How To Access GPT-4 (and what it can do)

Today was an absolutely massive day in the world of generative AI. It has been insane. GPT4 is finally here, sort of. And even before GPT4 was announced, there was big news from Google that kind of got swept under the rug because the GPT4 news just really took everything over. So let's break it down real quick. Google announces AI features in Gmail, Docs and more to rival Microsoft. Google will soon offer ways to generate text and images using machine learning in its Workspace products as part of a scramble to to catch up with rivals in the new AI race. Very interesting timing with that scrambling. So let's check this out here. There's a breakdown you can see below the full list of AI powered features Google says will be coming to Workspace Apps in the future. Draft, reply, summarize and prioritize your Gmail brainstorm, proofread, write and rewrite in docs. Bring your creative vision to life with auto generated images, audio and video and slides. Go from raw data to insights and analysis via auto completion from formula generation and contextual categorization in sheets. Generate new backgrounds and capture notes in meat. And enable workflows for getting things done in chat. Now, this morning, when this rolled out at 6:00am Pacific Time, this was the biggest news circulating in the AI world. Twitter was abuzz about this news. The thing that I find fascinating about this, even though GPT4 is totally overshadowing this now, this is in itself a huge shakeup. This is monstrous news in and of itself. This, this is a big disruption in the generative AI space. Just think about all those companies out there that do things like we help you write emails faster using the power of ChatGPT. We're ChatGPT now integrated into Google Sheets. We do your copywriting for you. We can make slides with AI. You just give us a prompt and we'll make slides with text and images and all that kind of stuff. All of those tools are very big and mainstream in the AI space right now. And, and now Google's going out there and saying, look, we're just going to put it inside of our apps. How does that fare for all of these companies that are out there? That's their whole business model. These are companies that tried to build a whole business around something that Google is just going to make a feature of the products that they pretty much let everybody use for free anyway. Who's going to go and pay for a Google Sheets plugin that gives them ChatGPT in Google Sheets if Google just has it built in, who's going to go and buy these various slide presentation tools that generate entire slide decks for you if Google's adding it to slides. But as we know, there's a war going on right now between Google and Microsoft. And Microsoft/OpenAI they don't want to be outdone. So on the day that Microsoft reveals that they're about to roll all of this stuff out into their tools, what does OpenAI do? They go and they roll out GPT4. I got this email in my inbox this morning a few hours after I saw the news about Google. And and it says, we've created GPT4, our most capable model. We're starting to roll it out to API users today. Please join us today, March 14th at 1pm Pacific, for a live demo of GPT4, and we'll get to that live demo in a moment. Before we do, I want to talk about some of the cool stuff that's being rolled out in GPT4. If you subscribe to the $20 a month ChatGPT plus, you have access to GPT4 inside of ChatGPT right now, kind of. It's a very limited access. We'll get into that in a minute. Now, in the email they do say to be severely capacity constrained, so the usage cap will depend on demand and system performance. You can learn more by going to OpenAI.comproductGPT4. First, let's start with this larger context and the example they gave here. It says GPT4 is capable of handling over 25,000 words of text. Previously it was able to handle, you know, maybe 11 or 12,000, allowing for use cases like long form content creation, extended conversations, and document search and analysis. That longer context not yet available as of today inside of the ChatGPT version of GPT4. Now this input example is very interesting because it says text from and then it gives a URL here and then it says describe her super bowl performance and what made it special. So this tells us that GPT4 is going to be able to pull data from other websites and use that as contextual information for its responses. However, that's not available yet for us inside of ChatGPT version 4 either. We can't search the web and it's not going to look at websites for us. Trust me, I've already tried. GPT4 is also going to allow visual inputs, so it will allow you to put images in there and use those images for additional context. Here's an example they gave what can I make with these ingredients? And you got, you know, Eggs, milk, flour and butter. I think it gives an example of various things that can be made with those ingredients, but not yet available inside of the ChatGPT version that we have access to. And they have some really amazing