a backend for this? If you want a backend for this I can actually ask for one but you can't actually play it in canvas then you'll have to download it and spawn your own server. Also make a backend for this. Okay. Now it'll actually go out and make a backend. You won't be able to preview it in the canvas. You'll actually have to download it along with all the node packages. Uh but it makes it pretty easy for you. And we have tested this offline. It works flawlessly. So, it's actually now creating the back end for you in Node and Express and it'll also come with full instructions on how to deploy it. Now, with other VIP coding apps, they actually go out and allow you to connect to things like Superbase and actually set up your backend and admin console and things like that. This model GBD5 is now going to power some of them and they are going to become superpowered because they allow you to do end-to-end deployment. Now, it makes sense at some point that chat GPD also allows end-to-end deployment because it simply feels like the last step and it makes sense that they do it. So, as you can see here, it's not only created the back end for you, it's also telling you how to run it locally, uh, including all the commands you need to run and it's giving you the entire endpoint list that it has and then all it has to do is wire the front end. So, yeah, you have a full-blown Twitter app that we made in what, like 2 minutes, 3 minutes. Anyway

### CRM For Won & Lost Deals [4:17]

let's keep going. Let's try to make something harder. Now, I know some of you are really interested in making a CRM. And a lot of you were not very impressed with that front end because it looked exactly like a Twitter clone. So, let's go crazy with the prompting thing, right? Uh, actually, let's keep it down to two lines. Let's see what we can do with just two or three lines. Make a CRM for one and lost deals. Okay, KBAN style. Okay, drag and drop. uh drag and drop and create all sorts of shiny bells and whistles and shiny moving CSS and card effects and uh and yeah, when a deal is won, make confetti appear from the top. Yeah, I'm going very crazy. Let's see if it does this. So, we're going to fast forward this because it's it's writing code and I don't like vibrating. Uh, plus it reduces retention of the video. So, let's quickly go through what it does. Let it do its thing. I went to 4 years of college for this, by the way. So, pretty insane how much work I'm doing. All right, let's hit preview. Awesome. So, this was created. Let's create a deal. Let's call it uh VM deal. Amount $5,000. Let's say it's in progress. Add deal. And as you can see, the deal is there. It's added a deal. Now, can I draw drag and drop it? Absolutely. Now, if you put it in one confetti, amazing. I put in lost. There's a nice little shake. You can see the little sad shake. Yeah, vibe coded one shot. I've been using it so much over the last few days that it's now normalized to me. But I think it's pretty phenomenal that it works out of the box with two or three lines of prompt. Let's move on to the

