Everything You Need To Know About GPT-4

Okay, welcome to today's live stream, guys. All right, let's do this. So quick, warm up one minute and we'll get right into it. Okay? Welcome, welcome, welcome. So this is. It's happening, it's happening. What else can I say? I mean, look at this. We have access. So quick overview. Look, while the stream warms up, okay, we're gonna talk all the improvements, and not just that, we're also going to test them. I have Chat GPT plus right here. And we're going to compare some outputs between chat GPT 3.5 and GPT 4. Okay, so we'll just wait for a few more people to stream in here. We're live, and then we're going to take your suggestions and implement those. Good to have you here, guys. All right, so much, so much happened. I. I went for everything and prepared for this. Now I summarize the live stream for YouTube. So this is. If you didn't see the live stream, if you didn't read the blog posts, if you didn't see their emails, this is gonna be your all in one summary right here. Okay? Excited to be doing this. A little hype, not gonna lie, A little hype. This is a big moment. We've been waiting for this since months. All right, let's get into this. So, okay, yeah, I know music is a little bit loud. That's fine. That's fine. In the intro, you know. Okay, let's get into this. Yeah. We good now? We're good now. Ladies and gentlemen, it happened. GPT4 wasn't just announced, but also just launched right here. If I switch over to my screen, you'll be able to see that. With ChatGPT plus, we are able to switch to GPT4. Now the question presents itself. What does this mean? Okay. What does this mean for us, the users? What use cases are going to be better? What new features does it have? What new features does it have that are not available yet? What about the development side of things? We're going to talk about all this. So if you didn't watch this live stream that just happened, if you didn't watch the trailer that just was released, they sent out an email today. If you didn't read the blog post, this video is going to catch you up on everything, plus give you some concrete use cases. So, look, let's get right into it. No time to waste. This is a live stream, so feel free to comment and bring your own questions and your suggestions. But I'm going to blaze through this, okay? I want this to Be valuable even to the people that are tuning in that are not here live with us. This is so exciting. Okay, so first question. Can you get four on the free version? No. Very, very simple. You need to have ChatGPT+ to access this right now. That has been answered and with that being said, let's dive right into it. So all the links are in the description below. You can check out the blog post, you can check out the trailer, kind of a trailer. You can check out the live stream. It's all valuable stuff. But let's get into it. Let's talk about what changed. And the answer is a lot. But not as much as you might think. Not as much as you might think. Okay, so let's talk in practical terms. Let's talk in practical terms. Okay, I have some notes here and let's go through this. So what did I add? Well, first of all, one of the biggest and most obvious ones is a so called vision module. This is not accessible yet, so we won't be able to test that today, as you can see right here. But there's a wonderful preview of this in the blog post. So let's have a look at this together. There is this part where you can actually send that image to ChatGPT now, or you'll be able to with the developer tools for now and then they'll be rolling it out slowly. So this is not accessible for the browser yet again, but you'll be able to send it an image and ask tricky questions. Okay, so the key here is it's not just a language model anymore. They're adding a vision model to it too. It's going to be both in one, so you're going to have to reasoning in the wisdom of the language model. That is even better now. And we'll talk about that in a second. Plus you have this vision model. So for example, here he posts this picture, right? And it's obviously it's kind of a funny meme like picture from Reddit. And you can ask it what is funny about this image? Describe it panel by panel. And then it goes ahead a smartphone with a VGA connector plugged into its charging port panel to the package it describes it. And then the humor in this image comes from the absurdity of plugging a large outdated VGA connector into a small modern smartphone charging port. So you take all the insane capabilities from ChatGPT and now you can apply that to images too. So that is the first like massive update. Like, seriously guys, anybody that has them spam these emotes Spam these. This like Peppo, that is. His eyes are popping out. If you have it, all the members, let's, let's do this. This is incredible. Okay, you're gonna be able to interface with your pictures, with your photo libraries, obviously, eventually, right? You're gonna be able to load that in and be like, hey, so tell me about, you know, this person. And then it analyzes all the pictures, takes all the data, all the knowledge, all the reasoning contained in those pictures and you can interface, you