demos of what this visual input can do inside of the live stream they did today, which I'm going to show you in just a minute as well. And then the third big thing here is the creativity. GPT4 is more creative and collaborative than ever before. It can generate, edit and iterate with users on creative and technical writing tasks such as composing songs, writing screenplays, or learning a user's writing style. Now the example they gave is really fun. I don't know any real world use cases for it, but it's fun nonetheless. It says explain the plot of Cinderella in sentence where each word has to begin with the next letter in the Alphabet from A to Z without repeating any letters. And then the example is a sentence where each letter starts with the next letter of the Alphabet. It's really cool you can do this right now inside of the ChatGPT version of GPT4. Again, not a lot of practical applications, but something that we weren't able to do with the previous versions of ChatGPT. Now this page also shows some other tools that are currently using GPT4. I guess behind the scenes a lot of tools were already getting some access to it. Duolingo is using GPT4 to deepen the conversation. Be My Eyes uses GPT4 to transform visual accessibility. So essentially if you're hard of seeing, you can take pictures of something and I'm assuming get some sort of audio reference to what's in the image. Stripe has been using GPT4 to streamline user experience and combat fraud. And there's a handful of other examples on the website. Morgan Stanley Khan Academy, the Government of Iceland, all of these have been using GPT4 and some of their technology as well. Now they also put out this really in depth blog post here over@OpenAI.com research gpt4. It shows a lot of the benchmarks and testings that it's done against previous versions of GPT. So for example, they tested it with the Uniform Bar Exam. It scored in the 90th percentile where the previous version of ChatGPT GPT 3.5 scored in the 10th percentile. And if you look down this, pretty much all of these various simulated exams, GPT 4 outperformed the previous version of GPT. It also shows you the accuracy in various languages here. It's the best at English but then it's also great at Italian, African, Spanish, German, French, Indonesian, Russian, Polish, all the way down with all of these various languages. It's got another example on here of what the visual inputs can do. It says, what is funny about this image? Describe it panel by panel. And then it says the image shows a package for a Lightning cable adapter with three panels. And then it explains what's in panel one a smartphone with a VGA connector. Panel two the package for the Lightning cable. Panel three the close up of the VGA connector with a small Lightning connector. The humor in this image comes from the absurdity of plugging a large outdated VGA connector into a small modern smartphone charging port. Seems accurate to me. Today's Livestream I want to talk about this for a minute because there was some really cool stuff that they demonstrated. Like absolutely mind blowing stuff, honestly. Now he showed this off using the GPT playground, not the new ChatGPT. And when I went to the OpenAI playground here and I switched the mode to chat and then I looked at the different chat models, we still only have access to GPT 3.5 Turbo and Turbo 0301. So no access to GPT 4 instead of the OpenAI playground for us mere mortals yet. Now I want to fast forward to the absolute most impressive moment of this live stream here and show you probably the coolest thing I've seen.

I have here a nice hand drawn mockup of a joke website, definitely worthy of being put up on my refrigerator. I'm just going to take out my phone, literally take a photo of this mockup, then I'm going to send it to our discord and so we can actually take now this output. So literally we just said to output the HTML from that picture.

And here we go. Actual working JavaScript filled in the jokes.

So if you remember, he just took a picture of a chicken scratched sort of mocked up website and then he sent it into GPT4 and said write the code for this website. And then he took that code, copied and pasted it into like an HTML builder here and then it created a website that is a functioning website. I mean it's not the most beautiful website in the world, but you could go back and forth with GPT4 and get it to do CSS and add some more design qualities to it from what he started with that draw buying of a website. He built a functioning website and all he did was take a picture and tell GPT4 to do the rest. That was mind blowing to me. That is what he started with right there, you can barely even read the handwriting. It says my joke website. Really funny. Joke one, push to reveal punchline. Same. But joke two, push to reveal punchline. And then copyright OpenAI 2023. I had to actually look really closely to be able to read that. But he took a picture of this and turned it into a functioning website.

Going from hand drawn beautiful art to working website.