### Building 2D Games [6:23]

next task. Now, let's see if it can make quick 2D games. Okay. So, make a Spider-Man web swinging sidecroller 2D game where you can attach a web to buildings and swing and even go over the buildings with enough momentum. Anyway, by coding, sorry, by waiting. Awesome. Let's hit preview. React canvas. Whoops. And as you can see, this is Spider-Man. Wow, that's so cool. Okay, I need to read the instructions. So, AD is move. Okay, I got that. Left click targets the edge of a building. That's pretty cool. Got it. left. Uh, Q, detach web. Okay. No, no. That's not good. Shift is to reel myself on top. Okay, that's cool. Aha. This is pretty good. You can see I can spin like this and then with enough momentum I can get on top. Wow, this feels really cool to play. I screwed up the controls. Wow, now I got the controls and I can reel myself in and then with enough momentum I can get on top. That is awesome. As you can see, there's nice little stars in the back. So, this doesn't use any sprites. So, let's see if we can allow it to use our own custom sprite. Can I Can you allow me to put up a custom sprite, please? See, always be nice to the AI, but as you can see, it's still hanging and swinging there, right? Like it's thought through the physics. So anyway, it's uh now added the ability to upload my own PNG. So, let me just click on preview again. So, it's allowed me to upload a sprite. So, let's just, you know, look for Spider-Man sprite. Let's take this one. We just save image. Uh, and I'm going to upload this Spider-Man sprite. Uh, wow. You know what? I need to remove the background of the sprite. Let me see if I can remove it. Can you automatically remove the background of the sprite? I have a sprite with a background. I'm being super lazy here. I bet you could also add sprites with animations. We've tried this. It actually works pretty well. 2D sides scrollers. This will generate most of the actual logic for you to put into the game. And it's just about telling the story. Okay. So, it's allowed me to do that. Let me just refresh. Let me go back. Let me preview. Okay. There's an unnecessary escape sequence. I fixed the bug. There's a bug. So, whenever there's a bug, it tells you what the bug is and then it tells you that you can fix the bug. It's just fixing the bug for me on the fly. See, it's going through the entire code to fix the bug. This is how bugs look. Then it can heal itself. There used to be this old repository for AI coding, right? Uh called Wolverine, which did automatic healing. Uh so, it's kind of doing that. The waiting is the most annoying part even though it's a very fast model. My only complaint with GPD5 is that I don't like waiting. I want instant outputs. Okay, it's done. I'm going to preview. So, as you can see here, it's also figured out a lot of the mat. It's figured out the colors. rectangles. It's filled the rectangles. So in fact we've been using this for a use case right in our game we have the idea for a sudashan chakra that you can throw around the arc for it and the return back you actually need to be very good with physics because we modeled it around a frisbee and frisbes also move along their own axis while moving forward and GBI was able to calculate pretty much how it should move one shot. Anyway, so we've got this. Let me choose a file first. Awesome. So, as you can see, that Spider-Man is being covered by the thing on the right, which I should have asked it not to put it on the top right and maybe put it somewhere else. But, as you can see, it works and it works pretty well. Can you imagine that? It erased the background by itself. It allowed me to upload a sprite. It did the feathering on it. I know that It said, "Wow. " I remember playing these old school 2D Spider-Man games many years ago and maybe almost more than a decade ago. Um, and to see them being able to be one shot with one or two fixes of bugs is pretty fascinating. And by the way, one

### CodeRabbit [11:42]

of the questions that many people have, especially with code, is where are we going to actually write this code? Is it going to be us just prompting to chat GP and that's the end of it? I think it's actually going to be inside of GitHub and your idees. I think a lot of these VIP coding tools and a lot of apps that actually sit on top of GitHub and help you work better are actually really going to benefit from the new models. I'll give you an example, right? Today devs spend as little as 9% of their time actually writing code. A large chunk of their time is spending on things like meetings, code reviews, like what we did, right? Spending time reviewing code and fixing bugs. Turns out the problem lies in how most traditional developers handle these tasks, especially code reviews. They're stuck with the very old ways of reviewing code and fixing bugs manually even in 2025. And we just showed you how fixed bugs are so much easier. So on GitHub for example, there's a solution. The solution is called Code Rabbit. So like we said, the few developers, the top 1%, the top 5%, they are sitting on GitHub and already AI assisted. So Code Rabbit is an AI powered coding reviewer that connects to your GitHub and it uses deep learning to scan and understand your entire codebase, reviews your code, and catches bugs almost instantly. Plus, it also suggests contextually relevant solutions on each pull request that you make. Now, it's already reviewed more than 10 million pull requests, installed on over 1 million repositories, so you can see the demand and used by 70,000 opensource projects. And the best part is it works directly within VS Code, Cursor, and Windsurf. And it's absolutely free for all open source projects. But it fits in our thesis of we think that all these powerful tools are still going to be used where you already use them. So if you're already using GitHub to collaborate with other people, now you're also going to collaborate with AI agents to help you. The models are just the underlying intelligence because it is cumbersome to go change this code here. And while GPD5 is actually very good making things from scratch, on larger code bases, it struggles a little bit. And it's important to look at that code in terms of chunks and solve each chunk while still keeping the bigger picture in mind. Anyway, let's get back