will be able to interface with it. This is honestly, honestly, this is impressive. Okay, so that's number one, right? We have the vision module coming in. Very impressive. We can't use it yet. What can we use though now? It's better at reasoning. What does that mean, Igor? It's better at reasoning. Well, it's gonna be better at solving your problems. Now what is a concrete example of that? They show a few in here. And as I said, we'll get to the part where we'll be exploring this together and we'll be testing ChatGPT versus GPT4 in several use cases. And I, I have to admit, I ran one already. Okay, I ran one, the good old Penguin prompt. Write me an essay about 2000 word essay about penguins. And although it's better at reasoning, it hasn't managed. I said 2000 words. Look at that, 1164 words. It's not. So, it's not perfect. But the way to think about this, the way I understand this right now, is it's an incremental upgrade. The vision model is a massive upgrade, alright? But everything else is kind of better, you know, 50% better, 80% better. It's not 100 times better. It's not what those graphs used to suggest. You remember those, that little, you know, the little. It used to be like, ooh, this is like ChatGPT and this is like GPT4. This is the data it has been trained on. So it's not that type of difference in terms of results. But what it is, what is this different in, is that it's safer. Okay. And this is a. They have their way of expressing this, but they essentially say it's way safer. All the guardrails are reliable now. They learned, they learned from the mistakes they made in the past. And you can, yeah, rely upon it a little more. Okay, so this accuracy and adversarial questions. This got better by 82%. So the guardrails are up. They're there. I'm guessing the damn prompt will be kind of hard to run. Here. Okay. But either way. So I think that's an interesting fact. It's safer. And here's a big one. Here's a big one. Do we have some good sound effect here to like? No, I'm not sure. Maybe. Maybe we don't. Here's a big one. Let's do it with the robot voice because that just feels appropriate. It just feels appropriate. Okay. You can process long documents. If you didn't get that, you can process long documents. Okay. They increased the cap on the tokens that you're able to feed it as a developer now, meaning all the capabilities where it could read your PDFs and the apps that people build, they were very limited. You couldn't feed it like a knowledge base of, you know, 100 pages. Now there's limits on it too. There's limits on it, mind you. And they're all outlined here in detail, if you care about all the details. As I said, this is going to be about the use cases, but it way expanded, it's way larger and you're going to be able to feed it a lot more of your data. All right, So I think that's fascinating, but this is, again, this is something for the developers out there. What can us normal people that just use this day to day inside the browser get from it? Well, I think one of the most interesting ones is they claim that it's more creative. That's another one. Okay, so it's going to be more creative. What does that mean, more creative? Well, what they say. And we're going to test this soon again. Yes, the info is still until 2021. Yes, yes, yes. But it's more creative. So what does that mean? They say right here again, we're going to test this very soon. But they say it's way better at writing songs, screenplays and learning your writing style. Okay, that's a big one that I'm trying to get here. Get to here. Creativity. Look at this. So this is a fantastic example and I want to demonstrate it on this. And we're going to try this ourselves. Right, but effectively it's. If you go ahead and say explain the plot of Cinderella in a sentence where each word has to begin with the next letter in the Alphabet from A to Z without repeating any letters. That's quite challenging. And right away we can, you know what, we're live. We're just going to do this, we're going to run this through GPT 3.5 and we're going to see what we get. And this beautifully shows the Problem. A, B, C D E, F, G H, I why? Okay, that's, that's actually pretty good. That's actually surprisingly good, I would say. Let's see what GPT4 comes up with. And this is the first one, this is the first prompt we send on the channel. So let's have a look. And look at that. That actually did not work well. See the perils of being live. Let's try again. Wait. A, B, C, D, E, F, G. I see some more creative words, I would say, but essentially it's kind of a. It's kind of a shame that it doesn't save the outputs here. Let's just screenshot it. So let's go back to GPT 3.5, run that again and then let's compare the results. Let's see. So already you can see, okay, this works to abcd. But if you directly compare these two results again, it's just another incremental upgrade. It's more creative. Does this mean it's 10 times more creative? No, but it's just a little more creative. And what all of this that we just talked about, right? Except of the vision model, that is like a mind blowing upgrade. What all of the other things, more creativity, better reasoning, it's