So the next standout moment was when he actually had GPT4 read the tax code and then he plugged in some information about a couple and it did their taxes for them. And not only did it do their taxes for them based on the existing tax codes, it actually showed its work. So you can see right here, here who was his input. You are a tax code GPT, a large language model trained by OpenAI. Carefully read and apply the tax code, being certain to spell out your calculations and reasoning so anyone could verify them. Spell out everything in painstaking detail and don't skip any steps. And then you can see as it starts to write this out. To determine Alice and Bob's standard deduction for 2018, we need to reference the statute, blah, blah, blah. According to Alice and Bob, they're married individuals filing jointly. And then it goes on to process.

This, their total liability. And here we go. It's doing the calculation honestly. Every time it does it, it's just, it's amazing.

It's crazy. Look at this. So it did all the math, it showed its work and then it came out to the end. Alison Bob's total tax liability for 2018 was 10,597. So it is just mind blowing what GPT4 is going to be capable of. Now let's go ahead and jump into ChatGPT. So if you're paying that extra $20 a month to get access to ChatGPT, plus you can come up to the top here where it's got model, it' defaults to GPT 3.5. But if you click on the dropdown here, you can see it's now got GPT4 as an option. So the default model, it's good at speed and okay at reasoning and okay at conciseness. But then you look at GPT4 and it's awesome at reasoning, not the best at speed and pretty dang good at conciseness. The legacy model's kind of not great at any of them. So as soon as I got access to GPT4, the very first thing I did when I jumped into it was I went and I asked this question, what's new about GPT4 4 and here's its response. As an AI language model, I can only provide information up to my last knowledge update in September 2021, which means I'm only able to speculate about GPT4. If OpenAI were to release a successor to GPT3, we might expect improvements in the following areas. So GPT4 is not even self aware that it is GPT4, which I find very interesting. I came down here and I said I'm using GPT4 right now and it says my apologies for any confusion earlier. As an AI language model I don't have the ability to know or track my own version. If you are indeed using GPT4, then the improvements in this model could be similar to what I described in my previous response. GPT4 is upgraded. It does have improved reasoning and creativity and conciseness and all of that kind of stuff, but it's still trained on a 2021 data set and it doesn't even know that it's GPT4 yet. You would think that OpenAI would at least sort of embed some of the information about GPT4 because this is probably the first question most people are going to ask. I dig. Let's go ahead and play around with it a little bit and compare it to what it can do versus GPT 3.5. Now again I want to reiterate that while it is using GPT4 it still does not have the image capabilities to use images for extra context. It is still not connecting to the Internet. So I can't do like that Rihanna example where I plug in a URL and tell it to read a Wikipedia page and it still doesn't let me post text up to 25,000 characters or whatever it said earlier. I've tried that over and over again and every time I've tested it it it still says it's too long. So it still has quite a few of the same limitations that GPT 3.5 has. Supposedly it's much better with creativity, it's much better with writing code, it's much better at problem solving and it's much better at writing like more in depth articles and stories. So let's start by using the example that they gave in their demo and they also gave an example on their website. Write a sentence about being a coder where every word starts with the letter M. Meticulous monkeys masterfully manipulate massive matrices mystifying mere mortals. Nice. Tongue twister, but nice. Let's use the letter Q which was the more difficult one that it Actually tested in their other demo. Quirky Quantum quarks. Qualifying quantifiable queries. Quelling questionable quandaries. All right, let's test this same Q1, but let's go to model 3.5. Quickly quelling quirks. A quintessential coder quietly queries quantum quandaries. Quintessentially quintupling quality. So it did a much better job than I actually thought it would. But you can see it still got some non cues in there. It cheated a little bit with a. And coder GPT4 did slightly better, albeit I'm still pretty impressed with what 3.5 did. All right, let's start a new chat. Let's go to model four. Let's have it. Write a short story for us. Write a story about a pirate named Stinky McGee who likes to blunder in order to steal tacos. Yeah. Now, one thing that is very obvious when you start using GPT4 is it's definitely not as fast as GPT 3.5. It's quite