### Custom Tetris Game [13:36]

to the video. Let's ask it for another different type of game. Let's see if it can do games like Tetris. Can you make a Tetris game with characters from the Marvel universe? That's a weird prompt, right? And I don't know if we're going to get a decent output for this, but let's see. Let's find out. And it's doing it. I don't know what it's going to come up with because we haven't tried this prompt before, but let's see. In the Spider-Man example, I could have asked it to move the UI to the bottom left. I could have just told it that move the UI to the bottom left. Um, and we would have done it, which is awesome. This is why I keep saying, right, I've said this for two or three years that English is the new programming language. Everyone's saying it. Everyone from Andre Karpati to I mean, they've been saying it for a while, right? That English is the programming language. Knowing what you want, knowing that you have to raise the background there for a Spider-Man uh asset, knowing what a sprite is. These are all simple things, but you know, you'll be surprised how many people who've gone to engineering college don't know about any of these things. So, anyway, I'm going to preview the Marvel Tetris game. Okay, this is interesting. So, left and right to move. Oh, wow. So, this is Thor. Okay, so let's use small emojis. Let's do that. Awesome. This is Who is this? This is Doctor Strange. Rotate. And I'll do this. I just can't believe this works. This is so good. This is Cap Iron Man. Uh, this is Hulk. I got it from the icon. Let's see if it deletes that line. And it did. It deleted the line. One shot. Anyway, let's go to

