LARRUZ shows how to use GPT-4 free on nat.dev and compares it against ChatGPT, Claude, and other models. Tests include an alphabet-constrained Cinderella story, a scheduling puzzle, and a coding task to build a multi-tab joke website. GPT-4 handles instruction-following better, but the playground lacks conversational memory.
Questions this video answers
Channel started on January 14, 2023; first subscriber on January 18, 2023.
The channel crossed 50,000 subscribers in just over two months.
nat.dev.
An email address and a phone number are required to sign up.
Noon to 12:30.
Chapters
You most probably clicked on this video because you have heard of GPT4 and want to try it out. But like many people, you don't have access to ChatGPT plus, not only I'm going to show you how to access GPT4, but how to compare it with other models which are not publicly available, such as the Cloud from Anthropic. But before that, a big milestone for the channel we just crossed 50,000 subscribers in just over two months. This is just unbelievable for me. I started this channel on January 14, 2023 and got my first subscriber on the 18th. I'm simply blown away by the support and can't thank you enough for making this journey so special. Your comments, likes and shares keep me motivated to bring you the best possible content. Thank you for being part of our journey and here's to the next 50k. So back to the video. If you go to the playground on OpenAI's website and you look for GPT4, it is not available there for free users, but there is a website which actually gives you access to GPT4. It's called nat.dev so first you need to sign up using your email address and phone number. You actually need both of them, so just go ahead and sign up. Once you sign up and log in you will see an interface like this. So they have a playground, then a compare and then there is a link to a Discord server. So before exploring the platform, a quick shout out to Nat Friedman. He's actually the one who made this website available to everybody but you might want to have a look at his cloud build. Just make sure you use this website responsibly and don't exploit it. It's simply for comparison of different models and how their output looks like. So if you come here to the right hand side there are a number of models available. So for example these are the open source models. Then some from hugging face. Then for example you have the Cloud Instant from anthropic or cloud 1.2 some models from CO here and then you have the OpenAI models including GPT4, this is ChatGPT and some other for example the text DaVinci3 models as well. So you can simply type in whatever query you have. It's very similar to the OpenAI playground. Submit it and you will get a response. But I'm really interested in comparing different models. So you can come to the Compare tab and then here you can select the models that you want to compare. So for example I have selected the cloud instant version 1.0. So let's say, let's remove that. Okay, so I want to select 1.2. This is Chat GPT Cloud 1.2 OpenAI and I think they also have both the Llama and Alpaca models as well. So let's see, let's look at check for llama. Right, so llama 65 billion parameter model as well. Okay. And then you have a few options that you can set so you can play around with those. So what I want to do is I want to compare these four models for different tasks, including a programming task. Use it as a prompt generator for mid journey. All right, so the first test that I want to do is to run this prompt. So this is one that was shown in the demo of GPT4. So explain the part of Sandrana in a sequence where each word has to begin with the next letter in the Alphabet from A to Z without repeating any letters. So let's see how good these models are at doing this task. Right, so all you have to do is write your prompt and then click submit. And the great thing about this is you actually see what's the speed. So how many characters per seconds are being generated now in this case. So Chat GPT definitely did not follow the instructions at all because you already see that after Cinderella. So that's like A and then it's supposed to be a letter starting with a word starting with letter B but it completely ignored it. Same for the cloud model, but if you look at GPT4 so it's a have like abused beauty Cinderella. So like A, B, C, D, E, F, G, H, I and it seems to be have actually followed it pretty accurately. So that's pretty nice. So this shows you that you are actually using GPT4 and the llama model. It's still trying to come up with something so I'm going to stop this. That's just one quick comparison that we did. Okay, now let's look at the advanced reasoning capabilities. So for this to compare these different models, I'm going to copy the prompt from GPT4 website and let's go back and let's paste it here. All right, so we pasted it. One thing I noticed is on the website, if you simply copy something and try to paste use it control Ctrl C. It doesn't seem to be work but okay, let's submit this and let's see what the responses. The answer is supposed to be noon to 12:30. So that's the only time when you have all three of them free. So let's see. Chat GPT, there are two possible windows within which all three participants are available noon to 