safer and it can process more documents. All of these four things boil down to the fact that it's upgrade. It's a incremental upgrade. And it's not just upgrade at the simple use cases those chatgpt handled really well. It's upgrade in the deep end of the prompts. It's upgrade for the advanced prompts, for the advanced workflows we've been talking about here on this channel. It really steps up the game when you push its limits. Okay? So the more powerful your prompts and the more complex your prompts, the better you're going to get better results now. Okay? They pushed the limits because there was a certain limit. You couldn't Write like a 1000 word prompts that just went extremely crazy. It would get confused and it just didn't work perfectly. You kind of had to either sequence it in multiple prompts or make it more concise or not demand too much of it. That just expanded. Okay. And I think from everything I'm reading in between the lines here and look, this is, this is no like official like information. But from everything I'm reading here, it seems to be about 30 to 80% better at these various tasks. So whenever you ran into this wall with your prompts, they might have just broken down that wall or Moved it is maybe a better way to visualize it. They move that wall. It still exists. It's still not omnipotent. It's still not God. Right? But it's just a whole lot better at the things that we wanted it to be. So, with that being said, let's do a quick comparison of these two outputs. Let's do a quick comparison of these two outputs. For that, I would need to find my desktop, which I'm sure I will be able to do. Maybe not. Maybe not. Where did the screenshot go? We're just going to do this, But let me tell you this. It failed at the Penguin challenge. It didn't get me exactly 2,000 words. Although I asked for it. Okay. How in the heck, out of all the things that I could get stuck on, I get stuck on finding my desktop. It's kind of hilarious. Okay, I think I've got it. So you know what? Let's do this. Let's screenshot this response. And now that we have this response, we can open it up right here. And then we go and we say, new chat. Oh, this goes over here. We say, new chat. We switch to GPT4. Right, like that. And then we give it the same prompt and we compare the outputs like this. All right, There you go. Perfect. Perfect. So curious to see what it gave us here. So I think this is a good way to look at it. So in one of them, explain the plot of Cinderella, maybe. Wait, one last idea on how to optimize this. All right, this is great. This is great. So the black logo, GPT4, this, GPT3. So let's have a look. So both of them kind of fulfill the thing we've been going for here, right? It works. It works. A, B, C, D, E. Both of them represent the Cinderella story. But let's look at the details here. What changed? So this began with an actual word. Instead of just saying A, it says abandoned. Bullied. Cinderella dances elegantly. Fairy godmother helping. Instant kismet. Lost magic, novelty. Opulent prince quickly rescues seeking true, unconditional, vivacious, wedded Xanadu yielding zenith. Interesting. And what happened down here? It falls apart in the middle of it. If you look closely. So there's an A in here, right? With a vivacious, wondrous young Zephyr arrived. And what's going on here? Here? It just went crazy. This is not what we asked for at all. And if you just like me, try to push ChatGPT to its limits, you will realize. You will realize ChatGPT 4 pictures working? No, not yet. Not yet. We don't get to test that yet. But a lot of other things which we'll test now, you will realize that it falls apart at a certain point. You can't push it infinitely. You can't ask it. There's. There's only so many more times. You can ask it for 10 more ideas. Right? You all know that you ask it for like, give me 10 ideas, then you say 10 more. It does a good job. It gives you 20 in total. But then once you ask for like, give me 10 more, it starts falling apart. And if you tell it 10 more like five more times, it starts regurgitating the same answers. So this is the one that I've been excited to explore here. Okay. I really want to look at Ideation as one of the main use cases. How does this stack up against GPT 3.5? Okay, so let's do just that. Very simple prompts we're going to test here today. Okay. All right, so let's just go with give me 10 ideas for YouTube video just because that's familiar about artificial intelligence. And then say concretely, focus on language models and their applications for everyday use cases. Quite simple prompt. This is GPT 3.5. So let's see what this comes up with and let's evaluate. All right. 