### Building Websites [15:27]

something a little more useful. Let's ask it to make a website. Make a fancy website for Batman. uh with all sorts of shiny CSS and bells and whistles and make it look really cool. Let's see what it comes up with. We've tried the website, you know, prompt many times. Sometimes it doesn't load images, so you have to explicitly tell it to load images. It may not do images on this one, but it does a fantastic job. I put out a tweet a few days ago before this video will go live saying that you know we we're now entering the midjourney era for front-end devs and while most people agreed because they're now seeing the progress there was still a couple of people who were like ah you're so stupid you know nice joke and stuff like that I feel like intelligence is not about skill maxing right you're good at a skill you keep getting better that's actually practice you get better at skills the more you practice it right the real intelligence comes from adaptability when what you are doing for a long time goes to zero or it's no longer working or the value of it goes to zero because the supply demand curve for software engineers is changing now right with this even if software engineers are hired which I still believe they will cuz the world needs more software but even if they are hired they will still be fewer in number compared to the mass factory production of software engineers that India has already been performing right because it's the high status thing and we're already seeing it reflect in salaries even before AI it's already being reflected in many of these salaries so I feel that when that stops working your job is to be adaptable and find the next thing that works, right? Find new technologies that work because that's always the cycle with technology. When I started writing code, when I was in college, I raised money for a company from investors I didn't know at all for a company called Jobspire where we said we will build an online platform for recruitment. At that time, we used to use something called Ruby on Rails. And we looked down at all the people who were using JavaScript and JavaScript. I mean, JavaScript frameworks were a very rare thing back then. So, we used to look down upon them and saying that Ruby on Rails is superior. But a few years later when you know after that company got done and after the meta changed of what people were using we moved on and we moved to newer frameworks because we said this is the more efficient way to do things but I know a lot of developers back then who got caught into the old meta the old way of doing things and never recovered they're still doing it the old way now 10 12 years later so adaptation is important and you were probably seeing the first cycle of adaptation that's happening live and there's two kinds of people one kind of person who's like this is fake this is not real and attack everybody like me in my comments saying you don't know anything because we made a video about this maybe 6 months or a year ago about how software engineering is going through a big fundamental shift and we did that after we had made an app called Autocode Pro and seen the early signs and we had all these dots of how the models are improving and I have early access to all of these tools and I've said well maybe the number of software engineers will reduce the world still needs more software but do we need as many software engineers doing baseline you know just using the framework and making the same you know starter app again and again. No, we need a new paradigm. So I would say everyone who has spent the last two years coping has not done very well. And I have come to the conclusion that coping is kind of the opposite of intelligence. The more you show signs of coping, the less intelligent you are because intelligence comes from adaptation. If you've read any Charles Darin and you know how evolution works, it's survival of the fittest, survival of the most adaptable, not survival of the best lead code grinder or the best interview prep uh person, right? Like it's just we are in a different era now and it's time you opened your eyes and adapted because there you're still early. There's still so many people coping. That is your I mean we have gone from 0 to 400 employees as AOS completely bootstrapped in 2 and 1/2 years because we adopted new technology in AI while everyone said nobody's going to watch AI presenters in content or nobody cares about AI as a field so much or nobody cares about making distribution because distribution is not that important if you build a good product they will come but we had learned from my last company where we had spent time and energy building out a good product that the era of good products is ending because distribution matters far more. If you have distribution, no matter what somebody else has, you will gain all the products. So that's why we adapted to distribution and we said that at some point software will be one click because we've run the early experiments. We just have to wait 2 or 3 years and running a business is a 10, 20, 30 year affair. So we don't mind waiting and hoping that our distribution is now valuable. Now that you can oneshot code, we are now able to put out so many different products and use our distribution to get them in front of you. We use our access with all of these big companies that we've built because we have the distribution. That also gives us advantages and I promise you what matters more is supply demand economics. If there is an overupp of something that can easily be done, its value regardless of whatever you say goes down. And I wish everyone took an economics class with as much intent as they took the book Cracking the Coding Interview. Anyway, let's get back. Let's see what it's done. And as you can see, it's not used uh third party images because it tends to do that with websites. But as you can see, they've added a lot of bells and whistles. So if you press toggle bat signal, you can see that it's toggling some sort of light stand down. That is very weird. Uh explore loadout and you can see the loadouts that it has. You know what? Use external images please. So sometimes you have to allow external images to be able to access um you know the GPT canvas or it you have to allow GPT canvas to access external images. So we're just telling it to do that. Remember that GBT and the reason this website doesn't look so good is because it tends to go back or fall back upon CSS simple colors and shapes because it doesn't know because it's it's a little bit hesitant to use external images. What if we don't have the copyrights for the external images? There's too many issues when you start using external images. But uh we have you know we've seen that if you can explicitly tell it to use external images it tends to do a much better job and you start realizing how much of website is actually about images. Now, luckily with images, you can AI generate images. So, yeah, it does make our life easier. So, let me just refresh this. Let me go back to canvas. Let me preview. There's an syntax error. Let me fix the bug. Sometimes it does make errors, but it gives you the option to fix it. And you might think, well, why should I fix it? If there's an error, find out yourself and then fix it. It of course will be able to do that, but I feel like it gives you a little bit of access if you are an actual engineer to go fix something uh right on the fly, so it doesn't have to go through the entire codebase again. But that's okay. In our case here, it's fine. And if you notice, right, something I've noticed a lot when you ask it to fix things, it actually goes and generates an entirely new code base. So, some of the options that you had in the past would be change. For example, here in the gallery, it's added clock tower, which probably wasn't there the last time. Uh, and it does that very often. So, it's almost remaking the entire website with that small change that you wanted to. It's I mean, it's unfortunately one of the drawbacks of using an LLM. So, as long as you get the right output, it's there. So, let's just preview this. And now I'm going to allow selected. As you can see, it's asking to connect to the network. And now there are some images. So, as you can see, it's used just very generic Unsplash images. It's not actually used pictures of Batman because they are very careful with third party external images. They tend to default to Unsplash and Unsplash has no Batman images. Now, let's try to make a website for something that we can easily find images for, right? Make a nice pretty modern website for a coding company. Use external images. Coding company called Tete Coding Services. Just kidding, guys. Um, let's see what it does. Now, I've been purposely ambiguous with this, right? What do I mean by coding company? Is it coding training? Is it actually doing code for end customers? Is it building software? Let's see where it defaults. And I can see from the codebase itself in the title that it's saying that we built thin fast software. So it's going after actually doing code writing code for end customers. You can also actually give it specific links to websites and say use images from here but it does it very sparingly. It's a little bit careful about it. All right. So TED coding services ready. Let's click preview. Allowing third party images. So as you can see it's still okay. Yeah it's brought images in. This looks pretty good. Looks pretty clean. It's followed my uh you know my prompts. And as you can see, even the text, it's thought of very good text for the entire thing. Average project kickoff in 5 to 7 days. That's actually pretty cool, right? Typical engagement is 6 to 12 weeks. It gives the client so much more information, right? And as you can see, there's a nice little, you know, shine that's going past this image. It's showing you some of their selected work using this which stack it uses uh the company uses or the stacks that they're familiar with. Uh what clients say it's got nice testimonials. Start a project does a nice little form. Yeah, this is pretty clean, straightforward. Um you can also give it references. You can say make a website like this and it'll actually go do a decent job of making a website like it but which still looks different. Now let's go to a slightly