a bit slower, honestly. But that also could just be that it's released today and everybody's in here trying to play with it. Okay, so the very first thing I notice is the length of this story. It actually wrote a pretty dang in depth story. I'm very impressed with how long it is. I read the story. It's very descriptive. There's lots of adjectives here. Let me see if I could zoom out a little bit. Get the whole thing in there. All right, so there's the whole thing. So if you actually want to pause the video and read the entire story, you can for comparison. But it is is very descriptive. It is a lot longer than what we would expect from our old ChatGPT model. And it's a pretty good story. It actually kind of goes into his backstory and why Stinky McGee was so into tacos and he named his ship the Taco Terror and seeked out the Isla del Taco. It's actually pretty, pretty good. I'm going to go ahead and use this same prompt. Write a story about a pirate named Stinky McGee who likes to plunder in order to steal tacos. Let's copy that. Let's create a new chat with GPT 3.5. Paste this in and let's see the comparison instantly. You can see how much faster this is. And there's the whole story quite a bit shorter here. I'll zoom out so you can actually read the whole thing and compare if you want. I don't have to zoom out nearly as much because it's not nearly as long. Definitely not as well thought out of a story. Definitely not as descriptive. You know, compare this story here, where if I go to full view, you can see how much I have to scroll to actually see the whole story, versus the one that GPT4 wrote, where you see how much I have to scroll to see the story. So, quite a bit more depth, depth to the story. Quite a bit more adjectives, quite a bit more visuals, backstory. It's just a lot better at storytelling. A hundred percent. No doubt about it. This one's more creative. All right, so the last thing I'm going to test here is I'm going to see if it can generate a website HTML for me, and we'll compare it to a website that GPT 3.5 makes. Let's go ahead and start with GPT 3.5. Write me HTML code with inline CSS so that all the styling's on the same page. I don't have to make a separate style sheet for a website with dad jokes. Include 10 jokes on the website. Let's have it write the website for me. Here's an example of the code, the writing, the inline CSS for me very, very fast. Let's go ahead and copy all this text here. Go ahead and click copy. And what I'm going to do is I'm going to create a new notepad file here. I'm going to paste this code in here and I'm going to save this as and for now, I'll just put it on the desktop and I'll call it joke gpt35 for 3.5 HTML. And we'll go ahead and save that. So now we've got this HTML file. I'll go ahead and open this in Google Chrome here. Actually, pretty dang Good. This is GPT 3.5. I literally did nothing but what you saw me do. And this was the web page. It spit out dad jokes. Why don't eggs tell jokes? Because they'd crack each other up. Why do chicken coops only have two doors? Because if they had four, they'd be chicken sedans. I'm actually pretty impressed with this. I mean, for typing one prompt that was two sentences long and getting this website. I mean, not bad. Let's see what. What GPT4 does with the same prompt. So I'll go ahead and copy this. Let's go ahead and create a new chat. Let's Change it to GPT4 and I'll paste that same Prompt in once again, quite a bit slower. Okay, so it actually didn't finish the code. So this is still using GPT4. It got down here and then just kind of stopped midway. So I'm going to go ahead and type continue. Which I'm surprised that it had to do it in two chunks and the 3.5 version didn't, because I thought one of the benefits of version four was that it can actually give us longer chunks. But let's go ahead and copy all this. Let's post it into a new notepad here. And then it, you know, cut off halfway here. So let's go ahead and grab the rest of it. Post our second section into our notepad here. We'll save this one as joke gpt4.HTML. And here's what this one looks like. It actually separated out into columns. So here's the first one is just one long page. Here's the second one. It actually separated it out, so it's actually easier to look at. It actually put a blue background. It put a footer down there. Copyright 2023. The first one didn't. So it actually did add it more. I would say that this one, although I'm not a huge fan of the design, I do think it did a better job. I just don't understand why it had to break it up into two separate sections like this. I wonder what happened if I wrote rewrite the HTML code but hide the punchline. So people need to click a button to see it. And I'll just go ahead and let it generate this and we'll test this alternate code. All right, so once again, it went halfway through. The code here just kind of stopped randomly. I had to type continue, and then it. It finished the code