3.5 has the better result, though. I don't think so. I think GPT4 right there kind of hit it on the head with that Cinderella example. It was more creative, the words were better, it didn't use filler words, and it legitimately fulfilled your request. So now or just give me catchy YouTube titles for your ideas. Okay, so let's see. This is still GPT 3.5 chatgpt, right? Is chatgpt for. Only available on Playground Pro. It's also available on Playground Pro, but it's also available in here, as you'll see in a second. So just to test this, let's see. So we have like texting just got smarter and really what we want to be. Give me 30 more. Right. If we do this, what we're looking for here is for repetitive examples, right? So I guess I would have to read all of this. But let's see. Customer service, better healthcare, personalized language learning, content creation revolution. Okay. The next big thing, language models. This is repetitive. The next frontier in AI, natural landing understanding. That's kind of the same thing we heard before. Art of conversation and the science of language. This is. This is decent. It's pretty good. What happens if we say 30 more. And then we're going to test the same prompts in GPT4. All right. So I'm curious to see if it can go further because this was one of the biggest limitations when brainstorming ideas. You really ran into this wall where it was like, okay, it just keeps repeating itself, understanding their limitations. The battle against fake news. That's a new one. Can they create art? That's a new one. So let's just go ahead and be like, let's see. How often does social media occur here, for example? One time. Okay, that's decent. How often does science. One time. Okay, that's pretty good. Health care. Health. Four times. Right. So twice in one title. Here in one title. So across all the ideas, it used health four times. Okay, interesting. Now, it's a little bit of a shame that we don't have the history here, but I think what we can do is we can just copy this one to one. So I have a notepad open here on the second screen, and we'll just use the identical prompts and we'll simply see if it again repeats some ideas like health. It repeated multiple times. I believe it used it in free titles. And yes, we're pushing it, of course, but that's what this is about. So we did that. Then we said, Give me 30 more. And I suppose that the temperature of this is set to a level where it's quite creative, quite out there. So if you didn't know, the temperature setting regulates how much variation it brings into the prompts, how creative it gets, or how out of the box it kind of thinks. So let's try the same thing in GPT4 here, shall we? Okay, GPT4 and let's begin with the same thing. So let's see if there's any obvious differences here. So I'll post the first prompt. Let's see what it does. Oh, wow. So look right away, it leads with the title. That's amazing. I love that. I love that. Okay, two tabs. One for GPT 3.5, one for 3. 4. That's actually a good idea. Let's see if that works. That's a great idea. Appreciate that. Miguel, how did you get GPT4? So all you need is ChatGPT Pro, and then you can switch to GPT4 in here. If you. If you just tuned in and you want to hear all the key features, go back to the beginning of the video about one minute in, I recap everything. The biggest upgrade is probably the Vision model, which we don't have access to right now. Access to right now. I signed up for the API early access. Okay. So I've run this prompt like a gazillion times myself. I always look for YouTube ideas and I always tested on ideas around artificial intelligence to begin with. Okay. I can already tell you from a non data driven perspective, purely subjective. Yes, this is more creative. Like as simple as that. I don't know if you're looking for like some complex research 10 minutes after the release of this thing, but just from a subjective and first impressions perspective. This is better. This is better. And there's more variety in here, so. Interesting. And it's still writing Meme generation with AI. Look at that. This is one I've never seen before in the suggestions. Interesting, interesting. What else are you getting for in Playground? Yes, they have playground for now where you can play with it. Yeah. On jobs and layoffs. That's an interesting question, Chandler. How much effect this technology is having on jobs and layoffs. I just recorded a massive video today talking about strategies to save yourself from this, you know, new AI dominated economy we're kind of entering right now. So I put all my thoughts into that actually. Yeah, it's a great video, so you can look forward to that. I also had a researcher helped me research that one. So that's really going to be an excellent video finish now. Editing will be done in a week. Okay, so this is just to get back on track here so I can tell you just subjectively decoding fake news. This is the quality of this is just higher. Just straight up. I've run this prompt so many times. The quality of this is excellent. Let's see what it does if we hit it with the follow up prompt. Let's see what happens here. So give me a catchy. Give me catchy YouTube titles for your ideas. It kind of did that already by itself, right? It kind of did it by itself because it's smarter now. All right. It has better reasoning, it has better problem solving as I mentioned. Okay, so it's a little slower, but I can't hold that against it. Probably all of the 90,000 people that watch that live stream or like at least half of them are trying this out right now. Right. Just went live. I don't think anyone is safe in my honest opinion. That's a good point. But I think some people.