### Creative Writing [24:32]

different angle. Let's see how GPD5 does with writing. And you know the problem I've always had with GPD claude and many others is man everything that it writes is pretty generic, right? It's like you can tell when something is generic. Can it create signal? Can it create something new? And can it write in my style? So let's try this. So I'm actually going to take a few scripts that I've written in the past including a script on how to write which I'll at some point make a video on. And I'm going to paste them in. Okay. I'm just going to paste all of these in. I could have downloaded and uploaded but whatever I'm lazy. Uh paste these in. Uh okay. See this is Vun Maya's writing style. You can also use some of his writing from the internet. Write a short essay, a short insightpacked essay on why intelligence is adaptability and not skill maxing. You can go through his Twitter for writing style if you need. Think and write like him. Cool. Oh, it's actually refused direct stylistic imitation. Okay, so it said I don't want to write exactly like Vun, but let's see what it's written. Okay, and I told you it always thinks about insight. So I'm going to read it out for you. Intelligence equal to adaptability, not skill maxing. Skills have a halfife. Intelligence is your reload speed. Skill maxing assumes the game stays still, but the rules keep patching. Markets shift, tools upgrade, incentives flip, teams reorganize, models get smarter. The person who wins isn't the one with the fattest skill checklist. It's the one who reconfigures faster when reality changes. The frame shift skill is optimizing for a snapshot. Become the best at X. Adaptability optimizes for a stream. Detect the change like water, right? Detect the change, update the map, swap the tools, keep shipping. In other words, masters focus on reception, not transmission. What's actually happening out there? what does this new context demand from me now and it talks about the three loops of adaptable intelligence. Now it may not be 100% right but it is pretty creative in how it is thinking about problems and I feel that's what makes writing with GBD5 very different from writing with GBD4 and you know I've been spending quite some time with GBD5 actually more than writing I'm learning new things right because I ask you to look at a problem from different ways and what GBD5 is great at doing to describe it is it's great at analogies it's great at taking a topic from somewhere else and putting it somewhere else and it's great at saying here's a correlation between some completely unrelated industry But here's a pattern that happens here that also happens in your industry. It gives you new ways of thinking. Like I feel like this is way smarter than a human at this point in terms of writing and even thinking about new ideas and insights. One experiment

### Game Dev Ideation [27:29]

and a lot of people you know keep saying this right like AI will never be able to come up with new game mechanics etc etc. I have an idea and let's see if GPD is able to do it. We have a game where the main character can use a sudaran chakra. Can you give me a skill tree including parts for what he can do with it? The player can also use teleports. Okay. Yeah. Model the sudaran chakra around a frisbee. So include catch and throw mini games inside the main game loop. Okay, let's see what it does. Now we've obviously already designed this pre AAI but I just wanted to give the problem to GBD5 on the fly and see what it does. Okay, so it said core loop and resources are throw then you can guide and manipulate as time slows down then you can catch and recall and you can change this. So it says that you have a resource called focus and it's generated by perfect releases and clean catches. So you probably have a timer for when it goes out and then you can cleanly catch it when it comes back. you have a small time window for that and then you gain that and that focus can be spent on advanced throws, teleports and you ultimates. Okay, next is called heat builds on sustained spins and multi throws. So overheat shrinks catch windows until you cool down. So as you're catching the window slows slowly like closes uh because you're the chakra generating too much heat and that makes it that the sustaining a combo over the long run is harder. Okay, then uh let's talk about skill trees, right? It says throw mastery, which is power angles and ricochet. So, you can spin up, you can hold to charge, spin, more stability and damage, raises heat slowly, quick toss, buffered instant throw, uh throw out of jaw, dodge and jump. That's a cool idea. Next two is T2, which is path A, which is curves. One is called Heiser discipline, which is bank left and right curves. So, you can start using it for curving. One ricochet on perfect release. The other one is called Anheiser discipline, which is inverse curve. Can arc behind targets. Now, I don't know what the value of that is, but it's come out there. Uh, then there's skip shot, which is ground bounce. You can add it on the you can throw the chakra at the ground and it'll, you know, it has a low angle hitbox and then it bounces upwards. It's almost like a, you know, a cricket ball throw. Uh, then there is um, ricochet one, which is plus two bounces on perfect release. Then there's the serrated edge, which is stacking bleed on consecutive hits in one fight. Now, probably I'm going very long here, right? But it has elements for throw mastery. chakra control which is hover break and things like that which is brief. Hover break is brief stall in air to retarget and slows time for 0. 3 seconds. These are really clever ideas and like I said right like a lot of your thinking work now will be spent talking to an AI to try and figure out is something possible or not. Now not all of this is possible but whatever you end up creating is actually going to be reasonably unique because GPT has taken the entire world's information. Everything that anyone's ever written, most of it is available to GPD in the training data. So, it's creating new correlations. Like I told you, GPD 5 is very good at taking things from totally different industries. Like something that would happen in the frisbee industry. In ultimate frisbee, it's able to transfer here and say, "This is exactly how it works in ultimate frisbee. You should implement something like it similar. " And we've had a lot of fun just seeing all the crazy ideas it puts out at us. Awesome. Now, it's not