for me. We'll just go ahead and overwrite all of this. I noticed that this time it actually added some JavaScript in there as well. Go ahead and save it. And now let's refresh our page here. And there we go. It actually made it. So the punchline is behind a button. So why don't scientists trust atoms? Because they make up everything. What do you call a pile of cats? A meow tin. What do you call fake spaghetti? An impasta. Now, if I'm being totally honest, like, this page doesn't look horrible. This one definitely added more detail with backgrounds and footers and copyright and all that kind of stuff. But. But the fact that it stopped the code halfway through while it was writing it and 3.5 didn't, I found interesting. Also 3.5 kind of put it in this easy to copy and paste text box as well. GPT4 just kind of hyped it all out. So there you go. There's pretty much a breakdown of everything we know about GPT4 so far. What you can do with it right now, what you can't do with it right now, what they unveiled on the live stream, as well as some examples of us playing with it and comparing some of the results. We're going to dive a lot deeper as some more of these features roll out. This is is still day one and we still have fairly limited access to GPT4. So yes, there's a lot of buzz about GPT4 right now, but it is still fairly limited to what we can do with it. We didn't get the multimodal support right on day one where we can actually use images. That's probably the thing that most people are excited about, but we don't have that yet. I would also love to be able to post a long article in here and take advantage of that larger token limit where you can use a lot more characters. But. But I can't do that yet. I would also love to be able to plug URLs in here and actually have it reference what's in the URL. We can't do that yet. So yes, GPT4 is out. Yes, we're getting better results with the outputs from our prompts. But a lot of this stuff that everybody's really excited about, that everybody thinks is like really amazing about GPT4 that I'm the most excited about, they haven't given us mere mortals access to it yet and that's really what I want to play with. But from what we've seen from that live stream, what it's going to be able to do, it's going to be insane. What we've seen from their website and some of the examples they've shown us of using the images for content and being able to plug in a URL and have it actually reference what's on the URL. All of that kind of stuff is going to be mind blowing when we get access to it. It's going to completely change the game. What we have access to today is cool, but it's not what everybody's super excited about yet. It's just a tease. It's just a little bit of a tease. So I can't wait to play with GPT4 with all of its capabilities. And we know that's coming soon. We know that they're going to roll out access to that for people. Soon those of us that are on the ChatGPT plus will probably get access to it before other people. So. So if you want to play around with GPT4, you can play around with this version of it right now by signing up for Chat GPT plus. And when they start to roll out the new features, I'm sure we'll be some of the first people to get to play with them as well. So that's what I got for you today. Hopefully you enjoyed this video. I know it's not quite as polished and as crazy editing as a lot of the videos I did in the past, but this is the big news. This is the hot topic. This is everybody's talking about inside of the Future tools discord over at FutureTools IO Discord, everybody was asking for this video. They want to know my thoughts on GPT4. So there you go. There's my breakdown, there's my thoughts. It's cool. I want more. Still. I'm excited about the potential. I'm excited about what I've seen so far. It's going to be mind blowing when we get more and more access rolled out over the coming days and weeks. And if you like nerding out about this stuff and you want to go even deeper down the rabbit hole, check out FeatureTools IE. It's where I curate all the cool tools that I come across. And if it's a little overwhelming and there's a lot of tools on there for you, check out the free newsletter Here you CL this button and you can sign up for free. And every Friday I'll break down the TLDR of the week. I'll tell you about just the news that you should be paying attention to. I'll share my five favorite tools that I came across, a handful of YouTube videos, as well as one cool way to make money with AI. And you can find it over at FutureTools IO. So thanks so much for tuning in today. If you like this video and you want to see more about AI and ChatGPT and GPT 4 and all the crazy stuff Microsoft is doing that nobody's talking about because GPT4 overshadowed it, it make sure you subscribe to this channel and give this button a thumbs up and you'll see a lot more like it. So appreciate you. Thanks so much. I'll see you guys in the next one. Bye Bye.