12:13. Okay, so this is actually incorrect because one of the person is not free at this time. I actually did the calculation. So what does cloud says? Let's break it down step by step. Right. So it's looking at the availability of everybody. Okay, so it's actually able to figure this time out. Okay. And then here's the response from GPT4 as well. So this is actually the correct answer. So it still, I think got it wrong. But if you run it again, you might see a different answer. And the llama model is still trying to process this, so it's not really coming up with anything. So let's stop this now. The real beauty of this website is actually to compare these models as well as understand the effects of these parameters. So you can actually play around with these. I might do another video on this. Now let's see how these, you know how good these models are at following instructions in terms of if you want to create a prompt for something like midjourney. One thing I did was I actually replaced the llama model with this command large X nightly from cohere, because that wasn't doing anything at all. All right, so here I'm defining the structure of what a mid journey prompt should look like. Right? So it's supposed to have a subject of the image, then 5 to 7 descriptive keywords and camera type and some other parameters that I want to look at. I found the basis for this prompt on the Internet and then modified a few things in there. Right. And then I'm asking it based on the above structure, generate five prompts for the following subject. So now I'm going to define the subject. Let's say subject is going to be classic cars. And one thing you might have noticed so far is it doesn't have the. It doesn't have the ability to converse. Right? So it's very different from what you expect from ChatGPT. Right, so here you can simply run a command and get an output because it's simply a tool to compare the output from all these different models. All right, so it seems like all the models have generated the results. So I'm going to copy each one of them and let's see how the results look like in mid journey. All right, just one thing to notice. The Cohab model actually didn't return anything. It simply rewrote the structure formula that we gave it. Chat GPT, give us a prompt. One thing was that the Aspect ratio. I'm defining in the width to height, right? So I think I might need to actually add the dash ash AR here. Okay, so let's look at the results. Okay, so here are the results. This is based on the prompt that was generated by GP ChatGPT. The second one is based on the cloud from Entropic. The and the third one is based on the one generated by GPT4. So these all looks pretty good. The goal of this comparison was simply to look at the ability of these models to generate prompts for AI image generators. Not ready to compare the output results. But all three of them are able to do the task pretty nicely. UAD has seen that both ChatGPT and GPT4 able to code. So let's look at a very simple example and see what the output is going to be. So I'm asking it write a code for a website that has multiple tabs and each tab displays a random joke. Let's see what these models comes up with. Now all four of the models were able to generate codes, but only ChatGPT and GPT4 were actually able to kind of explain it a little bit. So it did put some explanation within the code blocks. Right. For example, it says that this creates a simple websites with three tabs. When you click on a tab, it displays a random joke from the jokes array. Okay, so let's look at actually the final output of each of these code. And so the first website that we are looking at that is from the Cohere model. It did put a couple of links, but that don't seem to be working. The next is the clone model. Instead of tabs, it actually put three, I think, different links and simply displayed all three jokes here. So it doesn't really follow the instruction that we provided. Here's a chat GPT, it did put two tabs, but the jokes are missing and GPT 4, so it actually followed all the instructions. So there are tabs. Then you have one joke. If you click on tab two, it generates another joke, tab three, another joke. And if I actually click on these multiple times, so each time it will generate, it will pick a random joke from that array of jokes that it has. Okay, so I selected chat GPT4 and let's ask it what's the name of the model that it is currently running? Let's see what things. Okay. All right, good. So it knows that it's GPT4. That's a pretty good indication because some people actually had trouble with it. When was it last updated? Let's see is the language. I don't have the real time information. Please refer to openzi. Okay. It seems to actually remember the context by saying what is the cut off update? In your training you might encounter this because in the Background it's using GPT4 API and there seems to be a limit on it because of the plot that I showed you, like how much money it's costing the developer. So it seems like I was wrong when I said in the in the beginning that you cannot really chat with it. I think in the playground you can potentially chat with it. So check it out, have fun with it, experiment with it. Hope you like the video. Thanks for watching. See you in the next one.