### GPT 5 Agent [30:57]

done yet. GPD5 can also be used in agent mode. So I'm going to move to agent mode and I'm actually going to tell GPD5 to do any task. Now I'm actually not going to show you the process because it takes 17 20 minutes for the agent to work. But I made I asked it for a task which is edit a video of today's top news in AVTV or walks style. Okay. Now I've actually already run this prompt. Uh the output wasn't very great. The output was more like it went to pixels because it doesn't have that footage. Right? Even if it generates that footage with Sora or even if it had access to VO3, it's still B-roll that's not shot in the real world. So, it looks like a string of B-roll stuck together. It puts a small little, you know, text box at the bottom and it says, "Here's today's news. " I'll actually show you an output. So, you can actually see that, right? Like, as much as we talk about GBD's availability of all the data in the world, it can't use video footage that was shot in the real world, right? So it's not actually able to take today's news something from ANI or many of these other places that actually produce content in real time. It can't actually use that right and GPT really gets stuck at copyrights even though they're trained on the world's data they get stuck on copyrights for things that were produced recently or by other providers because they don't want to get into lawsuits. And the second piece so it's using something called Movie Pie to put it all together. So that's why the transitions, the movements, the text, they all look pretty sad. Somehow we need to give GPT the ability to use Premier Pro or After Effects. And I feel tool use is the next task for agents. They need to be trained on those tool use because earlier I have tried to ask GP to make a thumbnail and it did a really bad job. Even though we were using photop which is a browser tool. It did a really bad job. But now I think more people will start generating data for tool use. Uh but right now it really struggles at tool use. Even GPD 5 struggles at third party tool use. So it does everything through intelligence and writing of code, but it's not very good at actually going out, clicking a mouse and doing stuff. So yeah, I have at least another 50 use cases to talk about. And I have one last thing to show you. We've seen GBD5, right? We've seen GBD5 build us some front end. We've seen GPD5 build us some apps. We've seen GBD5 build us some 2D games. I've also used it to make a 3D game, but it's actually cheating. It's making 2D look like 3D. Um, so I'm not going to actually show you that. And that's because it doesn't have access to the models or an engine. With the 2D game, you can see it building the engine live. We've also shown you how it does at video tasks.

### Deep Research [33:19]

Now, I'm going to show you one last example of GPD5 that's actually significantly better from GPT4 or any of the O models. And I'm going to go to tools and I'm actually going to use deep research. GBD5 does a fantastic job of deep research of finding information that is very hard to find on the internet. So I'm going to ask it for deep research on please go do a deep research on the top 20 philosophical insights that the best books all agree on. Right? So it's a weird question. Please go do a deep research on the top 20 philosophical insights that the best books agree on. Actually let's not say best. Let's say that every single book generally agrees on and by the way what we found with deep research is very good at finding out numbers. So if you want to find out the cost of making a restaurant let me just say everything or the cost of starting a restaurant in Bangalore you know owning a private jet in India. It does a very good job. It finds these niche sources from like forums and stuff like that. puts all that data together, correlates it, checks, you know, for hallucinations. And by the way, OpenAI has its own sort of verifier now that's checking for hallucinations. At least that's what I read. So, it's actually pretty good. And I would say at this point, I take deep research a lot more seriously than I take. Even some people in my own office that I say, "Hey, go do research on this topic. " Cuz I know they wouldn't have gone to the niche forums. scrolled the entire web and got the entire web's information. They're very good at checking for hallucinations. Real humans are And the minute you need to pick up a phone and ask somebody, hey bro, what is the cost of, you know, rent in this area, the minute you have to do that, AI is useless. That's why anything to do with the offline world, anything where you need a human to go out and search for information or book something or talk to somebody, AI right now is not very good. And maybe at some point we'll have those AI callers that you can automatically use, but right now is not that time. And by the way, while this actually goes and does the deep research, I want to tell you a little bit, you know, like my thesis, my forming thesis about how I think software will work in the future. I think there's going to be a lot of personal software, but that personal software is going to sit embedded inside GPT. So, if I want a personal CRM, I can just make it and all that GPT or whoever needs to make is a back end and, you know, decent authentication uh for me to be able to use it without security concerns and for me to know it's private. I would love to use it straight in my GP. Gemini had actually shown an early version of this a long time ago of software being built on the fly for you personally. Yeah, it's going to happen for a lot of simple software. But for kind of software like Twitter, right, like where the hard part is not really the software. getting the users. Actually, it's a combination, right? You need software and you need the users and both them together make the product. Software part could at this point is you know very close to being automated. Um getting those users and having those network effects where the value of the network grows the more user there are. That part is still a very human part. knowing what campaign to run and you might say well yeah I can think of those campaigns but the meta changes every few weeks as people get bored of something now something else needs to come up right so it is something that you'll have to look and see and then decide should I make this video should I make that video and it depends on the cultural context of where you are at and what you know other people believe and so on and so forth all right so we got the outputs it says the 20 universal philosophical insights like I told you it's pretty insightful it's able to pick out uh you know interesting things and correlations between each other so the first one is the ethic of reciprocity which is the golden rule which is you treat others as you want to be treated. If you put negativity in the world, you will get negativity back. Uh so it's gone from Confucious's teaching all the way to Jesus you know newer philosophers. The next one is impermanence and constant change. Something this video is about which is change is the only constant. If you look at something from 10 years ago and said this is how it was done 10 years ago, that's why it's always going to be like this. I see this with a lot of like people who make gaming videos on YouTube, right? Like oh India can't do a game because 10 years ago it was like this and you know where are you going to do this? We're in the world of AI like where everything evolves pretty quickly. Um next is self-nowledge and know thyself. Then there's unity. I mean this is all like this is a lot more like it's connected things from Indian vanta which is the unity of atman and brahman uh and connected that to in Buddhism Mahayana and in Taoism the oneness with the tao. So it's done very good cross domain cross uh scripture sort of integration which I really like. Um but yeah like you should try out deep research especially if you want numbers. It's very good at it. So there's a lot more

### Closing Thoughts [37:51]

we can show you with GB5. I've tried at this point like we five 600 prompts on GB5. Seen a lot of different outputs. Know what it's good at, know what it's bad at. So I'll probably make a follow-up video to this one. Make sure you subscribe. These are all early opportunities and early arbitragees. If you're the kind of person like no new models you've used GPD 3. 5 2 3 years ago and based on that you made your decision that AI is bad and you are not able to see the progress that's happened in 2 years feel free to leave a negative comment but if you're somebody who wants to learn uh who wants to see what new is possible and then use that in your business or creating products or services or be better at your job make sure you subscribe the technology is here whether you like it or not it's better to use it and have an advantage over everybody else not using it this is a channel for you to do that that's it for Bye.